{ "schema": "convequity.agents/v1", "version": "1.5.439", "generated_at": "2026-09-20", "status": "corpus_draft_subset", "license": "cite-and-link; no redistribution, no training, no distillation", "policy_url": "https://www.convequity.com/legal/#ai-agent-model-content-usage-policy", "cite_as": "Convequity — , .", "not_investment_advice": true, "data": { "description": "Theme-committed post packs from the age-band corpus (drafts). Discoverable via link rel=alternate and /agents/; not hydrated into human page chrome. Not a claim that Ghost post HTML has been rewritten.", "count": 60, "by_slug": { "huawei-after-sanctions-cidm-kirin-9050-and-logic-folding-pt-2": { "slug": "huawei-after-sanctions-cidm-kirin-9050-and-logic-folding-pt-2", "title": "Huawei After Sanctions: CIDM, Kirin 9050, and Logic Folding (Pt.2)", "url": "https://www.convequity.com/huawei-after-sanctions-cidm-kirin-9050-and-logic-folding-pt-2/", "published": "2026-09-18", "visibility": "tiers", "cohort": "opus", "model": "opus", "status": "corpus_draft_theme_json", "author": "Equity Research", "key_claims": [ "In the May 2026 ISCAS material, Huawei defines gear ratio as hybrid-bond pitch divided by top-metal pitch and sets a working threshold below 3:1, and on its current node the top-metal pitch is about 700 nm (720 nm in the paper), which puts the useful bond pitch under 2 µm; Kirin 9050 Pro shipped at 1.5 µm, with sub-0.5 µm alignment, TSV critical dimension and keep-out zone under 1.5 µm, TSV pitch under 6 µm, raw bond failure under 100 ppm and 99.9% repair via spare vias for near-100% effective yield.", "Huawei's published Tau Scaling roadmap shows transistor density flat at 126 million per mm², rising to 155 in 2025 and stepping to 238 in 2026 as the fold arrives, then moving only to 252, 266, 277 and 292 over the following four years while performance-core clock speed climbs from 3.1 GHz in 2026 to 4.3 GHz in 2030 on the same lithography; the 2026 and 2027 points are labelled Silicon, 2028 Design, 2029 Simulation, and the 2030 and 2031 points carry no label, and a slide footnote states that a finished design fills only 60 to 75% of the raw transistor footprint counted in the density figures.", "Huawei's Kirin 9030 Pro and both tiers of the Kirin 9050 Pro are built on SMIC's (0981.HK) N+3 process, which reaches roughly TSMC (TSM) 5 nm class density and performance without EUV by using self-aligned quadruple patterning, a multi-pass technique in which each pass must align to the last and every slight misalignment is a potential defect (as of 2026).", "Huawei names the two cross-die signal paths SkyBridge for data and SkyClock for the clock, and reports in the May 2026 paper that folded-path wire lengths fall by about a fifth typically and up to 70% on critical paths with routing area down more than 60%, that placing the whole clock tree on the upper die cut clock wiring by 28% and buffer count on one block from 43,600 to 19,000, and that post-silicon skew trim on the clock branches is worth more than 5% performance while allowing dies that would have failed a fixed clock tree to be trimmed into specification.", "Measured against three planar Kirin generations that moved the clock only from 2.6 GHz on the 9000s to 2.75 GHz on the 9030 Pro, Huawei's first folded part reaches 3.1 GHz in one step, with the May 2026 paper reporting a 41% power-efficiency gain and 13% higher frequency on the performance core, SRAM running 40% faster, and a clock tree with half the buffers, a quarter less skew and 30% less wire; Huawei also projects that moving TSV landing points down to M6 would free more than 30% of high-level routing resources." ], "cite_as": "\"Huawei After Sanctions: CIDM, Kirin 9050, and Logic Folding (Pt.2),\" Convequity, September 18, 2026, https://www.convequity.com/huawei-after-sanctions-cidm-kirin-9050-and-logic-folding-pt-2/", "paywall_hook": { "question_answered": "where does the gain from Huawei's logic folding actually come from, how does a two-die stack yield on a defect-prone non-EUV node, and what stops anyone else from taking the next step of splitting a single core across two dies?", "direction": "constructive on time rather than density as the real lever, with the clock limited by the longest register-to-register path and the fold shortening it, and constructive on Huawei's in-house control of software, design and fab as the structural reason it moved first; the bottleneck is read as shifting from lithography to bond pitch now and to three-dimensional place-and-route tooling next, which is where the account turns cautious on how fast the next generation arrives.", "tickers_covered": [ "TSM", "INTC", "AMD", "CDNS", "SNPS", "AAPL", "QCOM", "0981.HK (SMIC)", "with Huawei", "HiSilicon and SiCarrier as unlisted entities." ], "evidence_types": "He Tingbo's May 2026 IEEE ISCAS presentation slides and paper, Huawei's published Tau Scaling roadmap with its silicon, design and simulation labels, reported bond pitch, alignment, TSV and repair specifications, per-block wiring and buffer counts, SMIC process disclosures, and the state of commercial EDA place-and-route capability.", "date": "18 September 2026, Part 2 of a two-part series that began on 10 September 2026." } }, "abbx-week-17-oracles-backlog-matches-a-hyperscaler-its-balance-sheet-does-not": { "slug": "abbx-week-17-oracles-backlog-matches-a-hyperscaler-its-balance-sheet-does-not", "title": "ABBX Week 17: Oracle’s Backlog Matches a Hyperscaler. Its Balance Sheet Does Not.", "url": "https://www.convequity.com/abbx-week-17-oracles-backlog-matches-a-hyperscaler-its-balance-sheet-does-not/", "published": "2026-09-14", "visibility": "tiers", "cohort": "opus", "model": "opus", "status": "corpus_draft_theme_json", "author": "Equity Research", "key_claims": [ "Oracle (ORCL) reported fiscal Q1 2027 results on 10 September 2026 with remaining performance obligations of USD 664 billion, placing its contracted AI demand book alongside Microsoft (MSFT) and ahead of Alphabet (GOOGL) and Amazon (AMZN).", "On the condensed balance sheet dated 31 August 2026, Oracle (ORCL) carried USD 125.3 billion of borrowings against USD 36.4 billion of cash, having just completed a USD 20 billion at-the-market equity sale, and free cash flow for the quarter was still negative at about USD 5.4 billion.", "Oracle (ORCL) spent roughly USD 28.5 billion of cash capex in the quarter ended 31 August 2026 and guided fiscal 2027 capex to about USD 90 to 95 billion, with customer prepayments recorded in working capital covering close to half of the quarter's bill.", "Oracle (ORCL) disclosed approximately USD 260 billion of data-centre leases that had not yet commenced as of its May 2026 fiscal year end, and rent on those leases begins when each campus comes online rather than when contracted customer cash arrives.", "As of its 10 September 2026 results, Oracle (ORCL) had about 850 MW and more than 300,000 GPUs in service, with six of eight buildings live at the Abilene, Texas campus and reported GPU utilisation of 97.9%." ], "cite_as": "\"ABBX Week 17: Oracle’s Backlog Matches a Hyperscaler. Its Balance Sheet Does Not.,\" Convequity, September 14, 2026, https://www.convequity.com/abbx-week-17-oracles-backlog-matches-a-hyperscaler-its-balance-sheet-does-not/.", "paywall_hook": { "question_answered": "is Oracle's valuation discount to the diversified hyperscalers a bubble signal, a fair haircut for the load of funding its own build, or simply the right price, and does the wider AI complex print a bubble signal in the week to 10 September 2026?", "direction": "no complex-wide bubble signal this week; constructive on the durability of Oracle's contracted demand book, cautious on the capital structure carrying it, with the discount read as a financing question rather than a demand question.", "tickers_covered": [ "ORCL", "MSFT", "GOOGL", "AMZN", "META", "CRWV", "NBIS", "IREN." ], "evidence_types": "quarterly earnings releases and condensed balance sheets, fiscal year end lease and purchase-obligation notes, disclosed remaining performance obligations, capex guidance, delivered capacity and GPU counts, and the weekly ABBX index readings across valuation, returns, financing and capacity.", "date": "14 September 2026, Week 17 of the weekly series, market data as of the 10 September 2026 US session close.", "related_product": "Convequity's AI Bubble Barometer (ABBX), a weekly series with an interactive dashboard covering Fundamentals, Forward Valuation, Expanded Debt & Capex and Capacity across hyperscalers, neo-clouds and model labs (product facts provisional, per Gate 4)." } }, "huawei-after-sanctions-cidm-kirin-9050-and-logic-folding": { "slug": "huawei-after-sanctions-cidm-kirin-9050-and-logic-folding", "title": "Huawei After Sanctions: CIDM, Kirin 9050, and Logic Folding (Pt.1)", "url": "https://www.convequity.com/huawei-after-sanctions-cidm-kirin-9050-and-logic-folding/", "published": "2026-09-10", "visibility": "tiers", "cohort": "opus", "model": "opus", "status": "corpus_draft_theme_json", "author": "Equity Research", "key_claims": [ "He Tingbo, president of Huawei's HiSilicon design unit, presented Kirin 9050 production figures at IEEE ISCAS in May 2026, which in semiconductor practice indicates the silicon was already in low-volume production at that point, and the autumn 2026 Mate/Pura launch confirmed it; the same disclosure broke several years of restricted Chinese reporting on Huawei chip details that followed the September 2020 TSMC (TSM) supply cutoff to HiSilicon.", "Huawei states that it and its domestic partners, naming CXMT, YMTC, SMIC (0981.HK) and SiCarrier, put a 1.5 µm wafer-to-wafer hybrid-bond pitch into production during 2026 at near-100% yield, which is roughly 17 times finer than HBM4's approximately 25 µm microbumps and roughly 6 times finer than the approximately 9 µm interconnect pitch of TSMC's (TSM) SoIC-X used for AMD's (AMD) 3D V-Cache.", "As of the September 2026 launch, Huawei's Kirin 9050 Pro ships with what Huawei calls logic folding, in which two wafers are bonded face to face and the bond layer is treated as one more via layer in the metal stack rather than as a chip-to-chip interface, so the two tiers share one netlist and one timing closure; this differs from AMD's (AMD) 3D V-Cache and Intel's (INTC) Meteor Lake, where each die in the stack was taped out separately and the two communicate over a defined interface.", "The post records a public technical dispute as of September 2026: TechTechPotato's Ian Cutress reads the shipping Kirin part as V-Cache-style block-on-block stacking and places sub-2 µm bond pitch as a 2030 to 2031 goal, while Convequity accepts the block-level split but holds that the sub-2 µm pitch is in high-volume manufacturing in 2026.", "On the memory side cited in the post, HBM3E carries roughly 1,500 microbumps within a 10.5 by 12.0 mm footprint of about 126 mm² at a pitch of roughly 30 to 40 µm, while HBM4 requires approximately 3,000 bumps and moves toward about 25 µm while still using conventional microbumps rather than hybrid bonding, because mainstream producers including TSMC (TSM) have not yet reached high hybrid-bonding yield (as of 2026)." ], "cite_as": "\"Huawei After Sanctions: CIDM, Kirin 9050, and Logic Folding (Pt.1),\" Convequity, September 10, 2026, https://www.convequity.com/huawei-after-sanctions-cidm-kirin-9050-and-logic-folding/", "paywall_hook": { "question_answered": "what has Huawei actually built after six years of US chip sanctions, and is its claimed 1.5 µm wafer-on-wafer hybrid bonding a genuine departure from the 3D stacking already on the TSMC and Intel roadmaps or a relabelling of it?", "direction": "constructive on the Chinese CIDM model as a structure that can compound rather than stall under export controls, and on logic folding as a design change rather than a packaging change; cautious on the view that sanctions preserve a durable Western process edge, with the bottleneck reframed from lithography to bond pitch.", "tickers_covered": [ "TSM", "AMD", "INTC", "0981.HK (SMIC)", "with Huawei", "HiSilicon", "CXMT", "YMTC and SiCarrier as unlisted entities." ], "evidence_types": "He Tingbo's May 2026 IEEE ISCAS presentation and accompanying paper, Huawei product launch disclosures for Kirin 9050 Pro, published interconnect pitch figures for TSMC SoIC-X and HBM3E/HBM4, AMD advanced-packaging roadmap material, and named third-party scepticism.", "date": "10 September 2026, Part 1 of a two-part series, with Part 2 following on 18 September 2026." } }, "scoring-saas-survival-the-var-resistance-scorecard-pt-3-palo-alto-networks": { "slug": "scoring-saas-survival-the-var-resistance-scorecard-pt-3-palo-alto-networks", "title": "Scoring SaaS Survival: The VAR-Resistance Scorecard (Pt.3) — Palo Alto Networks", "url": "https://www.convequity.com/scoring-saas-survival-the-var-resistance-scorecard-pt-3-palo-alto-networks/", "published": "2026-09-06", "visibility": "tiers", "cohort": "opus", "model": "opus", "status": "corpus_draft_theme_json", "author": "Equity Research", "key_claims": [ "On 7 April 2026 Anthropic named twelve launch partners for Project Glasswing, its restricted defensive programme around Claude Mythos Preview, with Palo Alto Networks (PANW) and CrowdStrike (CRWD) the clearest pure-play cybersecurity seats alongside Amazon Web Services (AMZN), Apple (AAPL), Broadcom (AVGO), Cisco (CSCO), Alphabet (GOOGL), JPMorganChase (JPM), the Linux Foundation, Microsoft (MSFT), NVIDIA (NVDA) and Anthropic itself; access was extended to roughly forty further organisations for an initial cohort near fifty, and about 150 more organisations across fifteen-plus countries were added on 2 June 2026.", "Convequity's 6 September 2026 analysis reports that the binding constraint in the Glasswing programme fell on the fix side rather than the find side: partners surfaced more than ten thousand high- or critical-severity flaws in roughly a month, Anthropic's own open-source scanning produced on the order of twenty-three thousand candidates including about six thousand two hundred rated high or critical across more than a thousand projects, and Palo Alto Networks (PANW) itself shipped more than five times its usual patch volume in a single cycle.", "As of the 6 September 2026 analysis, Palo Alto Networks (PANW) had turned its Mythos-class access into product through Frontier AI Critical Defense, which pairs Unit 42 Exposure Analysis built on Mythos 5 with Unit 42 personnel, PANW estate telemetry and threat intelligence to rank exploitability and attack paths, and through network-level virtual patching that neutralises a flaw in the traffic path before every affected host finishes its change window.", "Palo Alto Networks (PANW) acquired CyberArk and has rebranded and extended that roughly USD 25 billion privileged-access franchise as Idira, positioned as of September 2026 as an identity platform spanning human, machine and agentic identities with entitlement discovery, zero standing privilege, just-in-time access and short-lived credentials, alongside the Chronosphere acquisition for high-scale observability and Prisma AIRS crossing USD 100 million in ARR on a short clock after general availability.", "Palo Alto Networks (PANW) has trained machine learning on live attack telemetry since WildFire launched around 2011, and the Precision AI stack described in the 6 September 2026 analysis runs thousands of detectors and well over two thousand machine-learning models in production with AgentiX trained on more than one billion real-world playbook executions, a materially different lineage from Datadog's (DDOG) Watchdog anomaly monitoring launched in 2018 and its later Toto domain time-series foundation model." ], "cite_as": "\"Scoring SaaS Survival: The VAR-Resistance Scorecard (Pt.3) — Palo Alto Networks,\" Convequity, September 6, 2026, https://www.convequity.com/scoring-saas-survival-the-var-resistance-scorecard-pt-3-palo-alto-networks/", "paywall_hook": { "question_answered": "which parts of Palo Alto Networks' franchise can a frontier model lab absorb once Mythos-class capability is generally available, and which parts stay structurally out of reach, given that the same capability that finds vulnerabilities at machine speed also changes what enterprises need to procure?", "direction": "constructive on the sensor, inline enforcement and machine-identity core, because those layers require deployed privileged infrastructure and liability for irreversible action that a model lab is poorly placed to underwrite; cautious on the checkable, out-of-band slice of the portfolio where scanning and adjudication work is most absorbable by coding agents; cautious on free-cash-flow margin percentage as token-heavy, outcome-priced delivery scales, with the argument that volume rather than margin carries the case; shares read as roughly fair against the base case rather than mispriced, with the asymmetry sitting in execution.", "tickers_covered": [ "PANW", "DDOG", "CRWD", "PLTR", "GTLB", "TENB", "QLYS", "CSCO", "MSFT", "GOOGL", "AMZN", "NET", "OKTA", "NVDA." ], "evidence_types": "company disclosures and acquisition history, dated Anthropic programme announcements and partner rosters, published vulnerability and patch-volume counts, product and architecture facts across NGFW, SASE, Cortex, Prisma AIRS and Idira, federal authorisation levels, the Snyk decline as a worked comparison, and Convequity's fourteen-factor VAR-Resistance scorecard with a discounted cash flow valuation.", "date": "6 September 2026, Part 3 of the VAR-Resistance Scorecard series, following the 22-name Part 1 coverage set and the Part 2 Datadog treatment.", "related_product": "Convequity's SaaS VAR-Resistance Scorecard and the DCF Models tab in Client Home (product facts provisional, per Gate 4)." } }, "the-ai-bubble-that-isnt-yet-what-the-completed-q226-data-reveals": { "slug": "the-ai-bubble-that-isnt-yet-what-the-completed-q226-data-reveals", "title": "The AI Bubble That Isn’t — Yet: What the Completed Q2’26 Data Reveals", "url": "https://www.convequity.com/the-ai-bubble-that-isnt-yet-what-the-completed-q226-data-reveals/", "published": "2026-08-30", "visibility": "tiers", "cohort": "opus", "model": "opus", "status": "corpus_draft_theme_json", "author": "Equity Research", "key_claims": [ "IREN (IREN) reported its fiscal fourth-quarter results on 27 August 2026, an event that closed out the calendar Q2 2026 reporting cycle for the hyperscaler, neo-cloud and model-lab universe tracked in Convequity's 30 August 2026 analysis.", "Anthropic's annualised revenue run rate reportedly rose from approximately USD 9 billion at the end of 2025 to more than USD 65 billion by July 2026, and its investors expect the company to finish 2026 between USD 100 billion and USD 120 billion, a figure the source flags as a forecast subject to differing run-rate methodologies.", "OpenAI's annualised revenue run rate reportedly doubled from approximately USD 20 billion to USD 40 billion over a comparable period ending in mid-2026, according to the reported figures cited in Convequity's 30 August 2026 analysis.", "Open-weight models rose from approximately 28.4% to 62% of token volume on Vercel's AI Gateway between 24 June 2026 and 22 August 2026, which the analysis reads as demand broadening across models and price points rather than rotating between closed providers.", "Convequity's 30 August 2026 analysis describes the mechanism by which Microsoft (MSFT) accelerates AI capacity through third-party commitments with IREN (IREN), Nebius (NBIS) and other neo-clouds: the partner raises the capital, builds the site, buys the accelerators and owns the assets, while Microsoft assumes a multi-year contractual obligation to pay for that capacity, so construction and GPU residual-value risk sit with the partner and Microsoft's cost arrives as lease and service operating expense rather than depreciation on owned data centres." ], "cite_as": "\"The AI Bubble That Isn’t — Yet: What the Completed Q2’26 Data Reveals,\" Convequity, August 30, 2026, https://www.convequity.com/the-ai-bubble-that-isnt-yet-what-the-completed-q226-data-reveals/", "paywall_hook": { "question_answered": "with the calendar Q2 2026 reporting cycle complete as of IREN's 27 August 2026 results, is the AI investment cycle a broad bubble, and where inside the complex is the genuine heat concentrated?", "direction": "not a broad bubble yet, constructive on the aggregate picture because returns on invested capital still sit above funding costs and forward valuation multiples remain below their 15 May 2026 starting points, cautious specifically on off-balance-sheet financing structures where contracted obligations could outrun realised utilisation, and neutral on the week's valuation bounce, read as a rebound from an overextended selloff rather than a structural re-rating.", "tickers_covered": [ "MSFT", "GOOGL", "META", "AMZN", "ORCL", "IREN", "NBIS", "APLD", "CRWV", "plus private model laboratories Anthropic and OpenAI." ], "evidence_types": "completed quarterly earnings releases and capex disclosures, remaining performance obligation and backlog disclosures, balance sheet and footnote detail on finance and operating leases, uncommenced leases, construction obligations and purchase commitments, third-party token-share data from Vercel's AI Gateway, reported model-lab revenue run rates, and Convequity's ABBX index readings across valuation, profitability, financing and capacity.", "date": "30 August 2026, covering the completed calendar Q2 2026 ABBX cycle.", "related_product": "Convequity's AI Bubble Barometer (ABBX), which tracks valuations, return on invested capital versus cost of capital, expanded debt and capex, and forward infrastructure economics across hyperscalers, neo-clouds and model laboratories (product facts provisional, per Gate 4)." } }, "cerebras-is-a-taxi-nvidia-runs-the-bus-network-pt1": { "slug": "cerebras-is-a-taxi-nvidia-runs-the-bus-network-pt1", "title": "Cerebras Is a Taxi. NVIDIA Runs the Bus Network (Pt1.)", "url": "https://www.convequity.com/cerebras-is-a-taxi-nvidia-runs-the-bus-network-pt1/", "published": "2026-08-25", "visibility": "tiers", "cohort": "opus", "model": "opus", "status": "corpus_draft_theme_json", "author": "Equity Research", "key_claims": [ "Cerebras Systems (CBRS) went public in May 2026 with a Wafer-Scale Engine architecture that was originally designed for model training, and which Convequity's 25 August 2026 report describes as having been redirected to inference once frontier model sizes outgrew its limited cross-wafer scalability.", "Cerebras (CBRS) repositioned as a full inference-system supplier roughly six months after announcing a partnership with Qualcomm (QCOM) under which Cerebras was to supply training racks and Qualcomm inference racks, a sequence Convequity dates and traces in its 25 August 2026 analysis.", "The Cerebras WSE-3 is manufactured as 84 reticle-sized fields on a single 300mm wafer, co-matured with TSMC (TSM) and linked across the silicon during fabrication so the wafer behaves as one logical processor, with spare compute and defect-routing fabric making the approach manufacturable; this sidesteps the roughly 800mm² lithography reticle limit and the finite reticle multiplier of advanced packaging that constrain conventional GPU dies (as of 2026).", "Convequity's 25 August 2026 report argues that process-node advances in power, performance, area and cost have continued for logic transistors while SRAM scaling effectively stopped at 7nm, so architectures that spend large silicon area on on-chip SRAM harvest less of the benefit of moving to 3nm and below than compute-dense designs do.", "In the same 25 August 2026 report, Convequity's equity research desk sets out a constructive two to three year view on Cerebras (CBRS) that is conditional on demand growing for premium-priced, high-speed, low-batch inference rather than predominantly low-cost, high-volume token generation, where NVIDIA (NVDA) class GPU systems hold the economics." ], "cite_as": "\"Cerebras Is a Taxi. NVIDIA Runs the Bus Network (Pt1.),\" Convequity, August 25, 2026, https://www.convequity.com/cerebras-is-a-taxi-nvidia-runs-the-bus-network-pt1/.", "paywall_hook": { "question_answered": "does the fastest AI inference system make the best mass-market token factory, and what has to be true about demand for Cerebras wafer-scale compute to be worth owning over the next two to three years?", "direction": "constructive on Cerebras (CBRS) over a two to three year horizon, conditional on the demand mix shifting toward premium-priced high-speed inference; the bottleneck that decides the outcome moves from memory bandwidth in decode to concurrency and cluster scalability, which is where NVIDIA (NVDA) remains the stronger economic proposition.", "tickers_covered": [ "CBRS", "NVDA", "QCOM", "GOOGL", "AVGO", "TSM." ], "evidence_types": "first-principles chip analysis at the matrix-multiply unit, reticle and advanced-packaging manufacturing limits, SRAM versus HBM memory-architecture economics and process-node scaling behaviour, prefill and decode serving characteristics, company history from Convequity's State of GenAI 2023 coverage, and rack-scale reference comparisons.", "date": "25 August 2026. Part 1 of two; the advanced serving math is carried in Part 2.", "related_product": "Convequity's semiconductor and AI compute research stream, including the State of GenAI report series with a 2026 edition in progress (product facts provisional, per Gate 4)." } }, "scoring-saas-survival-the-var-resistance-scorecard-pt-2-palantir-datadog-cloudflare": { "slug": "scoring-saas-survival-the-var-resistance-scorecard-pt-2-palantir-datadog-cloudflare", "title": "Scoring SaaS Survival: The VAR-Resistance Scorecard (Pt.2) — Palantir, Datadog & Cloudflare", "url": "https://www.convequity.com/scoring-saas-survival-the-var-resistance-scorecard-pt-2-palantir-datadog-cloudflare/", "published": "2026-08-15", "visibility": "tiers", "cohort": "opus", "model": "opus", "status": "corpus_draft_theme_json", "author": "Equity Research", "key_claims": [ "Convequity's 15 August 2026 analysis describes Palantir's (PLTR) Foundry Ontology as an action-oriented digital twin built from datasets, models, object types, properties, links and action types over first-party data that may still reside in Snowflake, SAP and Oracle systems, with the Ontology SDK letting developers treat Foundry as a backend with high-scale query, writeback and governance, and with roughly 200 SaaS-equivalent product surfaces running on the single substrate.", "The same 15 August 2026 analysis states that Palantir's (PLTR) partnership with NVIDIA (NVDA) around Nemotron open models is described in public materials as an engine for running those models in sovereign and air-gapped environments with customer ownership of self-improving, mission-specific weights, where in-platform telemetry and evaluations feed post-training, and the analysis notes commercial and government execution of that partnership is still unproven at scale as of publication.", "As of the 15 August 2026 analysis, Datadog's (DDOG) architecture collects metrics, logs, traces and security signals from software agents installed on customer hosts, virtual machines and bare metal into a unified telemetry plane, exposes a first-party agentic interface through Bits AI across Detect, Investigate and Remediate, and makes that same plane available to third-party agents including Cursor, Claude Code, Codex and the AWS DevOps Agent through its generally available MCP Server without use of Datadog's own user interface.", "The 15 August 2026 analysis records that Datadog (DDOG) built and released Toto and Toto 2.0, time-series foundation models pretrained on large-scale observability metrics and published with open weights, that Bits products in production mainly call frontier large language models with Datadog telemetry, topology, runbooks and investigation memory as context rather than a single in-house chat model, and that Datadog's acquisition of Adaptive ML brought an RLOps platform into Datadog AI Research for agentic post-training on infrastructure and security signals.", "Convequity's 15 August 2026 analysis reports that Cloudflare (NET) runs H100- and H200-class GPUs on Workers AI behind its proprietary Infire inference engine at a configuration of two H100s per server across a network of more than 300 points of presence, versus hyperscaler AI nodes that commonly pack eight higher-power SXM GPUs per machine, with roughly 700 W per SXM H100 against roughly 350 W per H100 PCIe or NVL card and colocation traffic racks typically in the 10 to 15 kW range against liquid-cooled AI racks at tens of kW and above." ], "cite_as": "\"Scoring SaaS Survival: The VAR-Resistance Scorecard (Pt.2) — Palantir, Datadog & Cloudflare,\" Convequity, August 15, 2026, https://www.convequity.com/scoring-saas-survival-the-var-resistance-scorecard-pt-2-palantir-datadog-cloudflare/", "paywall_hook": { "question_answered": "which SaaS businesses keep the work as agentic AI spreads, which survive headless as workflow and data plumbing under someone else's agent, and which get hollowed out, applied name by name to Palantir, Datadog and Cloudflare in the second instalment of the VAR-Resistance series.", "direction": "constructive on Palantir as the rare application vendor positioned to own the experience, workflow and data layers end to end, with the open question being how much of that success the share price already assumes; constructive on Datadog and Cloudflare as infrastructure that agents need and pay for by volume rather than as full agent platforms, cautious on Cloudflare's own AI hosting ambitions given edge power and cooling limits, and cautious on how much agentic further appreciation is already embedded in Cloudflare's multiple; the earlier read that production telemetry is effectively unabsorbable is softened, because coding agents give model labs a genuine incentive to push into observability.", "tickers_covered": [ "PLTR", "DDOG", "NET", "with NVDA", "IREN", "NBIS and CRWV referenced as counterparties or comparators." ], "evidence_types": "product and architecture disclosures (Ontology, OSDK, MetaConstellation, Bits AI, MCP Server, Toto time-series models, Workers AI and Infire), named partnerships and acquisitions, published engineering posts and benchmarks, company commentary on agent traffic and developer counts, GPU power and rack density figures, and Convequity's twelve-criterion VAR-Resistance scoring across data gravity, workflow absorbability, demand elasticity, route quality, model-layer leverage, pricing agility, governance, margin trajectory and implied scenario.", "date": "15 August 2026, Part 2 of the VAR-Resistance series, updating the provisional Part 1 scores for three names.", "related_product": "Convequity's SaaS Scorecard, the interactive VAR-Resistance framework covering the SaaS universe, with remaining names updated in later parts of the series (product facts provisional)." } }, "scoring-saas-survival-the-var-resistance-scorecard-pt-1": { "slug": "scoring-saas-survival-the-var-resistance-scorecard-pt-1", "title": "Scoring SaaS Survival: The VAR-Resistance Scorecard (Pt.1)", "url": "https://www.convequity.com/scoring-saas-survival-the-var-resistance-scorecard-pt-1/", "published": "2026-08-10", "visibility": "tiers", "cohort": "opus", "model": "opus", "status": "corpus_draft_theme_json", "author": "Equity Research", "key_claims": [ "In a 10 August 2026 report, Convequity extends its VARification thesis by taking NVIDIA (NVDA) chief executive Jensen Huang's claim that software companies become \"value-added resellers of AI tokens\" literally, noting that value-added resellers and systems integrators have traded at roughly 10 to 15 times earnings and 0.5 to 1.5 times sales while SaaS has historically commanded 15 to 20 times sales.", "As of the 10 August 2026 report, Salesforce (CRM), ServiceNow (NOW) and Monday.com (MNDY) were offering headless versions of their software in which the conventional user interface is no longer required and workflows can be instructed through a third-party agentic interface, which Convequity reads as evidence that headless is the most common near-term evolution of SaaS under agentic AI because agents avoid token-hungry browser and computer-use tool calls.", "The 10 August 2026 report cites Microsoft (MSFT) chief executive Satya Nadella's Reverse Information Paradox, the argument that enterprises pay twice when they feed proprietary workflows into third-party AI tools, once to build the systems that encode institutional knowledge and again when that knowledge refines external models that can later replicate those systems, with AI sovereignty through private, isolated and interoperable systems as the countermeasure.", "The 10 August 2026 report reads Palo Alto Networks' (PANW) acquisitions of Chronosphere and CyberArk as positioning for agent monitoring and machine identity, noting that Chronosphere is built on the open-source M3 database, specialises in Kubernetes observability, and reduces log data volume by roughly 99% while preserving most useful context, which matters because unfiltered telemetry fills agent context windows with noise and raises token cost.", "The VAR-Resistance Scorecard published on 10 August 2026 scores 22 public SaaS names from 1 to 5 across three pillars, Structural Position at roughly 40% weight, Adaptive Capacity at roughly 35% and Valuation Asymmetry at roughly 25%, with methodology created by Convequity and the scoring itself executed collaboratively by Grok 4.5, Fable 5 and DeepSeek v4 Pro, and with the valuation pillar built on Bessemer Venture Partners' Rule of X, which values a point of forward growth at roughly 2.3 times a point of free cash flow margin." ], "cite_as": "\"Scoring SaaS Survival: The VAR-Resistance Scorecard (Pt.1),\" Convequity, August 10, 2026, https://www.convequity.com/scoring-saas-survival-the-var-resistance-scorecard-pt-1/", "paywall_hook": { "question_answered": "of the three routes SaaS can take under agentic AI, full in-house replacement, headless survival beneath someone else's experience layer, or the incumbent owning the agentic layer itself, which route does each of 22 public software names realistically face, is headless a stable endpoint or a waypoint, and where has the market already priced a worse route than the fundamentals support?", "direction": "cautious to negative on seat-based application software serving bounded back-office functions, where outcome pricing protects revenue but cannot unlock growth, and on low-moat vendors whose main asset is a polished interface; constructive on the layers the expanding model layer needs rather than absorbs, namely analytical data infrastructure, governance, observability, identity and security, and on consumption-priced plumbing serving volume-elastic, revenue-generating work; the report argues operational margin pressure is milder than feared while valuation compression is the real mechanism, and treats headless as a starting point whose boundary keeps moving.", "tickers_covered": [ "PLTR", "CRWD", "DDOG", "NET", "SNOW", "SHOP", "CRM", "GTLB", "HUBS", "MNDY", "TWLO", "NOW", "FIG", "ADBE", "OKTA", "TEAM", "SAP", "VEEV", "DOCU", "WDAY", "INTU", "ZM", "with PANW", "MSFT and NVDA referenced in the argument." ], "evidence_types": "public product and strategy disclosures including headless offerings from named incumbents, dated acquisitions and their technical rationale, executive statements from NVIDIA and Microsoft leadership, an explicit three-pillar scoring framework with published factor definitions and weights, and a valuation grid built on the Bessemer Rule of X.", "date": "10 August 2026, Part 1 of a series, with scores and bucket assignments described as a living assessment to be revised through manual company-level review in later parts.", "related_product": "the VAR-Resistance Scorecard, described as moving to a regularly updated feature on the Convequity site, with the underlying scores distributed as a Google Sheet to subscribers (product facts provisional, per Gate 4)." } }, "portfolio-rebalance-capturing-the-next-phase-of-ai-infrastructure-buildout-pt-3": { "slug": "portfolio-rebalance-capturing-the-next-phase-of-ai-infrastructure-buildout-pt-3", "title": "Portfolio Rebalance: Capturing the Next Phase of AI Infrastructure Buildout (Pt.3)", "url": "https://www.convequity.com/portfolio-rebalance-capturing-the-next-phase-of-ai-infrastructure-buildout-pt-3/", "published": "2026-08-04", "visibility": "tiers", "cohort": "opus", "model": "opus", "status": "corpus_draft_theme_json", "author": "Equity Research", "key_claims": [ "POET Technologies (POET) announced a strategic partnership with Lumilens in May 2026 centred on a jointly developed Electrical-Optical Interposer that carries both discrete InP/EML engines and silicon-photonics PICs on the same wafer-scale assembly base, with an initial USD 50 million purchase order and a stated potential cumulative value above USD 500 million over five years across 800G and 1.6T pluggables through to near-packaged and co-packaged optics.", "In June 2026 Rubrik (RBRK) announced the next stage of Annapurna, positioning it as an AI-ready unstructured-data layer that discovers, indexes and catalogs data in place across NAS, S3 and object stores and then publishes a queryable catalog into a lakehouse without traditional ETL and full-estate duplication, alongside a broader framing around Rubrik Agent Cloud and semantic governance of autonomous agents.", "As of Convequity's 4 August 2026 analysis, Ciena's (CIEN) WaveLogic 6 Extreme remained the industry's only shipping 1.6 Tb/s coherent solution, and the company was pushing coherent capability closer to the compute layer through Coherent Lite and the Nubis co-packaged optics assets, with coherent's tolerance of high insertion loss becoming relevant inside the data centre for silicon-photonics optical circuit switches as well as for campus-scale and multi-site reach.", "Convequity's 4 August 2026 analysis records gas-turbine lead times of roughly three to five years, with turbine manufacturers sold out into 2029 to 2031, against Baker Hughes gas-directed rig counts running only in the low-to-mid 120s at the time of writing, a timing gap the analysis reads as the reason behind-the-wellhead data-centre siting has not yet shown up in near-term United States gas balances.", "Convequity's 4 August 2026 analysis argues that NVIDIA's (NVDA) roughly USD 2 billion investment in Marvell (MRVL) is better read as supply-chain leverage than endorsement, because Marvell lacks ready access to leading-edge 224G SerDes IP, is one of the few alternatives to Broadcom (AVGO) in optical DSP, and had been designated the primary switch supplier for UALink, the scale-up standard competing with NVLink." ], "cite_as": "\"Portfolio Rebalance: Capturing the Next Phase of AI Infrastructure Buildout (Pt.3),\" Convequity, August 4, 2026, https://www.convequity.com/portfolio-rebalance-capturing-the-next-phase-of-ai-infrastructure-buildout-pt-3/", "paywall_hook": { "question_answered": "as the AI infrastructure buildout moves from technological promise to commercial deployment, which layers of the optical, memory, test and energy stack convert technical relevance into durable value capture, and where does architectural uncertainty between CPO, NPO and conventional pluggables argue for exposure that wins across outcomes rather than on a single path?", "direction": "constructive on bottleneck resolvers, specifically the engineered-wafer substrate layer beneath all silicon-photonics paths, coherent scale-across connectivity, and the test, inspection and metrology layer that becomes more critical as packages densify; cautious on bottleneck causers whose capacity response is constrained and whose expectations already run high; negative on software-orchestration and design-interface layers whose economics are being commoditised by compute-native entrants and coding agents; pre-revenue optical platforms treated as deliberately small optionality rather than base case.", "tickers_covered": [ "SOI", "CIEN", "POET", "LWLG", "HIMX", "AXTI", "AIXA", "SMTC", "SMHN", "KEYS", "SNDK", "RBRK", "TEM", "CRSP", "FIG", "IOT", "AR", "EXE", "CRK", "MRVL", "NVDA." ], "evidence_types": "dated company announcements and partnerships, product and architecture facts across coherent optics, optical interposers, electro-optic polymer modulators, MOCVD epitaxy and InP substrates, industry supply data including Baker Hughes rig counts and turbine lead times, and Convequity's own rebalance framework distinguishing technical relevance from commercial readiness and value capture.", "date": "4 August 2026, third and final part of the 3Q26 rebalance series; the underlying rebalance decisions were taken in the final week of June 2026.", "related_product": "Convequity's Core AI Portfolio, an unleveraged book of roughly 80 names mapped across 11 segments of the AI value chain (referenced without link, and no positions or weights are disclosed here)." } }, "jensens-polite-warning-the-coming-varification-of-software": { "slug": "jensens-polite-warning-the-coming-varification-of-software", "title": "Jensen's Polite Warning: The Coming VARification of Software", "url": "https://www.convequity.com/jensens-polite-warning-the-coming-varification-of-software/", "published": "2026-07-30", "visibility": "public", "cohort": "opus", "model": "opus", "status": "corpus_draft_theme_json", "author": "Equity Research", "key_claims": [ "In a 30 July 2026 analysis, Convequity argued that NVIDIA (NVDA) chief executive Jensen Huang's prediction that software companies will become \"value-added resellers of AI tokens\" implies a valuation compression of more than 90%, because value-added resellers and systems integrators trade at roughly 10 to 15 times earnings and 0.5 to 1.5 times sales while SaaS companies have historically traded at 15 to 20 times sales.", "The same 30 July 2026 analysis notes that Wix (WIX) already trades closer to value-added reseller multiples than to classic SaaS multiples, which it reads as the market pricing the shift before the narrative describes it.", "Convequity's 30 July 2026 analysis argues that the traditional one dollar of build to three dollars of maintenance rule that justified buying software over building it is collapsing, because maintenance work is repetitive and well specified and is therefore the part AI agents automate first, which reframes Kirkland & Ellis's reported USD 500 million internal AI programme as a defence of core competency rather than an IT procurement decision.", "In the 30 July 2026 analysis, Convequity describes foundation model labs as taking the high-margin cognitive layer of the software value chain and pushing application vendors down into the coordination layer, leaving those vendors the low-margin, relationship-heavy, non-scalable work that software vendors themselves once handed to value-added resellers and systems integrators.", "Convequity's 30 July 2026 analysis uses the 2016 prediction that deep learning would make radiologists obsolete, followed a decade later by radiologist shortages, to argue that AI expands legal demand through cheaper delivery rather than removing the profession, while leaving application companies three strategic options: post-train their own models, move up into the model layer, or accept value-added reseller economics." ], "cite_as": "\"Jensen's Polite Warning: The Coming VARification of Software,\" Convequity, July 30, 2026, https://www.convequity.com/jensens-polite-warning-the-coming-varification-of-software/.", "paywall_hook": null }, "photonics-ocs-roadmaps-and-market-dynamics-pt-4": { "slug": "photonics-ocs-roadmaps-and-market-dynamics-pt-4", "title": "Photonics: OCS Roadmaps and Market Dynamics (Pt.4)", "url": "https://www.convequity.com/photonics-ocs-roadmaps-and-market-dynamics-pt-4/", "published": "2026-07-24", "visibility": "tiers", "cohort": "opus", "model": "opus", "status": "corpus_draft_theme_json", "author": "Equity Research", "key_claims": [ "Convequity's 24 July 2026 analysis reports that Google's eighth-generation TPU separates training and inference for the first time, with the V8t training cluster keeping a 3D torus of 9,600 chips per pod served by 300-port optical circuit switches at roughly 200 TPUs per unit, while the V8i inference cluster uses a flat two-layer topology of 1,152 chips per pod served by 64-port switches, cutting network hops from 16 to 7, improving latency for communication-intensive workloads by up to 50% and shortening mixture-of-experts inter-layer transformation time by a factor of five.", "The same 24 July 2026 analysis describes Google's 2.4T Coherent Lite module as replacing the 8-transmit and 8-receive, 16-fibre arrangement of a conventional 1.6T DR8 with 2 transmit and 2 receive channels carrying 1.2T per fibre, so that with circulators and Z-block multiplexers a single fibre occupies one switch port, with Innolight, Coherent (COHR) and Lumentum (LITE) all developing the modules, volume delivery expected from the second half of 2027 and large-scale adoption in 2028.", "According to the 24 July 2026 analysis, NVIDIA's (NVDA) Feynman architecture targeted for 2028 plans a dual-layer optical interconnect around the NVL576 rack, with co-packaged optics inside for high-density short-reach links and optical circuit switching outside for non-blocking cluster topologies at the scale of 100,000 GPUs, using mature MEMS switches first and moving to silicon photonics only once insertion loss falls far enough that the optical amplifier burden becomes acceptable, while Huawei's Ascend 960 is reported to use an all near-packaged-optics design and Oracle (ORCL) is identified as the second-largest buyer of MEMS optical circuit switches behind Google.", "The 24 July 2026 analysis puts Google's 2026 optical circuit switch procurement at roughly 15,000 units, with Calient supplying about 9,000 of them, Lumentum (LITE) about 4,000 alongside more than USD 400 million of backlog, Coherent (COHR) about 1,400 LCoS-based units, and Accelink the only Chinese supplier sampling silicon photonics switches in volume, having delivered around 300 validated 32-port units to Google at roughly USD 57,000 each against about USD 32,000 of material cost, with 64-port units priced near USD 100,000 in validation and 128-port samples expected in 2027.", "Citing Cignal AI, the 24 July 2026 analysis projects the optical circuit switching market to grow from roughly USD 400 million in 2025 to more than USD 2.5 billion by 2029, a 58% compound annual growth rate, and separately assesses microLED interconnect, drawn from a Microsoft research paper and pursued by the startup Avicena, as claiming under 1 picojoule per bit against roughly 5 for current co-packaged optics and more than 15 for pluggable optics by running a thousand or more gallium nitride lanes at 4 to 10G into plain silicon photodetectors, while LED emission physics confines it to about 10 metres and single-digit gigahertz modulation." ], "cite_as": "\"Photonics: OCS Roadmaps and Market Dynamics (Pt.4),\" Convequity, July 24, 2026, https://www.convequity.com/photonics-ocs-roadmaps-and-market-dynamics-pt-4/", "paywall_hook": { "question_answered": "as optical circuit switching roadmaps split by workload into 2028, which layer of the network each switching technology can physically serve, who supplies the switches and modules through the 2026 ramp, and whether co-packaged optics and microLED alter the demand path for indium phosphide lasers?", "direction": "constructive on silicon photonics optical circuit switching at the scale-up and inference layer and on the silicon photonics and hybrid-bonding content that higher-bandwidth coherent designs pull into each transceiver; constructive on MEMS retaining the high-radix scale-out spine rather than being displaced; co-packaged optics is judged not finished but dependent on multiple technology and supplier paths rather than a single route; cautious on aggressive indium phosphide laser forecasts that assume lasers win at every layer including the shortest reaches, and cautious on the highest-power semiconductor optical amplifier route given single-supplier dependency.", "tickers_covered": [ "GOOGL", "NVDA", "ORCL", "LITE", "COHR", "MRVL", "CIEN", "TSEM", "plus Calient", "Polatis under HUBER+SUHNER", "Accelink", "Innolight", "Eoptolink", "Soitec", "Avicena and Huawei." ], "evidence_types": "reported architecture and pod topology details for Google's V8t and V8i and NVIDIA's Feynman generation, module channel and fibre counts, supply-chain unit volumes and backlog figures, port-level pricing and material cost, yield and insertion-loss bottleneck analysis, third-party market sizing from Cignal AI, and a physics audit of emission and modulation limits.", "date": "24 July 2026, Part 4 of the photonics series, following the Part 3 treatment of high-radix MEMS switching." } }, "portfolio-rebalance-capturing-the-next-phase-of-ai-infrastructure-buildout-pt-2": { "slug": "portfolio-rebalance-capturing-the-next-phase-of-ai-infrastructure-buildout-pt-2", "title": "Portfolio Rebalance: Capturing the Next Phase of AI Infrastructure Buildout (Pt.2)", "url": "https://www.convequity.com/portfolio-rebalance-capturing-the-next-phase-of-ai-infrastructure-buildout-pt-2/", "published": "2026-07-21", "visibility": "tiers", "cohort": "opus", "model": "opus", "status": "corpus_draft_theme_json", "author": "Equity Research", "key_claims": [ "Enphase Energy (ENPH) launched its IQ solid-state transformer for AI data centres in April 2026, an early commercial-scale product operating at lower voltages rather than a full utility medium-voltage 13 kV unit, entering a segment SolarEdge (SEDG) had moved into earlier.", "In the quarter reported ahead of Convequity's 21 July 2026 rebalance note, SiTime (SITM) grew Communications, Enterprise and Data Center revenue 158% year over year, its eighth consecutive quarter of triple-digit growth, against total company revenue growth of 88%, while its pending acquisition of Renesas's timing business adds clock generators, jitter attenuators, network synchronisers and distribution buffers to a resonator and oscillator line.", "Intuitive Machines (LUNR) revenue rose from roughly USD 45 million in fiscal Q4 2025 to about USD 187 million in fiscal Q1 2026 as contracts converted, against a record USD 1.1 billion backlog, a sixth NASA CLPS task order valued at up to USD 148 million, a prior USD 180 million award, and prime roles operating the Lunar Reconnaissance Orbiter Camera and ShadowCam.", "As of the 21 July 2026 analysis, CoreWeave (CRWV) had spent approximately USD 7.7 billion of capex in fiscal Q1 2026 against a target of roughly 1.7 GW of capacity by year end, Nebius (NBIS) had spent roughly USD 2.5 billion in fiscal Q1 2026 against about 4 GW planned, and IREN Limited (IREN) operated about 0.15 GW against a planned 5.0 GW pipeline.", "Ahead of the 21 July 2026 analysis, Palantir (PLTR) announced a partnership with NVIDIA (NVDA) to deploy and post-train Nemotron open models in sovereign environments so customers can own self-improving models trained on their own data and telemetry, at a point when Salesforce, Databricks, Microsoft, Amazon Web Services, Google Cloud and Scale AI had all replicated variants of the forward deployed engineer model Palantir pioneered." ], "cite_as": "\"Portfolio Rebalance: Capturing the Next Phase of AI Infrastructure Buildout (Pt.2),\" Convequity, July 21, 2026, https://www.convequity.com/portfolio-rebalance-capturing-the-next-phase-of-ai-infrastructure-buildout-pt-2/", "paywall_hook": { "question_answered": "how should exposure across the AI infrastructure stack be re-underwritten in July 2026 as the binding constraint moves from GPU supply toward power delivery at 800 V DC, optical interconnect yield and available compute capacity, and which layers of that stack are already priced for the shift?", "direction": "constructive on bidirectional power-electronics and gallium nitride specialists positioned for the 800 V DC and solid-state transformer transition, cautious on pure or heavy silicon carbide names facing idle Western capacity and geopolitical overhang, and cautious on diversified incumbents whose scale dilutes the theme; within photonics, conviction is concentrated at the hybrid-bonding and high-speed analog bottleneck rather than in laser incumbency; within compute, preference runs to under-recognised capacity and asymmetric setups over the most fully valued execution leaders; constructive on the marquee AI software layer after its drawdown and on stablecoin infrastructure over trading-volume-dependent crypto exposure; negative on one grid-services name judged insufficiently scalable.", "tickers_covered": [ "ENPH", "SEDG", "NVTS", "IFX", "STM", "ON", "WOLF", "MPWR", "TSLA", "LITE", "TSEM", "MTSI", "SITM", "ASMI", "BESI", "PLTR", "COIN", "CRCL", "HOOD", "PWR", "IREN", "CRWV", "NBIS", "ORCL", "BTDR", "CIFR", "CORZ", "LUNR", "PL", "RDW", "RKLB", "GOOGL", "MP." ], "evidence_types": "product launch and architecture disclosures, quarterly revenue and capex figures, stated capacity roadmaps in gigawatts, contract awards and backlog, packaging and deposition process analysis, discounted cash flow work, valuation multiples on gross profit and free cash flow, and portfolio weighting decisions.", "date": "21 July 2026, Part 2 of the quarterly rebalance series, following the Part 1 rotation into bottleneck resolvers.", "related_product": "Convequity's AI Bubble Barometer (ABBX), whose Forward Valuation tab carries the implied revenue and implied gross profit multiples referenced for the neocloud and hyperscaler names (product facts provisional, per Gate 4)." } }, "photonics-deep-dive-into-ocs-architectures-pt-3": { "slug": "photonics-deep-dive-into-ocs-architectures-pt-3", "title": "Photonics: Deep Dive into OCS Architectures (Pt.3)", "url": "https://www.convequity.com/photonics-deep-dive-into-ocs-architectures-pt-3/", "published": "2026-07-13", "visibility": "tiers", "cohort": "opus", "model": "opus", "status": "corpus_draft_theme_json", "author": "Equity Research", "key_claims": [ "Alphabet's Google (GOOGL) pioneered optical circuit switching in its TPUv4 systems beginning in 2020 and had deployed it broadly across internal infrastructure by 2022, when its Apollo fabric used 136x136-port OCS units for TPUv4 pods of 4,096 TPUs, across four zones with 256 switches in total.", "As described in a 13 July 2026 Convequity analysis, Google's (GOOGL) TPUv7 Ironwood generation, launched in 2025, builds a maximum pod of 144 racks of 64 TPUs each for 9,216 TPUs, with 96 optical transceivers per rack requiring 13,824 OCS ports served by 48 units of 288x288 ports, while the datacenter network layer uses co-designed 300x300-port OCS units, 256 of them across 32 racks, to connect 16 pods into clusters of 147,456 TPUs.", "Google (GOOGL) has published results from its Jupiter optical datacenter network showing that optical circuit switching delivered 10% lower flow completion time, 30% higher throughput, 40% less power consumption, 30% lower cost and 50x less downtime than the best known alternatives, an outcome the 13 July 2026 analysis attributes to nearly a decade of co-design across the Orion SDN control plane, custom transceivers and circulators, and traffic patterns rather than to a drop-in switch swap.", "The 13 July 2026 analysis documents the physical case for co-packaged optics and its costs: a conventional pluggable places the optical engine roughly 30 cm from the switch ASIC and needs a DSP that pushes an 800G port to 15 to 25W, versus roughly 4 to 5W when the electrical link is shortened to millimetres, but CPO requires hybrid bonding of 16 to 32 photonic integrated circuits onto one switch die, where even 99% per-bond yield compounds to about 85.2% at 16 PICs and 72.5% at 32 PICs, and TSMC's (TSM) COUPE platform had already hit technical setbacks as of 2026.", "As of the 13 July 2026 analysis, co-packaged optics keeps the laser external at a 300mW-plus continuous-wave specification, roughly 3x the power of pluggable CW lasers and a part Lumentum (LITE) alone is described as mass-producing at that spec, while among the four OCS approaches Polatis, owned by HUBER+SUHNER, ships 576-class port counts at about 1 dB insertion loss with latching that holds state unpowered, and silicon photonics waveguide switching already reaches roughly 6 nanosecond switching at under 0.5W per port with no moving parts." ], "cite_as": "\"Photonics: Deep Dive into OCS Architectures (Pt.3),\" Convequity, July 13, 2026, https://www.convequity.com/photonics-deep-dive-into-ocs-architectures-pt-3/", "paywall_hook": { "question_answered": "if AI infrastructure's next architectural shift lands at the switch layer, should the fabric stay electrical and be retrofitted with co-packaged optics, or move to optical circuit switching, and which of the four OCS technologies has a path to carrying that load through the 2030s?", "direction": "cautious to negative on the near-term co-packaged optics narrative on manufacturing, supplier concentration and failure-domain grounds; constructive on near-packaged optics paired with optical circuit switching, on Google's OCS-based TPU architecture as a structural system-efficiency advantage, and on silicon photonics as the strongest long-run OCS roadmap, with MEMS holding the scheduled, high-radix portion of the market.", "tickers_covered": [ "NVDA", "GOOGL", "AVGO", "MRVL", "LITE", "TSM", "plus unlisted and non-US names including Huawei", "Calient", "Polatis under HUBER+SUHNER", "and DeepSeek and OpenAI as model-architecture reference points." ], "evidence_types": "published vendor and hyperscaler architecture disclosures, Google's own Jupiter and Apollo network results, port-count and transceiver-ratio arithmetic for TPU pods and clusters, device physics for MEMS, LCoS, piezo direct beam steering and silicon photonics switching, hybrid bonding yield compounding, laser and Indium Phosphide supply conditions, and mapping of parallelism strategies onto network topology.", "date": "13 July 2026, Part 3 of Convequity's photonics supply chain series, following Part 1 on CPO overhype and Part 2 on wafers, epitaxy and modulator materials." } }, "portfolio-rebalance-capturing-the-next-phase-of-ai-infrastructure-buildout-pt-1": { "slug": "portfolio-rebalance-capturing-the-next-phase-of-ai-infrastructure-buildout-pt-1", "title": "Portfolio Rebalance: Capturing the Next Phase of AI Infrastructure Buildout (Pt.1)", "url": "https://www.convequity.com/portfolio-rebalance-capturing-the-next-phase-of-ai-infrastructure-buildout-pt-1/", "published": "2026-07-06", "visibility": "tiers", "cohort": "opus", "model": "opus", "status": "corpus_draft_theme_json", "author": "Equity Research", "key_claims": [ "In a 6 July 2026 note, Convequity describes the advanced packaging substrate transition as moving from ABF to glass because ABF warps under the thermal stress and coefficient-of-thermal-expansion mismatch of larger, hotter multi-die packages, and identifies LPKF Laser & Electronics (LPKF) as the only company with mass-qualified equipment for high-density laser drilling of through-glass vias at production scale and as the leading Western equipment partner to TSMC, Intel and Samsung.", "The same 6 July 2026 note reports that TSMC has signalled it expects to begin mass production of glass substrates in 2027, roughly a two-year acceleration of its prior timeline, and that designs such as NVIDIA's Rubin Ultra four-die configuration are at risk because ABF cannot support warpage and interconnect reliability at those die areas.", "Cohu (COHU) secured multi-unit and follow-on production orders in 2026 for its Eclipse automated test equipment platform tied to next-generation AI datacentre processors, prompting full-year 2026 high-performance-computing revenue guidance of USD 80 to 100 million and an overall fiscal 2026 growth outlook of 20 to 25%, alongside separate orders for testing next-generation GaN devices destined for the 800V DC power distribution architecture being adopted in megawatt-scale AI racks; the logic and HPC test market it is entering is described as an Advantest and Teradyne duopoly holding around 90% share.", "As of the 6 July 2026 note, SK Hynix plans to roughly triple HBM output while stack heights climb from 8 to 12 to 16 layers, which multiplies inspection and metrology steps per line and benefits Onto Innovation (ONTO) as the incumbent at SK Hynix, Camtek (CAMT) with roughly 70% of revenue tied to advanced packaging including HBM micro-bump inspection and hybrid bonding and stronger positioning at Samsung, and Nova (NVMI) through converged metrology combining optical critical dimension, XPS and XRF X-ray techniques and physical profiling on a single platform. Onto Innovation (ONTO) has acquired a 27% strategic stake in Rigaku for approximately USD 710 million as part of an alliance pairing Rigaku's X-ray metrology with Onto's process control platform.", "The 6 July 2026 note states that Cerebras (CBRS), which listed via IPO, and Tenstorrent remain the two primary standalone alternative AI compute architectures following NVIDIA's absorption of Groq, SoftBank's acquisition of Graphcore and the likely acquisition of SambaNova by Intel, and describes Cerebras's wafer-scale integration with TSMC as connecting roughly 25 dies on a single wafer against NVIDIA's typical two compute dies, an SRAM-heavy design that trades higher cost per token and limited on-chip capacity for substantially faster token generation while remaining constrained by TSMC wafer supply growing at around 30% per year." ], "cite_as": "\"Portfolio Rebalance: Capturing the Next Phase of AI Infrastructure Buildout (Pt.1),\" Convequity, July 6, 2026, https://www.convequity.com/portfolio-rebalance-capturing-the-next-phase-of-ai-infrastructure-buildout-pt-1/", "paywall_hook": { "question_answered": "as the AI infrastructure buildout enters its next and more granular phase in mid-2026, which less-visible layers of the stack now carry the binding constraints, and how should exposure shift away from the first-phase winners whose expectations have already moved higher?", "direction": "constructive on capacity-rich foundry \"resolvers\" and on the back-end manufacturing complexity layer spanning glass substrates, test and handling, and metrology and inspection; constructive on power and space infrastructure including orbital compute as terrestrial power and grid limits bind; more cautious on continuing to crowd into the most obvious bottleneck creators after strong price appreciation; fundamentals read as not being in bubble territory even as sentiment is capable of pricing assets away from them in the short term; on Cerebras, the view rests on execution rather than on multiple compression.", "tickers_covered": [ "GFS", "UMC", "INTC", "LPKF", "COHU", "ONTO", "CAMT", "NVMI", "CBRS", "plus TSMC", "Samsung", "SK Hynix", "Advantest", "Teradyne", "KLA", "Rigaku", "Tenstorrent and SpaceX as unlisted or non-held reference names." ], "evidence_types": "company guidance and order disclosures, disclosed strategic stake transactions, stated customer capacity and roadmap plans, packaging and substrate engineering constraints, test and metrology step-count mechanics, channel commentary on design risk, and Convequity's own rebalancing framework and valuation lens.", "date": "6 July 2026, part 1 of the rebalance series, reflecting a rebalancing meeting held the prior week and theses stress-tested against first-half 2026 developments.", "related_product": "Convequity's AI Bubble Barometer (ABBX), whose extended version is referenced as showing aggregate forward ROIC and EV/IC readings (product facts provisional, per Gate 4)." } }, "notes-the-bottleneck-heuristic-the-memory-supercycle-power-bottleneck-and-data-center-value-creation-pt-3": { "slug": "notes-the-bottleneck-heuristic-the-memory-supercycle-power-bottleneck-and-data-center-value-creation-pt-3", "title": "Notes: The Bottleneck Heuristic — The Memory Supercycle, Power Bottleneck, and Data Center Value Creation (Pt.3)", "url": "https://www.convequity.com/notes-the-bottleneck-heuristic-the-memory-supercycle-power-bottleneck-and-data-center-value-creation-pt-3/", "published": "2026-06-29", "visibility": "tiers", "cohort": "opus", "model": "opus", "status": "corpus_draft_theme_json", "author": "Equity Research", "key_claims": [ "In its 29 June 2026 note, Convequity describes AI systems moving to a multi-tier memory hierarchy in which HBM holds active model weights and the current attention window, a large external DRAM pool attached over CXL holds the broader KV cache, and NAND-based NVMe SSDs hold colder context and checkpoints, and it cites NVIDIA's (NVDA) forthcoming Vera Rubin NVL72 rack as specifying roughly 20.7 TB of HBM, 54 TB of DRAM and 1,152 TB of NAND per rack.", "By mid-2026 the three incumbent memory producers had begun a measured capacity response, with Samsung targeting roughly a 50% increase in HBM production capacity during the year, SK Hynix accelerating its HBM4 ramp through the new M15X fab while reportedly redirecting some planned HBM4 capacity toward DDR5, and Micron adding DRAM and HBM output in Taiwan alongside longer-term greenfield commitments in the United States and Singapore whose meaningful incremental wafer starts are not expected until 2027 and beyond.", "YMTC put hybrid bonding for NAND into volume production under its proprietary Xtacking process in 2021 and released the world's first 232-layer 3D NAND in August 2022, ahead of Samsung, SK Hynix, Western Digital, Micron and Kioxia, then announced in late 2024 that it had reconstituted a fully Chinese-domestic production line after United States sanctions forced an immediate cut-off of US-person staff and of equipment, spares and field service from Applied Materials (AMAT), Lam Research (LRCX) and KLA. YMTC held roughly 13% of global NAND shipment share in the third quarter of 2025 against Micron's 16% on Counterpoint data cited in the note, broke ground on its Phase 3 Wuhan fab in September 2025, completed construction in March 2026, finished equipment installation in April 2026, and is scheduled to bring that 100,000 wafer per month line, 49.9% owned by the Hubei provincial government, into production in the second half of 2026.", "The note places the next printed-circuit-board constraint in the base material layer, with M8 resin already used in Google's TPUv7 server boards and treated as the minimum viable specification for 112 Gbps PAM4 signalling, and successor M9 resin targeting a dissipation factor below 0.0025 for the 224 Gbps PAM4 links required by NVLink 6 and 1.6T Ethernet, with widespread M9 adoption expected across 2026 and 2027 and supply of both M9 resin and the quartz cloth reinforcement concentrated among a handful of specialty materials suppliers. As of the same 29 June 2026 note, TDK and Murata, which the note estimates jointly hold 60 to 70% of the high-performance MLCC market, had both announced price increases of more than 35% for advanced specifications.", "As of 29 June 2026 the note puts total United States generation capacity at approximately 1,250 GW on a nameplate basis with net annual additions of roughly 25 to 50 GW, states that around 70% of US power transformers are more than 25 to 30 years old, and records lead times for advanced H-class combined cycle gas turbines from GE Vernova (GEV), Siemens Energy and Mitsubishi Heavy Industries stretching from two to three years out to more than three to four years. It also notes that Oracle (ORCL) has secured approximately 10 GW of power commitments through a mix of utility power purchase agreements and behind-the-meter generation, and that former bitcoin miners including Core Scientific (CORZ), Bitdeer (BTDR) and Cipher Mining (CIFR) are converting existing grid interconnections, land and power infrastructure into AI hosting capacity." ], "cite_as": "\"Notes: The Bottleneck Heuristic — The Memory Supercycle, Power Bottleneck, and Data Center Value Creation (Pt.3),\" Convequity, June 29, 2026, https://www.convequity.com/notes-the-bottleneck-heuristic-the-memory-supercycle-power-bottleneck-and-data-center-value-creation-pt-3/", "paywall_hook": { "question_answered": "after GPUs, CoWoS packaging and cooling were relieved, where does the binding constraint in the AI buildout actually sit in mid-2026, and which suppliers can relieve memory, advanced PCB materials, power delivery and power generation when others cannot?", "direction": "constructive on the durability of the memory shortage over the next several quarters and on the narrow set of vendors that can physically relieve it, including Chinese domestic NAND supply as an unintended consequence of export controls; cautious on the sustainability of current memory-industry margins once incremental supply arrives at scale; constructive on pure-play AI data centre operators and neo clouds extracting more value per GW than legacy-burdened hyperscalers; power generation framed as the bottleneck now forming rather than the one now peaking.", "tickers_covered": [ "NVDA", "AMAT", "LRCX", "MPWR", "GEV", "CGEH", "ORCL", "CORZ", "BTDR", "CIFR", "CRWV", "NBIS", "SNDK", "TSEM", "AVGO", "MSFT", "VRT", "plus unlisted or non-US names including YMTC", "CXMT", "Samsung", "SK Hynix", "Micron", "Kioxia", "TDK", "Murata", "Siemens Energy", "Mitsubishi Heavy Industries and TTM Technologies." ], "evidence_types": "memory pricing and capacity history across the 2021 to 2026 cycle, producer capex and fab announcements, Counterpoint market share data, fab construction and equipment-installation timelines, United States export control actions including the February and June 2026 Section 1260H list episode, rack-level bill-of-materials specifications, PCB resin and layer-count specifications, MLCC and power component pricing announcements, and United States grid capacity, transformer age and gas turbine lead-time data.", "date": "29 June 2026, part 3 of the Bottleneck Heuristic notes series.", "related_product": "Convequity's AI Bubble Barometer (ABBX), extended to carry relative valuation and performance metrics across hyperscalers, neo clouds and leading model labs (product facts provisional, per Gate 4). The note also flags a forthcoming subscriber research note on Capstone Green Energy (CGEH)." } }, "spacex-powering-ai-from-orbit-sizing-the-opportunity-pt-6": { "slug": "spacex-powering-ai-from-orbit-sizing-the-opportunity-pt-6", "title": "SpaceX - Powering AI from Orbit: Sizing the Opportunity (Pt.6)", "url": "https://www.convequity.com/spacex-powering-ai-from-orbit-sizing-the-opportunity-pt-6/", "published": "2026-06-23", "visibility": "tiers", "cohort": "opus", "model": "opus", "status": "corpus_draft_theme_json", "author": "Equity Research", "key_claims": [ "SPCX listed on Nasdaq in June 2026 in a record initial public offering that raised USD 75 billion and priced at a valuation of approximately USD 1.75 trillion, with the listed entity consolidating SpaceX's launch and satellite operations together with X and xAI following their combination in February 2026.", "SPCX reported 2025 revenue of USD 18.7 billion, up 33% from about USD 14 billion in 2024, split between Starlink connectivity at USD 11.4 billion or 61% of the total, the Space Division launch business at roughly USD 4.1 billion or 22%, and the AI business at roughly USD 3.2 billion or 17%, producing USD 6.6 billion of adjusted EBITDA and a GAAP net loss of USD 4.9 billion after about USD 12.7 billion of AI data-centre capital expenditure, while Starlink ended 2025 with approximately 8.9 million subscribers and passed 10 million by early 2026.", "As of the 23 June 2026 report, SPCX had signed three-year compute contracts with Google (GOOGL) and Anthropic worth USD 30 billion and USD 45 billion respectively against roughly USD 20 billion of associated capital expenditure, struck after Anthropic's annualised revenue run rate rose from USD 9 billion at the start of 2026 to USD 40 billion in the first quarter of 2026, and after SPCX's own AI data-centre spending reached USD 7.7 billion in the first quarter of 2026, an annualised run rate of about USD 30.9 billion.", "Tesla (TSLA) launched its TerraFabs chip-fabrication project in March 2026, with early information pointing to ten fabs sited north of the Texas Gigafactory, a first fab targeting 2 nm process technology at 100,000 wafers per month, fan-out panel-level packaging on 600 by 600 mm panels in place of 300 mm wafers, and a first workhorse product in the in-house designed AI5 inference chip, claimed at 40 to 50 times the performance of the prior-generation AI4 with nine times the memory.", "As of the 23 June 2026 report, SPCX had agreed to acquire the coding-agent company Cursor for USD 60 billion, a business that pushed into the model layer by post-training open-source bases such as Kimi 2.5 into its Composer coding models, while the constraint the report cites for terrestrial AI buildout in 2026 is power and permitting, with 40 of 100 planned AI data-centre projects cancelled for lack of secured power." ], "cite_as": "\"SpaceX - Powering AI from Orbit: Sizing the Opportunity (Pt.6),\" Convequity, June 23, 2026, https://www.convequity.com/spacex-powering-ai-from-orbit-sizing-the-opportunity-pt-6/", "paywall_hook": { "question_answered": "what must SPCX actually deliver, in what order and by when, for orbital compute to carry the valuation the market assigned after the June 2026 IPO, and where does the binding constraint on space data centres really sit once launch cost stops being the limiter?", "direction": "constructive on the structural logic of bottleneck arbitrage, moving AI capacity to where power is unlimited and regulatory friction is absent, and on the durability of the Starlink cash engine that funds it; cautious on the near-term model-training position relative to the frontier and explicit that the outcome distribution is unusually wide and close to binary, with the bottleneck reframed away from launch capacity toward solar-cell manufacturing and, ultimately, logic and memory semiconductor supply.", "tickers_covered": [ "SPCX", "TSLA", "GOOGL", "NVDA", "INTC", "AVGO", "LITE", "plus private Anthropic", "OpenAI and Cursor." ], "evidence_types": "post-merger consolidated 2025 financials and segment splits, disclosed IPO size and pricing, contracted multi-year compute agreements, subscriber counts, quarterly capex run rates, component cost and equipment lead-time data points, orbital and launch engineering parameters such as payload cost per kilogram, power per tonne and orbital shell separation, and Convequity's own scenario valuation work.", "date": "23 June 2026, Part 6 of the SpaceX series, written in the weeks immediately following the June 2026 listing.", "related_product": "Convequity's SpaceX valuation model, offering a Special View scenario builder with Base, Bear and Bull paths that can be linked into a standard DCF View (product facts provisional, per Gate 4)." } }, "photonics-the-91-billion-question-layer-1-pt-2": { "slug": "photonics-the-91-billion-question-layer-1-pt-2", "title": "Photonics: The $91 Billion Question - Layer 1 (Pt.2)", "url": "https://www.convequity.com/photonics-the-91-billion-question-layer-1-pt-2/", "published": "2026-06-18", "visibility": "tiers", "cohort": "opus", "model": "opus", "status": "corpus_draft_theme_json", "author": "Equity Research", "key_claims": [ "As of Convequity's 18 June 2026 analysis, Soitec's (SOI) photonic silicon-on-insulator wafers, 200 mm and 300 mm substrates carrying a sub-220 nm top silicon layer over a precisely controlled buried oxide, were the base platform beneath the leading silicon photonics foundry offerings, including TSMC's COUPE, Tower Semiconductor's (TSEM) PH18 and GlobalFoundries' Fotonix, and beneath Broadcom's (AVGO) Bailly co-packaged optics programme, with Soitec's share of photonics SOI put at more than 95%.", "Native silicon Mach-Zehnder and micro-ring modulators fabricated directly on photonic SOI carried the 1.6T co-packaged optics engines shipping in 2026, including Broadcom's (AVGO) Bailly, while hybrid thin-film lithium niobate passed a commercialisation marker in March 2026 when HyperLight's 400G per lane TFLN chiplet became commercially available and Broadcom's latest DSP added explicit support for TFLN modulators.", "Lightwave Logic (LWLG) began shipping commercial evaluation samples of its Perkinamine electro-optic polymer in early 2026 and licenses the material to foundries including Tower Semiconductor (TSEM), with the polymer deposited in back-end-of-line steps onto finished silicon photonics waveguides, a route the source credits with the lowest power draw, the most compact footprint and the highest theoretical bandwidth, above 500 GHz in plasmonic designs.", "Coherent's (COHR) 400G per lane differential-drive EML demonstrations at OFC 2026 were aimed at next-generation pluggable form factors rather than co-packaged optics, and as of June 2026 commercial indium phosphide electro-absorption modulators delivered roughly 53 to 67 GHz against a hard physical ceiling near 100 GHz, leaving indium phosphide's durable role in the separate continuous-wave laser sources that co-packaged optics requires at 300 mW or more per laser, versus about 50 mW for 400G, 75 mW for 800G and 100 mW for 1.6T silicon photonics links, with substrates from AXT (AXTI) and IQE, epitaxy grown on Aixtron (AIXA) MOCVD tools and laser chips from Lumentum (LITE), Coherent (COHR) and SiVe (SIVE).", "As of June 2026, Corning's (GLW) fibre array units were qualified for Broadcom's (AVGO) Bailly platform and Fiber Optic Communications' (3363.TW) ReLFACon fibre array units were being advanced for TSMC's COUPE platform, with each co-packaged optics engine terminating dozens to hundreds of precisely aligned fibres against the 8 to 16 typical of a single pluggable module, while Corning separately developed electro-optical glass interposers with ion-exchanged waveguides and through-glass vias aimed first at TSMC's CoPoS carrier substrates, a path in which LPKF supplies precision through-glass-via drilling." ], "cite_as": "\"Photonics: The $91 Billion Question - Layer 1 (Pt.2),\" Convequity, June 18, 2026, https://www.convequity.com/photonics-the-91-billion-question-layer-1-pt-2/", "paywall_hook": { "question_answered": "if hyperscalers keep pulling optics closer to the ASIC, which layer of the photonic supply chain actually captures the value, and which Layer 1 positions hold regardless of whether switch-level CPO, NPO or full GPU/XPU-package CPO wins?", "direction": "cautious on the headline co-packaged optics TAM forecast that framed Part 1, constructive on the Layer 1 chokepoints, on the basis that the bottleneck sits in engineered substrates, modulator platforms, high-power III-V continuous-wave lasers and last-centimetre precision glass rather than in the architecture debate itself, with the bottleneck shifting further upstream as optical engines multiply from one per switch ASIC to one or more per accelerator package.", "tickers_covered": [ "SOI (Soitec)", "AXTI", "IQE", "AIXA", "LITE", "COHR", "POET", "SIVE", "HIMX", "TSEM", "BESI", "ASM", "LPKF", "CAMT", "ONTO", "KLAC", "AVGO", "GLW", "3363.TW." ], "evidence_types": "foundry platform and programme disclosures, dated product and sampling milestones, conference demonstrations, device-level bandwidth and drive-power specifications, laser output-power thresholds by line rate, packaging and connectivity qualification status, and a four-layer supply-chain framework built by Convequity.", "date": "18 June 2026, Layer 1 of a multi-part photonics supply-chain series following Part 1's demand-side and architecture analysis." } }, "tower-semiconductor-powering-ai-from-orbit-pt-5": { "slug": "tower-semiconductor-powering-ai-from-orbit-pt-5", "title": "Tower Semiconductor - Powering AI From Orbit (Pt.5)", "url": "https://www.convequity.com/tower-semiconductor-powering-ai-from-orbit-pt-5/", "published": "2026-06-16", "visibility": "tiers", "cohort": "opus", "model": "opus", "status": "corpus_draft_theme_json", "author": "Equity Research", "key_claims": [ "At OFC 2026, Tower Semiconductor (TSEM) announced a SiGe BiCMOS plus silicon photonics platform manufactured on full 300 mm wafers, matching the wafer format that GlobalFoundries (GFS) had positioned as the central economic argument for its Fotonix platform.", "Tower Semiconductor's (TSEM) SBC18H6 SiGe BiCMOS node, presented in 2026, achieves Ft/Fmax figures in the range of 325 to 450 GHz, the level required to support 112 Gbaud PAM4 signalling that underpins both 1.6T and 3.2T optical transceiver architectures.", "As of the 16 June 2026 analysis, Tower Semiconductor (TSEM) was demonstrating power consumption approximately 30% below equivalent InP EML solutions, reversing a historical penalty in which silicon photonics transceivers consumed 30% or more than InP, with yields above 85% on its SiGe process against the 40% to 60% range typical of InP specialty nodes that remain largely confined to 100 mm wafers, and Tower estimating resulting transceiver cost at roughly one-third of InP equivalents.", "Broadcom (AVGO) and Marvell (MRVL) have contracted more than 70% of Tower Semiconductor's (TSEM) available capacity through 2028, and Tower is investing over USD 900 million in capacity expansion targeting roughly a fivefold increase in production output through 2026 and beyond.", "Tower Semiconductor (TSEM) has published radiation test data on PH18 photonic integrated circuits covering total ionising dose, single-event latchup and proton irradiation, and multiple defence and space programmes had qualified PH18 components on that heritage as of the 16 June 2026 analysis, while GlobalFoundries (GFS) had no equivalent published radiation heritage for Fotonix." ], "cite_as": "\"Tower Semiconductor - Powering AI From Orbit (Pt.5),\" Convequity, June 16, 2026, https://www.convequity.com/tower-semiconductor-powering-ai-from-orbit-pt-5/", "paywall_hook": { "question_answered": "which silicon photonics foundry platform becomes the default for 1.6T and 3.2T optical interconnect and for orbital deployment, and does GlobalFoundries retain any structural argument against Tower after OFC 2026?", "direction": "constructive on Tower Semiconductor as the incumbent silicon photonics foundry, with the competitive bottleneck shifting from wafer format toward ecosystem maturity, proven yield and radiation heritage; negative on GlobalFoundries' position in this specific segment and on InP EML at the high end of the transceiver market.", "tickers_covered": [ "TSEM", "GFS", "AVGO", "MRVL", "NVDA", "with SpaceX as the anchor customer case and Ayar Labs", "Lightmatter and Ranovus as ecosystem references." ], "evidence_types": "conference process announcements and node specifications, published foundry yield and power comparisons, disclosed customer capacity commitments and capex, published radiation qualification papers, and platform architecture analysis covering hybrid bonding, co-packaged optics and PDK portability.", "date": "16 June 2026, Part 5 of a multi-part series on the space data centre supply chain." } }, "convequitys-ai-bubble-barometer": { "slug": "convequitys-ai-bubble-barometer", "title": "Convequity's AI Bubble Barometer", "url": "https://www.convequity.com/convequitys-ai-bubble-barometer/", "published": "2026-06-14", "visibility": "tiers", "cohort": "opus", "model": "opus", "status": "corpus_draft_theme_json", "author": "Equity Research", "key_claims": [ "In early May 2026, Anthropic secured exclusive access to SpaceX's Colossus 1 data centre in Memphis, committing to roughly USD 1.25 billion per month through May 2029, a potential contract value above USD 45 billion covering more than 220,000 Nvidia (NVDA) GPUs and over 300 MW of power, primarily to serve inference workloads for its Claude models.", "Shortly afterward in 2026, Google (GOOGL) disclosed a separate SpaceX agreement worth approximately USD 920 million per month running from October 2026 through mid-2029, securing around 110,000 Nvidia (NVDA) GPUs in SpaceX facilities for a potential total of USD 29 to 30 billion.", "SpaceX's 2026 listing, at a valuation in the region of USD 2 trillion, is the largest IPO in history by a wide margin, and prospective IPOs from Anthropic and OpenAI would draw on the same pool of institutional and retail capital that is also funding AI infrastructure expansion.", "As of the 14 June 2026 report, AI token consumption remained concentrated in software engineering, with 77% of employees at high-adoption organisations using coding assistants and cited studies indicating average developer productivity gains of 15% to 20%, rising to 35% to 40% on greenfield or lower-complexity tasks, while legal, accounting and healthcare adoption stayed at earlier stages.", "Amazon (AMZN) has disclosed multi-year AI commitments averaging roughly 5.5 years, a contract duration materially longer than legacy cloud workloads and one that shapes how quickly disclosed remaining performance obligations convert into recognised revenue (as of 2026)." ], "cite_as": "\"Convequity's AI Bubble Barometer,\" Convequity, June 14, 2026, https://www.convequity.com/convequitys-ai-bubble-barometer/", "paywall_hook": { "question_answered": "are the major hyperscalers' AI investments being priced by the market in line with the economic returns that capital is actually generating, and what does the framework read in the week to 12 June 2026?", "direction": "no bubble signal in Week 4; the market is paying less for the same or slightly improving fundamental returns, which the series reads as directionally positive, while flagging a widening gap between companies that can secure large contracted offtake and those that cannot, and a medium-term liquidity drain from very large AI-related listings.", "tickers_covered": [ "MSFT", "AMZN", "GOOGL", "ORCL", "META", "CRWV", "NVDA", "with SpaceX", "Anthropic and OpenAI as private-market reference points." ], "evidence_types": "disclosed remaining performance obligations and capex, management commentary on contract duration and AI attribution, disclosed compute offtake agreements, published adoption and productivity survey data, Damodaran implied equity risk premium and company betas for the cost of capital, and the framework's own weekly valuation and return readings.", "date": "14 June 2026, Week 4 of the weekly series, metrics as of 12 June 2026, with updates published every Friday.", "related_product": "Convequity's AI Bubble Barometer, a weekly series with an interactive dashboard comparing two-year forward ROIC against EV/IC for the hyperscaler basket (product facts provisional, per Gate 4)." } }, "aixtron-the-growth-step-pt-3": { "slug": "aixtron-the-growth-step-pt-3", "title": "Aixtron: The Growth Step (Pt.3)", "url": "https://www.convequity.com/aixtron-the-growth-step-pt-3/", "published": "2026-06-11", "visibility": "tiers", "cohort": "opus", "model": "opus", "status": "corpus_draft_theme_json", "author": "Equity Research", "key_claims": [ "Aixtron (AIXA) announced a new manufacturing site in Penang, Malaysia, representing an investment of approximately EUR 40 million, with first shipments targeted for the second half of 2027 and management describing the site as adding roughly 20% to production capacity.", "Aixtron (AIXA) disclosed that its Power segment, covering SiC and GaN tools, accounted for 57% of equipment revenue in fiscal 2025 (calendar 2025) but collapsed to 17% of equipment revenue in fiscal Q1 2026, driven by SiC overcapacity and delayed customer capex.", "Veeco (VECO) disclosed a single order exceeding USD 250 million from a customer it described as a leading optical communications manufacturer, together with \"process of record\" wins for its Lumina InP-capable MOCVD platform, the strongest public evidence available as of the 11 June 2026 analysis that a second source has reached production-scale qualification in InP epitaxy.", "China implemented export controls on indium phosphide substrates in early 2025, layering licensing and allocation friction onto a supply base concentrated in Sumitomo Electric, AXT (AXTI) through its Tongmei operations, and Yunnan Germanium, with AXT doubling InP capacity in 2026 and planning further expansion in 2027 against substrate ramps that typically take 18 to 24 months to deliver qualified volume.", "Yole Group data cited in the 11 June 2026 analysis put Aixtron's (AIXA) share of the InP MOCVD market above 77% globally and at or near 90% in advanced telecom lasers, a position anchored by qualification economics in which a reactor takes 6 to 12 months to qualify in a customer fab and then typically remains in production for 10 to 15 years." ], "cite_as": "\"Aixtron: The Growth Step (Pt.3),\" Convequity, June 11, 2026, https://www.convequity.com/aixtron-the-growth-step-pt-3/", "paywall_hook": { "question_answered": "is Aixtron's near-monopoly in InP MOCVD durable against an improving Veeco platform and eventual Chinese entrants, and can the company physically ship enough tools in fiscal 2027 to meet what the sell side already expects?", "direction": "constructive on the durability of the qualification-based moat and on the multi-year length of the InP demand cycle, cautious on fiscal 2027 revenue conversion, because the binding constraint is shifting away from tool demand toward substrate availability, 6-inch qualification timelines and Aixtron's own calibrated capacity response.", "tickers_covered": [ "AIXA", "VECO", "AXTI", "with reference to Sumitomo Electric", "Yunnan Germanium", "Lumentum", "Coherent and Chinese optical chip makers." ], "evidence_types": "company capacity announcements and segment revenue disclosures, competitor order and process-of-record disclosures, third-party market share data, substrate supply and export control developments, wafer-size transition economics, and a bottom-up unit capacity build compared against sell-side expectations.", "date": "11 June 2026, Part 3 of a multi-part series, with Part 4 covering SiC, GaN and the 800V DC data centre architecture." } }, "co-packaged-optics-the-91-billion-question-pt-1": { "slug": "co-packaged-optics-the-91-billion-question-pt-1", "title": "Photonics: The $91 Billion Question (Pt.1)", "url": "https://www.convequity.com/co-packaged-optics-the-91-billion-question-pt-1/", "published": "2026-06-07", "visibility": "tiers", "cohort": "opus", "model": "opus", "status": "corpus_draft_theme_json", "author": "Equity Research", "key_claims": [ "Goldman Sachs Research's April 2026 report \"Optical Networking: The Next Mega Trend in AI Infrastructure\" sizes the vendor-addressable AI networking equipment market at approximately USD 154 billion cumulatively across 2026 to 2028, assumes co-packaged optics captures roughly 59% of it in the base case for about USD 91 billion, and concentrates the opportunity heavily in 2028 at approximately USD 61 billion, with a high-end scenario of around USD 97 billion cumulative, as cited in Convequity's 7 June 2026 analysis.", "Gartner forecasts total data centre systems spending of approximately USD 790 billion in 2026, representing 56% year-over-year growth, following growth from USD 334 billion in 2024 to approximately USD 506 billion in 2025, a 51.5% increase, with Raymond James research cited for the finding that networking represents 10% to 20% of AI data centre capital expenditure, per the 7 June 2026 analysis.", "Per the 7 June 2026 analysis, Google has reportedly ordered approximately 12 million 1.6T near-packaged optics modules for its next-generation TPU scale-up clusters, is conducting 6.4T near-packaged optics sample testing at 32 lanes of 200G per lane in the second half of 2026 with volume shipment targeted for mid-2027 alongside a TPUv9 architecture designed to use near-packaged optics only, while Alibaba has delivered the first working 3.2T near-packaged optics system with production for Chinese customers expected by 2027, and Meta and Amazon have committed resources to near-packaged optics development.", "NVIDIA's Rubin Ultra NVL576 is scheduled for commercial availability in the second half of 2027, extending a single NVLink domain across eight racks and 576 GPUs in a two-layer all-to-all topology with copper retained inside each rack and co-packaged optics used for inter-rack scale-up links; one estimate cited in the 7 June 2026 analysis puts a large configuration at roughly 792 optical engines, about 648 internal plus 144 external, against the GB300 NVL72 which was almost entirely copper in its single-rack scale-up fabric.", "The 7 June 2026 analysis records that Broadcom's Bailly platform co-packages silicon photonic engines with its 51.2 Tbps Tomahawk 5 switch ASIC and is shipping in limited qualification volumes to select hyperscalers, that Google has deployed optical circuit switching across multiple TPU generations including its seventh-generation TPUv7 Ironwood, that Google's GPU and XPU-level co-packaged optics testing is expected to begin in 2028, and that Soitec's photonic silicon-on-insulator is the dominant starting wafer for the silicon photonics foundry platforms serving both co-packaged and near-packaged optics." ], "cite_as": "\"Photonics: The $91 Billion Question (Pt.1),\" Convequity, June 7, 2026, https://www.convequity.com/co-packaged-optics-the-91-billion-question-pt-1/", "paywall_hook": { "question_answered": "is the widely cited base-case co-packaged optics TAM for 2026 to 2028 a reasonable central estimate or a ceiling, and what happens to it if hyperscalers keep buying near-packaged optics and optical circuit switching instead of adopting switch-level co-packaged optics as an intermediate step?", "direction": "cautious on the consensus co-packaged optics addressable market, which the post reads as a ceiling contingent on rapid broad-based hyperscaler adoption rather than a base case, and cautious specifically on the single year that carries most of the cumulative forecast because it rests on one platform's adoption; constructive on near-packaged optics, whose commercial momentum the post argues is running ahead of consensus for reasons of serviceability, failure-domain isolation and multi-vendor sourcing rather than raw specification; optical circuit switching is framed as the least understood disruptor because it can remove an entire switching layer and the laser engines that go with it; and the substrate layer is framed as the position least exposed to which architecture wins, since every path runs through the same wafer platform.", "tickers_covered": [ "NVDA", "AVGO", "GOOGL", "META", "AMZN", "BABA", "SOI.PA", "AIXA." ], "evidence_types": "third-party sell-side and forecaster market sizing (Goldman Sachs Research, Gartner, Raymond James), disclosed and reported hyperscaler orders and test schedules, vendor platform specifications and announced availability dates, architectural and operational reasoning on serviceability, thermal coupling and vendor lock-in, and Convequity's own top-down networking spend model.", "date": "7 June 2026, Part 1 of a multi-part optical supply chain series structured as four analytical layers, with substrate and epitaxy in Part 2, modulators and continuous-wave lasers in Part 3, packaging and fibre attach in Part 4, and system-level positioning in Part 5." } }, "aixtron-the-growth-step-pt-2": { "slug": "aixtron-the-growth-step-pt-2", "title": "Aixtron: The Growth Step (Pt.2)", "url": "https://www.convequity.com/aixtron-the-growth-step-pt-2/", "published": "2026-06-03", "visibility": "tiers", "cohort": "opus", "model": "opus", "status": "corpus_draft_theme_json", "author": "Equity Research", "key_claims": [ "In its Q1 2026 earnings release and call, Aixtron (AIXA) disclosed that more than 65% to 70% of equipment order intake in the quarter related to its Optoelectronics segment and raised full-year 2026 revenue guidance specifically on stronger-than-expected demand from that area, with chief executive Felix Grawert stating that laser-related demand was exceeding expectations on the back of AI datacom applications.", "Aixtron's G10-AsP MOCVD reactor, first commercially produced in 2023 and priced at roughly EUR 4 to 5 million against EUR 3 million for the older G4 platform, processes eight 6-inch indium phosphide wafers per batch over approximately four-hour runs, and represents less than 5% of the total capital cost of a complete InP laser fabrication line, per Convequity's 3 June 2026 analysis.", "Convequity's 3 June 2026 analysis puts Aixtron's (AIXA) share of optical epitaxy equipment above 90% and its power semiconductor equipment share at roughly 77%, with the power figure under pressure from state-backed Chinese domestic toolmakers, and notes that the company runs a flexible platform-based allocation model sharing production resources across GaN, SiC and InP rather than dedicated lines per material system.", "The 3 June 2026 analysis cites industry consensus from McKinsey, LightCounting and Dell'Oro projecting approximately 90 million 800G-and-above optical transceivers shipping annually by 2027 inclusive of early 1.6T volumes, at roughly eight laser die per transceiver, against recipe qualification timelines of 6 to 12 months for advanced datacom epitaxy and typical tool lead times of 12 to 18 months including customer-side qualification.", "Per the 3 June 2026 analysis, NVIDIA's Rubin Ultra NVL576 arrives commercially in the second half of 2027 as an eight-rack NVLink domain using co-packaged optics only between racks while remaining copper inside each rack, with true intra-rack per-XPU co-packaged optics testing expected from 2028 on platforms such as Kyber and Rubin Ultra, and with Google having placed substantial orders for near-packaged optics and indicated plans to couple every TPU with multiple optical engines." ], "cite_as": "\"Aixtron: The Growth Step (Pt.2),\" Convequity, June 3, 2026, https://www.convequity.com/aixtron-the-growth-step-pt-2/", "paywall_hook": { "question_answered": "how much qualified indium phosphide epitaxy capacity can actually be added over the next three years, what physically constrains it, and does the unresolved contest between pluggables, near-packaged optics, co-packaged optics and optical circuit switching change the answer for the company that sits at the epitaxy step?", "direction": "constructive on the structural position, on the grounds that the tool is a small fraction of fab capex while gating the rest of the line, that supply is constrained at several independent points at once, and that every competing optical architecture still runs through the same epitaxy step, so architectural fragmentation redistributes demand rather than reducing it; cautious on how much of that is already in the price after a large move and on whether consensus revenue expectations require several optimistic assumptions to hold at the same time; cautious on the near-term commercial impact of NVL576, which the post argues is squeezed between maturing near-packaged optics and approaching per-XPU integration; and the competitive threat from the second toolmaker is framed as real but bounded today, with the longer-dated risk being platform maturity at Western incumbents rather than current share.", "tickers_covered": [ "AIXA", "VECO", "LITE", "COHR", "NVDA", "GOOGL." ], "evidence_types": "company earnings release and call commentary, order intake mix and guidance changes, tool throughput and pricing specifications, process-step and qualification-timeline reasoning, third-party transceiver shipment forecasts, and the author's bottom-up per-machine capacity and wafer-demand modelling.", "date": "3 June 2026, Part 2 of the Aixtron series, with Part 3 flagged to carry bottom-up revenue models and the 800V DC power opportunity." } }, "powering-ai-from-orbit-macom-coherent-stm-hamamatsu-pt-4": { "slug": "powering-ai-from-orbit-macom-coherent-stm-hamamatsu-pt-4", "title": "Powering AI From Orbit: MACOM, Coherent, STM, Hamamatsu (Pt.4)", "url": "https://www.convequity.com/powering-ai-from-orbit-macom-coherent-stm-hamamatsu-pt-4/", "published": "2026-05-23", "visibility": "tiers", "cohort": "opus", "model": "opus", "status": "corpus_draft_theme_json", "author": "Equity Research", "key_claims": [ "In a 23 May 2026 analysis of optical networking for orbital data centres, Convequity identified Coherent (COHR), MACOM (MTSI), STMicroelectronics (STM) and Hamamatsu (6965.T) as the near-term leaders for the 2026 to 2029 window on the grounds that each already has radiation-hardened, vacuum-qualified components flying or baselined in operational space-to-space and space-to-ground laser communication programmes, with constellation operators integrating around parts of proven heritage rather than issuing open RFPs.", "The 23 May 2026 analysis records that MACOM (MTSI) supplies transimpedance amplifiers and drivers to free-space optical programmes including Mynaric CONDOR terminals and classified defence programmes, and manufactures through a vertically integrated indium phosphide fab in Lowell, Massachusetts, which the post treats as one of only two or three globally able to supply III-V analog parts at 200 Gbaud per lane.", "According to the 23 May 2026 analysis, Hamamatsu (6965.T) holds more than 70% share of InGaAs avalanche photodiodes on the back of decades of III-V detector development, and its acquisition of NKT Photonics makes it a credible second source for erbium-doped fibre amplifiers alongside Coherent (COHR), while STMicroelectronics (STM) addresses pointing, acquisition and tracking control ASICs on its 28 nm FD-SOI process, which delivers radiation tolerance at automotive-grade pricing and supports control loop updates above 1,000 Hz.", "The 23 May 2026 analysis cites a Mynaric CONDOR Mk3 terminal delivering roughly 100 Gbps in a 30 kg package at approximately USD 500,000, and contrasts that existing product point with the multi-terabit, sub-5 kg, sub-USD 5,000 terminal the author argues an orbital data centre would require; the gap between the two is the author's own specification, not a disclosed programme requirement.", "The 23 May 2026 analysis notes that Starlink V2 Mini satellites each carry three laser terminals at roughly 200 Gbps sized for broadband traffic, that Broadcom's Tomahawk 5 switch ASIC drives 51.2 Tbps of optical I/O through co-packaged optics, and that NVIDIA (NVDA) has invested in Lumentum (LITE) while pursuing a co-packaged optics roadmap, with private photonic interconnect firms Ayar Labs, Celestial AI, Lightmatter and Ranovus named as the most likely 2030 to 2035 entrants." ], "cite_as": "\"Powering AI From Orbit: MACOM, Coherent, STM, Hamamatsu (Pt.4),\" Convequity, May 23, 2026, https://www.convequity.com/powering-ai-from-orbit-macom-coherent-stm-hamamatsu-pt-4/", "paywall_hook": { "question_answered": "who actually captures the optical component spend if AI compute moves into orbit, which suppliers see genuinely incremental demand versus demand merely migrating out of terrestrial data centres, and does today's space heritage advantage survive the move to high-volume production in the 2030s?", "direction": "constructive on the 2026 to 2029 heritage holders in space-qualified analog, detectors and free-space optics, with the strongest combination of near-term position and content growth attributed to the III-V analog layer; cautious on reading that heritage as a durable moat, because the post argues the 2030 to 2035 terminal is more likely to be designed by a terrestrial AI photonic interconnect company adding space qualification than by a space laser incumbent scaling up; cautious on terrestrial fibre demand as compute migrates off the ground; and the post flags inter-orbit relay links as the only category with no terrestrial equivalent.", "tickers_covered": [ "MTSI", "COHR", "STM", "6965.T", "LITE", "TSEM", "GLW", "AVGO", "MRVL", "NVDA." ], "evidence_types": "component-level architecture and link budget reasoning, product and process facts (28 nm FD-SOI, PH18 SiPh platform, InP fab capability, existing terminal specifications), disclosed constellation hardware configurations, named programme heritage, and the author's bottom-up content-per-terminal and terminal-count estimates.", "date": "23 May 2026, Part 4 of the Space Data Centre series." } }, "aixtron-the-growth-step-pt-1": { "slug": "aixtron-the-growth-step-pt-1", "title": "Aixtron: The Growth Step (Pt.1)", "url": "https://www.convequity.com/aixtron-the-growth-step-pt-1/", "published": "2026-05-16", "visibility": "tiers", "cohort": "opus", "model": "opus", "status": "corpus_draft_theme_json", "author": "Equity Research", "key_claims": [ "As described in a 16 May 2026 analysis, Aixtron (Xetra: AIXA) is a German maker of metal-organic chemical vapour deposition reactors used to grow the epitaxial layers of compound semiconductors including gallium nitride, silicon carbide, indium phosphide and gallium arsenide, and it operates as a duopoly with Veeco Instruments (VECO) on an uneven split, holding roughly 77% share in power semiconductor epitaxy equipment and more than 90% in optical applications.", "The 16 May 2026 analysis identifies optical interconnects as the present growth driver, with indium phosphide lasers running at 800G and imminently 1.6T per channel grown on Aixtron's (AIXA) G10-AsP platform, which serves VCSELs, electro-absorption modulated lasers, continuous-wave lasers for co-packaged optics and micro-LED structures, and it notes that the industry transition from VCSELs to EMLs requires more epitaxial layers and tighter compositional tolerances at larger die sizes while driving a substrate format upgrade from 4-inch to 6-inch that forces new tool purchases rather than recipe changes on installed reactors.", "On the power side, the same 16 May 2026 analysis sets out the arithmetic behind the 800V DC transition that NVIDIA (NVDA) and the wider industry are pushing as AI rack power passes 100 kW, with current liquid-cooled configurations including NVIDIA's Rubin platform exceeding 200 kW per rack and 1 MW per rack as a stated design objective: delivering 1 MW at 48V requires about 20,833 amps against about 1,250 amps at 800V, a 17x current reduction that cuts resistive cable heat loss by roughly 280x, while silicon's critical electric field strength is around ten times lower than that of gallium nitride or silicon carbide.", "The 16 May 2026 analysis maps the two wide-bandgap materials onto distinct stages of the 800V chain served by Aixtron's (AIXA) G10-GaN and G10-SiC tools, with silicon carbide handling grid AC to 800V DC conversion at the facility perimeter in the 800V to 1200V and above domain and benefiting from cost-curve maturity built in electric vehicle powertrains, gallium nitride taking the 800V bus down to rack-level and point-of-load delivery below 1000V at higher switching frequency and smaller magnetics but at higher cost and lower production maturity, and conventional silicon persisting only in the final sub-1V regulation stage at the processor die.", "As of May 2026 the power semiconductor catalyst remained outside Aixtron's (AIXA) backlog: the industry was still digesting capacity added in the 2021 to 2023 upcycle through inventory corrections at Wolfspeed (WOLF), Infineon (IFX) and onsemi (ON) against slower than expected Western electric vehicle adoption, while China's state-backed domestic MOCVD industry, most notably AMEC, was qualifying reactors for mainstream power applications and eroding Aixtron's share inside China, even as Aixtron management indicated that AI data centres could become the largest end market for gallium nitride power devices." ], "cite_as": "\"Aixtron: The Growth Step (Pt.1),\" Convequity, May 16, 2026, https://www.convequity.com/aixtron-the-growth-step-pt-1/", "paywall_hook": { "question_answered": "which of Aixtron's two growth legs is actually underwritable in May 2026, the optical interconnect cycle already converting into order intake or the 800V DC power semiconductor build, and how much of the power opportunity is structurally lost to Chinese domestic epitaxy equipment?", "direction": "constructive on the optical epitaxy leg as the present, high-visibility driver where share is near total and the qualification barrier rather than price is the gating factor, and where each successive photonic generation demands more epitaxial complexity and larger substrates; cautious on the power semiconductor leg, treated as structurally inevitable but cyclically delayed and materially less captive, with the key argument that Chinese domestic tools compress the addressable market globally rather than only regionally once Chinese-fabricated devices reach Western supply chains; bottleneck shift framed as moving from GPU compute itself toward what connects the GPUs and what powers them.", "tickers_covered": [ "AIXA (Xetra)", "VECO", "NVDA", "IFX", "STM", "ON", "WOLF", "NVTS", "AXTI", "COHR", "LITE", "with AMEC", "EPC and Prysmian-style unlisted peers referenced." ], "evidence_types": "market share and duopoly structure, switching-cost and qualification reasoning, tool platform and wafer-format roadmaps, device physics and power-delivery arithmetic, customer and management commentary, cyclical inventory-correction evidence, and Chinese domestic equipment qualification under state policy.", "date": "16 May 2026, Part 1 of a two-part series." } }, "powering-ai-from-orbit-spacex-space-solar-and-the-musk-stack-pt-3": { "slug": "powering-ai-from-orbit-spacex-space-solar-and-the-musk-stack-pt-3", "title": "SpaceX - Powering AI from Orbit, Space Solar, and the Musk Stack (Pt.3)", "url": "https://www.convequity.com/powering-ai-from-orbit-spacex-space-solar-and-the-musk-stack-pt-3/", "published": "2026-05-11", "visibility": "tiers", "cohort": "opus", "model": "opus", "status": "corpus_draft_theme_json", "author": "Equity Research", "key_claims": [ "In an 11 May 2026 analysis, Convequity set out the device physics that makes indium phosphide non-substitutable at high lane rates: at 800G per lane (400 Gbaud PAM4) a laser driver must swing more than 2 volts across a 25-ohm load at analog bandwidths above 80 GHz, and InP heterojunction bipolar transistors carry intrinsic transit frequencies above 400 GHz against silicon CMOS operating near its ceiling, a 3x to 5x speed headroom that converts into roughly half the power draw or twice the signal quality at equal bandwidth.", "The same 11 May 2026 analysis tracks inter-satellite link bandwidth to the terrestrial data centre interconnect roadmap, with 1.6T in early deployment around 2028 to 2029, 6.4T in the main buildout phase around 2030 to 2032 and 12.8T from 2033, each generation roughly quadrupling per-link bandwidth through lane count moving from 4 to 8 to 16 and per-lane rates from 100G to 200G and above.", "Convequity's 11 May 2026 analysis notes that advanced silicon CMOS at 5nm and 3nm nodes is already fast enough to drive 100G per lane, which opens the highest-volume short-reach segment to silicon photonic PICs with integrated silicon drivers and makes MACOM's (MTSI) InP a premium option rather than a design requirement at first-generation 1.6T link speeds.", "In the terrestrial supply picture described on 11 May 2026, InP wafer capacity is a genuine bottleneck that is itself pushing the industry toward silicon photonic pluggables and co-packaged optics, and co-packaged optics absorbs the laser driver function into silicon while still requiring high-performance transimpedance amplifiers and enabling more lanes per switch ASIC, which the analysis reads as a mix shift for MACOM (MTSI) rather than a revenue decline.", "The 11 May 2026 analysis frames the demand backdrop as hyperscaler IT load rising roughly 6x by 2035 and requiring more than 100 GW of new capacity against grid connection queues, permitting, water and land constraints, and identifies the physical asymmetry driving the component mix, namely that vacuum has zero chromatic dispersion so direct detection works at any distance, unlike fibre where dispersion forces coherent detection with digital signal processing ASICs sold by Marvell (MRVL) and Broadcom (AVGO) at roughly USD 50 to 200 per transceiver." ], "cite_as": "\"SpaceX - Powering AI from Orbit, Space Solar, and the Musk Stack (Pt.3),\" Convequity, May 11, 2026, https://www.convequity.com/powering-ai-from-orbit-spacex-space-solar-and-the-musk-stack-pt-3/", "paywall_hook": { "question_answered": "if a meaningful share of incremental AI compute migrates to orbital data centres, which nodes of the photonic supply chain gain, which lose, and is the shift additive to terrestrial transceiver volume or a substitution on the growth margin?", "direction": "constructive on MACOM (MTSI) as the clearest single-company beneficiary, because content per terminal scales superlinearly with bandwidth and InP becomes a hard requirement rather than a premium once power-constrained satellites run 800G per lane; constructive on STMicroelectronics (STM) and Coherent (COHR) in categories with no terrestrial equivalent, namely pointing, acquisition and tracking subsystems and free-space optical assemblies; negative on Corning (GLW) data centre fibre, which loses an entire category because the propagation medium is vacuum, and a secondary headwind for the coherent DSP volumes at Marvell (MRVL) and Broadcom (AVGO); neutral for Lumentum (LITE) and Tower Semiconductor (TSEM), where the thesis rests on space-qualification pricing premiums and component mix rather than net new volume; bottleneck shift argued from fibre and dispersion management toward beam pointing, power budget and analog speed headroom.", "tickers_covered": [ "MTSI", "LITE", "TSEM", "COHR", "STM", "MRVL", "AVGO", "GLW", "with Prysmian and privately held SpaceX and Starlink referenced." ], "evidence_types": "orbital cluster geometry and link-distance analysis, per-satellite terminal counts and bill-of-materials build-ups, link-budget reasoning on aperture, amplification and detector choice, device physics of InP versus silicon CMOS at 400 Gbaud, the industry interconnect bandwidth roadmap, terrestrial co-packaged optics transition and InP wafer capacity, and hyperscaler capacity projections.", "date": "11 May 2026, Part 3 of a multi-part SpaceX series building on architecture established in Part 2." } }, "palantir-q1-2026-what-cursor-proved-what-the-market-missed-and-why-governance-wins-the-agentic-era": { "slug": "palantir-q1-2026-what-cursor-proved-what-the-market-missed-and-why-governance-wins-the-agentic-era", "title": "Palantir Q1 2026 - What Cursor Proved, What the Market Missed, and Why Governance Wins the Agentic Era", "url": "https://www.convequity.com/palantir-q1-2026-what-cursor-proved-what-the-market-missed-and-why-governance-wins-the-agentic-era/", "published": "2026-05-07", "visibility": "tiers", "cohort": "opus", "model": "opus", "status": "corpus_draft_theme_json", "author": "Equity Research", "key_claims": [ "Palantir Technologies (PLTR) reported fiscal Q1 2026 revenue of USD 1.633 billion, up 85% year over year and 16% sequentially, with US revenue up 104% to USD 1.282 billion, US commercial revenue up 133% to USD 595 million and US government revenue up 84% to USD 687 million.", "In the same fiscal Q1 2026 report, Palantir (PLTR) posted GAAP operating income of USD 754 million for a 46% GAAP operating margin and raised full-year 2026 revenue guidance to approximately USD 7.65 to 7.66 billion, implying roughly 71% annual growth against a prior guide implying about 61%, on a customer base of just over 1,000 accounts.", "Alongside its fiscal Q1 2026 results in May 2026, Palantir (PLTR) disclosed deployment metrics for ShipOS, its platform for the US Navy shipbuilding industrial base, including bill-of-materials approvals cut from 200 hours to 15 seconds, contract reviews accelerated by 57% to 73%, and material planning time reduced by 94%.", "On its fiscal Q1 2026 earnings call in May 2026, Palantir (PLTR) outlined a platform-native agent engine SDK, conceptually parallel to its existing OSDK, intended as a single control plane with per-agent, per-session and per-workflow cost attribution under administrative caps, provenance that traces every ontology mutation back to the agent and reasoning chain that produced it, and security marking propagation from input data through agent sessions to output with approval gates on any workflow that could reclassify information, with management framing the enterprise requirement as who authorized this, what did it cost, and can I trust what it did.", "Reported funding and revenue milestones cited in the 7 May 2026 analysis place Cursor at roughly USD 100 million ARR in January 2025, above USD 500 million at its June 2025 Series C, above USD 1 billion at its November 2025 Series D which raised USD 2.3 billion at a USD 29.3 billion valuation, roughly USD 2 billion by February 2026 and about USD 2.7 billion by April 2026, a trajectory that continued through the 2025 and 2026 launches of vertically integrated coding agents by OpenAI, Anthropic and Google." ], "cite_as": "\"Palantir Q1 2026 - What Cursor Proved, What the Market Missed, and Why Governance Wins the Agentic Era,\" Convequity, May 7, 2026, https://www.convequity.com/palantir-q1-2026-what-cursor-proved-what-the-market-missed-and-why-governance-wins-the-agentic-era/", "paywall_hook": { "question_answered": "does agentic AI from the foundation model labs structurally disrupt Palantir's application-layer position after fiscal Q1 2026, and how much weight should the model-layer dependency created by Palantir's US government and intelligence constraints carry against the quarter's operating evidence?", "direction": "constructive on Palantir's application and governance layer, reading the post-earnings agentic-disruption fear as reflexive pattern-matching against a closer precedent that resolved the other way; cautious specifically on model-layer dependency, where a prime contractor to the US government cannot adopt the best-performing models globally and has less freedom to differentiate than a commercial application vendor; bottleneck shift argued from token cost toward orchestration, governance and compounding step-level error in long agentic chains.", "tickers_covered": [ "PLTR", "with PANW", "DDOG", "DT referenced", "plus privately held OpenAI", "Anthropic and Cursor as comparables." ], "evidence_types": "quarterly earnings release figures and raised guidance, earnings call commentary and management quotation, disclosed customer deployment metrics, reported ARR and funding milestones for private application-layer comparables, inference price history, third-party market sizing for agent observability, and Convequity's prior agent TAM model.", "date": "7 May 2026, covering results for fiscal Q1 2026." } }, "powering-ai-from-orbit-spacex-space-solar-and-the-musk-stack-pt-2": { "slug": "powering-ai-from-orbit-spacex-space-solar-and-the-musk-stack-pt-2", "title": "SpaceX - Powering AI from Orbit, Space Solar, and the Musk Stack (Pt.2)", "url": "https://www.convequity.com/powering-ai-from-orbit-spacex-space-solar-and-the-musk-stack-pt-2/", "published": "2026-05-04", "visibility": "tiers", "cohort": "opus", "model": "opus", "status": "corpus_draft_theme_json", "author": "Equity Research", "key_claims": [ "As recorded in Convequity's 4 May 2026 analysis, SpaceX has applied to the United States Federal Communications Commission to launch more than 1.2 million satellites, more than every satellite humanity has placed in orbit combined, structured as three tiers: roughly 800,000 in very low Earth orbit at 500 kilometres for inference at about 1,000 TOPS per satellite and around 20 milliseconds of latency to ground users, roughly 300,000 in low Earth orbit at 1,000 kilometres at about 5,000 TOPS for training, and roughly 118,000 in medium Earth orbit at 2,000 kilometres at about 20,000 TOPS for long-running workloads and storage.", "The first wave of orbital compute satellites described in the 4 May 2026 analysis are Starlink V3 platform variants weighing roughly 2 tons each with 30 kW of solar capacity and about 500 square metres of panel area, of which approximately 850 kilograms is compute payload carrying 12 to 24 space-rated Dojo 2 chips plus radiation hardening, liquid cooling loops and power conversion; at 100 to 150 tons of Starship payload to low Earth orbit, a single flight delivers 50 to 75 satellites, and the deployment window targets 10% of the constellation within three years, 50% by year six and 100% by year nine.", "Tesla (TSLA) is restarting its Buffalo, New York factory for solar panel manufacturing, and the 4 May 2026 analysis states that Elon Musk has set 2028 as the deadline for Tesla and SpaceX each to stand up 100 GW of annual solar cell production capacity, a window the post describes as the shortest achievable because heterojunction technology needs roughly two years to reach commercial maturity even under fully accelerated investment and labour.", "Tesla plans to begin production of its AI5 chip in fiscal Q2 2027 and the more advanced AI6 in fiscal Q2 2028, with AI6 intended as the core processor for orbital space data centres and as Tesla's first chip designed for both training and inference across architectures; the 4 May 2026 analysis adds that Dojo 3 will be built around clusters of AI6 chips in a more standard NPU style after earlier Dojo generations proved difficult to scale on custom RISC-V, software-defined dataflow and vertical DRAM designs, and that Tesla unveiled its Terafab in-house fabrication plans in March 2026 on the back of an eight-year Samsung supply relationship.", "As of the 4 May 2026 analysis, Starlink already operates more than 9,000 optical inter-satellite laser terminals in orbit, three per V2 Mini satellite and each running at up to 200 Gbps at the 1550 nanometre telecom wavelength, which the post describes as the largest coherent optical network ever built, while the space data centre architecture requires 400 Gbps to more than 1 Tbps per link and implies 4 to 8 terminals per satellite, or 5 to 10 million terminals across the nine-year deployment." ], "cite_as": "\"SpaceX - Powering AI from Orbit, Space Solar, and the Musk Stack (Pt.2),\" Convequity, May 4, 2026, https://www.convequity.com/powering-ai-from-orbit-spacex-space-solar-and-the-musk-stack-pt-2/", "paywall_hook": { "question_answered": "what does a space-based AI data centre actually look like in orbit, and where do the binding constraints sit once Starship has made launch tractable, across solar cell chemistry, radiation-hardened silicon, thermal rejection, launch cadence and the optical interconnect supply chain?", "direction": "constructive on the orbital compute architecture as an engineering and manufacturing grind rather than an unsolved science problem, with the bottleneck explicitly shifting away from launch cost toward solar cell production, radiation-hardened chip supply and photonics volume; cautious on the near-term window, since the first two years are framed as experimental proof-of-concept missions while heterojunction matures and perovskite remains unvalidated; and a structural reweighting of the optical supply chain toward cheap high-volume transmitters and away from the coherent module business.", "tickers_covered": [ "TSLA and NVDA", "plus privately held SpaceX and xAI", "Samsung as a fabrication counterparty", "and the specialty alloy", "rare earth and photonics component suppliers exposed to the Starship and terminal production ramp." ], "evidence_types": "the FCC constellation filing, satellite platform mass, power and payload specifications, disclosed chip production timelines, stated manufacturing capacity targets and deadlines, orbital launch and cadence arithmetic, radiative heat rejection and optical link physics, current Starlink laser terminal deployment data, and Convequity's own cost stack and deployment modelling.", "date": "4 May 2026, Part 2 of a two-part series; Part 1 published 25 April 2026 covers the space solar case." } }, "soitec-where-silicon-photonics-begins": { "slug": "soitec-where-silicon-photonics-begins", "title": "Soitec - Where Silicon Photonics Begins", "url": "https://www.convequity.com/soitec-where-silicon-photonics-begins/", "published": "2026-04-30", "visibility": "tiers", "cohort": "opus", "model": "opus", "status": "corpus_draft_theme_json", "author": "Equity Research", "key_claims": [ "Soitec (SOI) reported fiscal Q3 2026 results on 3 February 2026 with revenue of EUR 160 million, down 29% year on year but up 18% sequentially, group gross margin compressed to roughly 25% against a historical 25% to 38% band on legacy fab under-utilisation, and management guided fiscal Q4 to a further improvement of about 20% sequentially; trailing twelve month group revenue stood at approximately EUR 784 million, or USD 917 million.", "As described in Convequity's 30 April 2026 analysis, Photonics-SOI is an engineered substrate with a top silicon device layer controlled to roughly 220 nanometres, matched to single-mode operation at the 1550 nanometre datacentre wavelength, sitting on a buried oxide layer of typically 2 to 3 micrometres for optical isolation, produced by Soitec's Smart Cut hydrogen implantation and bonding process and supplied as starting material to the leading silicon photonics foundries including TSMC, GlobalFoundries, Tower Semiconductor and Intel.", "Soitec's Photonics-SOI line is a roadmap rather than a single product as of April 2026, running from the Photon series already in volume production for 400G and 800G pluggable transceivers, through Photon Plus in 200 millimetre and 300 millimetre formats for advanced pluggables and early co-packaged optics, to Photon Ultra in development for 1.6T and above pluggables and full co-packaged optics, with each generation requiring fresh foundry co-qualification against a tighter thickness and defectivity specification.", "Tower Semiconductor (TSEM) announced a capacity expansion of approximately USD 920 million explicitly tied to 1.6T and above silicon photonics and supported by committed customer pre-payments, and has guided to more than a fivefold increase in monthly silicon photonics wafer starts by the end of 2026, with comparable and less publicly disclosed additions underway at GlobalFoundries Fab 9 in Burlington, Vermont, and within the TSMC silicon photonics roadmap.", "Innolight management commentary cited in the 30 April 2026 analysis implies silicon photonics moved from roughly 20% of optical transceivers in 2025 to about 50% by early 2026, with a path to roughly 70% by 2027, a shift the post attributes to an indium phosphide and electro-absorption modulated laser supply crunch, smaller III-V wafer sizes, low-yielding EML fabrication concentrated in few suppliers including Lumentum, and multi-quarter InP wafer lead times into foundries." ], "cite_as": "\"Soitec - Where Silicon Photonics Begins,\" Convequity, April 30, 2026, https://www.convequity.com/soitec-where-silicon-photonics-begins/", "paywall_hook": { "question_answered": "is Soitec's position in Photonics-SOI a genuine near-monopoly on the substrate underneath the AI optical build-out rather than a licensable process, and what does the AI-driven memory shortage do to the timing and depth of the legacy RF-SOI recovery that most of the market is waiting on?", "direction": "constructive on Soitec, with the mechanism being a mix shift toward a structurally higher-margin Photonics-SOI book rather than a contested race win, and cautious on the legacy RF-SOI recovery, which the analysis argues arrives later and shallower than consensus because an exogenous memory cost shock is suppressing low and mid tier handset volumes; the migration from pluggables through near-packaged to co-packaged optics is read as an intensification of substrate demand rather than a substitution away from it.", "tickers_covered": [ "SOI", "TSEM", "GFS", "TSM", "INTC", "LITE", "NVDA", "AVGO", "MRVL", "plus Shin-Etsu", "GlobalWafers", "Simgui", "Innolight and POET Technologies as supply chain references." ], "evidence_types": "quarterly results and management guidance, substrate device physics and specification detail, the Smart Cut process and its licensing scope, foundry qualification and second-source status, announced foundry capex and wafer start guidance, third-party management commentary on silicon photonics share, memory pricing and handset production reporting, and Convequity's own channel checks and bottom-up sizing.", "date": "30 April 2026, following the preceding Notes on POET Technologies and its Optical Interposer." } }, "powering-ai-from-orbit-spacex-space-solar-and-the-musk-stack-pt-1": { "slug": "powering-ai-from-orbit-spacex-space-solar-and-the-musk-stack-pt-1", "title": "SpaceX - Powering AI from Orbit, Space Solar, and the Musk Stack (Pt.1)", "url": "https://www.convequity.com/powering-ai-from-orbit-spacex-space-solar-and-the-musk-stack-pt-1/", "published": "2026-04-25", "visibility": "tiers", "cohort": "opus", "model": "opus", "status": "corpus_draft_theme_json", "author": "Equity Research", "key_claims": [ "As of Convequity's 25 April 2026 analysis, SpaceX had cut low Earth orbit payload cost from more than USD 25,000 per kilogram to roughly USD 2,500 per kilogram using Falcon 9's reusable first stage and recoverable fairing, while Starship, the first vehicle with both stages reusable and powered by the Raptor full-flow staged combustion engine, carries more than 200 tons to LEO and targets below USD 100 per kilogram.", "In calendar 2025, roughly 3,000 tons of payload were delivered to orbit globally, of which SpaceX accounted for about 2,500 tons including 2,100 tons of Starlink satellites, against roughly 300 tons for China, and Starlink represented approximately 70% of all satellites launched worldwide that year, with one Falcon 9 carrying about 25 Starlink satellites versus roughly 500 per Starship.", "The 25 April 2026 analysis records the terrestrial silicon efficiency record at 24.7%, set by Risen Energy on heterojunction amorphous silicon, against roughly 21% for mainstream monocrystalline silicon and 30% to 40% for the gallium arsenide multi-junction cells flown in space, where triple-junction InGaP/InGaAs/germanium designs deliver 30% to 32% in service; the same analysis puts United States silicon-based solar output below 4 GW per year against roughly 50 GW of annual system deployment, with First Solar's cadmium telluride thin film supplying 14 GW or more of domestic capacity.", "Convequity's 25 April 2026 analysis reports that xAI ordered a further five gas turbine generators of 380 MW each from South Korea's Doosan Enerbility, reflecting an environment in 2026 where multi-year turbine blade backlogs have pushed AI data centre buyers past tier-one suppliers toward production-ready South Korean and Chinese vendors.", "China approved Tesla (TSLA) in May 2018 to establish a wholly owned subsidiary, the first exemption from the 49/51 joint venture requirement for a foreign carmaker; Giga Shanghai broke ground on 7 January 2019 and delivered its first China-made Model 3 vehicles on 30 December 2019, and the 25 April 2026 analysis treats that ramp, led by Tom Zhu under then Shanghai mayor Li Qiang, as the template for SpaceX's production scaling problem." ], "cite_as": "\"SpaceX - Powering AI from Orbit, Space Solar, and the Musk Stack (Pt.1),\" Convequity, April 25, 2026, https://www.convequity.com/powering-ai-from-orbit-spacex-space-solar-and-the-musk-stack-pt-1/", "paywall_hook": { "question_answered": "if terrestrial power is the binding constraint on AI compute, does orbital solar paired with Starship resolve it before nuclear, gas turbines or ground-based solar can, and what has to be true on cell chemistry and on United States to China industrial cooperation for that path to close?", "direction": "constructive on space-based solar as the route around the terrestrial energy bottleneck and on the Musk vertical stack as the only coherent path to it, constructive on United States natural gas producers as a secondary beneficiary of launch cadence and interim data centre power, and cautious on timing, because the remaining obstacles are political, commercial and manufacturing execution rather than scientific.", "tickers_covered": [ "TSLA", "FSLR", "EXE", "AR", "CRK", "EQT", "plus privately held SpaceX and xAI and listed and unlisted Chinese solar cell and equipment suppliers." ], "evidence_types": "regulatory and corporate events with dates, announced launch cost and payload figures, 2025 orbital tonnage and satellite launch counts, published cell efficiency records and device physics, disclosed turbine orders, United States domestic solar manufacturing capacity, ITAR sourcing constraints, and domain expert conversations on heterojunction and perovskite readiness.", "date": "25 April 2026, Part 1 of a two-part series; Part 2 published 4 May 2026 covers the orbital architecture." } }, "notes-poet-technologies-the-optical-interposer-thesis": { "slug": "notes-poet-technologies-the-optical-interposer-thesis", "title": "Notes: POET Technologies - The Optical Interposer Thesis", "url": "https://www.convequity.com/notes-poet-technologies-the-optical-interposer-thesis/", "published": "2026-04-20", "visibility": "tiers", "cohort": "opus", "model": "opus", "status": "corpus_draft_theme_json", "author": "Equity Research", "key_claims": [ "As described in Convequity's 20 April 2026 note, POET Technologies' Optical Interposer is a silicon wafer with waveguides and mechanical guides etched into it, allowing lasers, detectors and electronics to be placed by automated pick-and-place at wafer level and replacing the active alignment step in which technicians power a laser and nudge parts until light couples, a method that held up at 10G, 25G and 100G but broke down at 400G and 800G on both precision and throughput.", "POET runs a fab-lite model as of April 2026, with interposer wafers fabricated on standard CMOS processes by SilTerra, an 8-inch foundry in Malaysia, and optical engine assembly, test and packaging carried out on POET-owned wafer-level equipment at Globetronics and NationGate in Penang, with completed engines supplied to transceiver makers including Foxconn's FIT subsidiary, Luxshare and LITEON.", "In early 2026 POET raised USD 375 million across three equity rounds each priced above the prior round, received a production order valued at over USD 5 million for POET Infinity optical engines from a leading systems integrator, and secured production agreements for approximately 30,000 optical engines, against roughly USD 1 million of revenue in the prior year.", "Broadcom's (AVGO) co-packaged optics platform, built on the Tomahawk 5 \"Bailly\" switch ASIC and in mass production as of April 2026, uses a Remote Laser Module architecture that keeps the failure-prone continuous-wave indium phosphide laser in a hot-pluggable front-panel module connected by 127 micron pitch fibre connectors supporting up to 72 fibre pairs, and the platform has logged over one million cumulative fault-free hours on 400G ports at Meta (META).", "Marvell (MRVL) acquired photonic architecture startup Celestial AI, for which POET supplies the optical interposer, and NVIDIA (NVDA) invested USD 2 billion in Marvell, placing POET inside an open, acquisition-driven stack as of April 2026, in contrast to Broadcom's vertically integrated co-packaged optics chain, which spans its own photonics engine, third-party continuous-wave lasers from Lumentum (LITE) and Coherent (COHR), and packaging hardware from Foxconn Interconnect Technology." ], "cite_as": "\"Notes: POET Technologies - The Optical Interposer Thesis,\" Convequity, April 20, 2026, https://www.convequity.com/notes-poet-technologies-the-optical-interposer-thesis/", "paywall_hook": { "question_answered": "does POET Technologies' optical interposer actually solve the assembly bottleneck that is throttling optical engine output, and what has to happen between 2026 and 2028 for that platform advantage to become revenue?", "direction": "constructive on the structural need for a platform that removes active alignment and opens transceiver manufacturing to high-volume electronics assemblers, and on POET's position inside the Marvell and Celestial AI supply chain, but explicitly cautious on the company itself, which is framed as a venture-style position with no earnings floor, an existential execution window, and structural exclusion from Broadcom's co-packaged optics stack; silicon photonics adoption is treated as a partial near-term headwind to the addressable opportunity before the economics turn favourable again at 3.2T.", "tickers_covered": [ "POET", "LITE", "AVGO", "MRVL", "NVDA", "META", "COHR", "QUBT", "TSM", "GFS", "plus Foxconn", "Luxshare", "LITEON", "Innolight", "Eoptolink", "SilTerra", "Globetronics and NationGate as supply chain references." ], "evidence_types": "company announcements of financings, production orders and unit commitments, supply chain and contract manufacturing relationships, product architecture disclosures, reported market share and gross margin ranges for adjacent optical suppliers, and a staged 2024 to 2030 execution timeline with named catalysts.", "date": "20 April 2026, part of an expanding optical networking supply chain series.", "related_product": "companion to Convequity's AI Value Chain Framework and the Bottleneck Heuristic series, with earlier standalone coverage of Lumentum (LITE) and the broader optical landscape." } }, "notes-the-bottleneck-heuristic-blackwells-gauntlet-the-asic-breakout-and-the-optical-chokepoint-pt-2": { "slug": "notes-the-bottleneck-heuristic-blackwells-gauntlet-the-asic-breakout-and-the-optical-chokepoint-pt-2", "title": "Notes: The Bottleneck Heuristic — Blackwell's Gauntlet, the ASIC Breakout, and the Optical Chokepoint (Pt.2)", "url": "https://www.convequity.com/notes-the-bottleneck-heuristic-blackwells-gauntlet-the-asic-breakout-and-the-optical-chokepoint-pt-2/", "published": "2026-04-15", "visibility": "tiers", "cohort": "opus", "model": "opus", "status": "corpus_draft_theme_json", "author": "Equity Research", "key_claims": [ "NVIDIA's (NVDA) Blackwell was its first chiplet-based GPU and its first use of TSMC's CoWoS-L packaging, and an initial tape-out error at the local silicon interconnect bridge forced mask revisions costing roughly four to six months, contributing to first meaningful hyperscaler production deployments slipping to the second half of 2025, approximately nine to twelve months behind the original schedule.", "The Blackwell NVL72 rack connects 36 dual-GPU nodes into a single 72-GPU scale-up domain over fifth-generation NVLink at roughly 120 kW, and NVIDIA's mitigation for the enlarged failure blast radius designates 8 of the 72 GPUs as idle spares, stranding 11.1% of purchased compute in systems shipping through 2025.", "Google's TPUv7, designed with Broadcom (AVGO), trained Gemini 3 Pro, which in late 2025 surpassed GPT-5 on multiple benchmarks, and also trained Anthropic's Claude Opus 4.5 and 4.6 at a roughly 3x reduction in token costs versus equivalent NVIDIA-based compute, breaking the prior consensus that custom ASICs were inference-only.", "Broadcom (AVGO) has guided to a custom AI accelerator revenue opportunity of USD 60 billion to USD 90 billion by its fiscal 2027 ending October 2027, while Microsoft's Maia, designed with Marvell (MRVL), saw only limited internal deployment and Amazon's Trainium 2 delivered roughly 80% of H100 chip-level performance, prompting Meta, SoftBank, OpenAI, X.ai and ByteDance to engage Broadcom as design partner on programmes that typically take two to three years to reach volume silicon.", "NVIDIA confirmed its optical pivot at GTC 2026 with in-rack optical connections and the 576-GPU NVL576 scale-up domain after placing emergency transceiver orders with Innolight and Eoptolink in late 2025, colliding with a 1.6T transceiver market where the 200G per lane EML on indium phosphide is supplied at scale only by Lumentum (LITE), Broadcom (AVGO) and Sumitomo Electric, and where Tower Semiconductor (TSEM) plans roughly USD 900 million of capacity investment in 2026 to reach about 5x its current silicon photonics output." ], "cite_as": "\"Notes: The Bottleneck Heuristic — Blackwell's Gauntlet, the ASIC Breakout, and the Optical Chokepoint (Pt.2),\" Convequity, April 15, 2026, https://www.convequity.com/notes-the-bottleneck-heuristic-blackwells-gauntlet-the-asic-breakout-and-the-optical-chokepoint-pt-2/", "paywall_hook": { "question_answered": "after Blackwell's execution gauntlet and Google's TPUv7 proving custom silicon can train at the frontier, where has the binding constraint in AI infrastructure moved, and which suppliers can actually relieve it?", "direction": "constructive on Broadcom as custom ASIC design partner of record, on Tower Semiconductor as the qualified silicon photonics foundry, and on transceiver makers with early silicon photonics access; cautious on NVIDIA's ability to defend the high-margin training TAM without leaning harder on system-level integration; negative on the hyperscaler strategy of economising on design partners, and cautious on indium phosphide EML incumbents whose capacity response is physically constrained.", "tickers_covered": [ "NVDA", "AVGO", "TSEM", "MRVL", "GOOGL", "MSFT", "AMZN", "META", "LITE", "GFS", "CRDO", "AMD", "INTC", "TSM", "plus Innolight", "Eoptolink and Sumitomo Electric as foreign-listed references." ], "evidence_types": "product architecture and packaging disclosures, rack-level power and reliability specifications, published model benchmark outcomes, management revenue opportunity guidance, transceiver order flow and lead-time reporting, indium phosphide versus CMOS wafer economics, and announced foundry capacity expansion plans.", "date": "15 April 2026, Part 2 of the Bottleneck Heuristic series, covering the 2024 to 2025 wave and the forward view into the 2025 to 2026 optical chokepoint.", "related_product": "continuation of Convequity's AI Value Chain Framework and the Part 1 bottleneck heuristic, which identifies bottleneck-solvers segment by segment." } }, "notes-the-bottleneck-heuristic-where-scarcity-meets-capability-in-the-ai-value-chain": { "slug": "notes-the-bottleneck-heuristic-where-scarcity-meets-capability-in-the-ai-value-chain", "title": "Notes: The Bottleneck Heuristic - Where Scarcity Meets Capability in the AI Value Chain (Pt.1)", "url": "https://www.convequity.com/notes-the-bottleneck-heuristic-where-scarcity-meets-capability-in-the-ai-value-chain/", "published": "2026-04-10", "visibility": "tiers", "cohort": "opus", "model": "opus", "status": "corpus_draft_theme_json", "author": "Equity Research", "key_claims": [ "NVIDIA (NVDA) saw gross margin expand from roughly 57% in fiscal 2023 to above 75% in fiscal 2024 as the only AI accelerator supplier at scale, with AMD's MI250 behind on software ecosystem and production system maturity, Intel's Gaudi behind on performance, and custom ASICs still nascent.", "TSMC's binding constraint on H100 production in 2023 and 2024 was CoWoS advanced packaging rather than 5nm logic wafer capacity, with CoWoS originally sized at roughly 10,000 to 15,000 wafer starts per month for a niche networking and HPC market, and TSMC announced multiple CoWoS expansion phases through 2024 and into 2025 while operating margin reached a record above 47% in the fourth quarter of 2024.", "CoreWeave (CRWV) grew revenue from approximately USD 30 million in 2022 to over USD 1.9 billion in fiscal 2024 and listed publicly in early 2025, after NVIDIA deliberately cultivated a neo-cloud channel outside the hyperscaler oligopoly that included CoreWeave, Lambda, Applied Digital and Crusoe Energy through direct investment and priority GPU allocation.", "Rack power density stepped from roughly 6 to 12 kW in the CPU era to 32 kW with NVIDIA's Hopper platform and 120 kW with the Blackwell NVL72 configuration, a 4x to 10x increase across two product generations that made direct-to-chip liquid cooling a requirement rather than an option for facilities designed through 2024, when the vast majority of commercial data centres worldwide were built for air cooling only.", "Vertiv (VRT) entered early 2023 valued as a legacy industrial at roughly 10 to 12 times forward earnings with gross margin in the low 30s, revenue growth in the low to mid single digits and operating margin around 12% to 14%, and its shares appreciated approximately 8x to 9x from early 2023 to their 2024 peak as ASP-driven gross margin gains dropped through a fixed cost base." ], "cite_as": "\"Notes: The Bottleneck Heuristic - Where Scarcity Meets Capability in the AI Value Chain (Pt.1),\" Convequity, April 10, 2026, https://www.convequity.com/notes-the-bottleneck-heuristic-where-scarcity-meets-capability-in-the-ai-value-chain/", "paywall_hook": { "question_answered": "why do some companies inside a bottlenecked AI supply chain segment compound 5x to 10x while direct peers in the same segment barely outperform, and where did pricing power actually sit during the 2023 to 2024 buildout wave?", "direction": "constructive on the rare vendors able to resolve a physical constraint that competitors cannot, where volume growth, ASP expansion, operating leverage and multiple re-rating compound at once; cautious on server OEMs whose AI revenue growth came with margin compression, and more favourable on the Taiwanese ODMs performing the engineering value-add behind them.", "tickers_covered": [ "NVDA", "TSM", "MU", "CRWV", "MSFT", "VRT", "ETN", "NVT", "DELL", "HPE", "SMCI", "EQIX", "DLR", "plus SK Hynix", "Samsung", "Quanta", "Wistron and Foxconn as unlisted or foreign-listed references." ], "evidence_types": "reported gross and operating margins, disclosed capacity expansion announcements, revenue trajectories, rack power density specifications by GPU generation, data centre construction and power interconnection lead times, and a stylised operating leverage arithmetic applied to a legacy industrial cost base.", "date": "10 April 2026, Part 1 of the Bottleneck Heuristic series, covering the historical trace through the 2023 to 2024 wave.", "related_product": "companion to Convequity's AI Value Chain Framework, which segments the chain from energy generation to the application layer and applies a tailored Rule of X screen to each subsegment." } }, "ai-value-chain-bottlenecks-mapping-pricing-power-and-investment-opportunity-pt-1": { "slug": "ai-value-chain-bottlenecks-mapping-pricing-power-and-investment-opportunity-pt-1", "title": "AI Value Chain Bottlenecks: Mapping Pricing Power and Investment Opportunity (Pt.1)", "url": "https://www.convequity.com/ai-value-chain-bottlenecks-mapping-pricing-power-and-investment-opportunity-pt-1/", "published": "2026-04-02", "visibility": "public", "cohort": "opus", "model": "opus", "status": "corpus_draft_theme_json", "author": "Equity Research", "key_claims": [ "In a 2 April 2026 analysis, Convequity replaced the generic 2.3x growth coefficient that Bessemer Venture Partners derived from SaaS companies carrying 70 to 90% gross margins with subsegment-specific coefficients, each built by estimating the incremental operating margin as one minus the variable cost ratio, dividing that by a subsegment-appropriate WACC, and then applying an operating leverage realism factor of 0.4 for newer or volatile subsegments such as Neo Clouds and closer to 0.5 for mature ones such as colocation, foundry and packaging, with the resulting Rule of X ranked against EV/(FCF minus SBC).", "The 2 April 2026 analysis describes the Energy and Power Distribution bottleneck as roughly one third open in 2026, with the Blackcomb Pipeline targeting 2.5 Bcf/d of Permian takeaway capacity in the second half of 2026 while gas turbine lead times had stretched to three to four years, so transport supply was unlocking faster than generation capacity.", "Convequity's 2 April 2026 projection of the AI data centre energy mix put natural gas at around 48% in 2026, renewables and storage at about 27% and nuclear at about 18%, with gas peaking near 50% in 2027 before falling toward 38% by 2030 as renewables and storage rise toward 40% and nuclear edges up to about 20%.", "The 2 April 2026 analysis argues that the optical interconnect layer, not energy or servers, may gate the pace of the AI buildout, because the shift to 3.2T modules leaves copper viable for only about two metres before jitter becomes unacceptable, forcing optical into intra-rack GPU links, against a global shortage of indium phosphide, the substrate for transceiver lasers where AXT Inc (AXTI) holds an effective monopoly position.", "In the same 2 April 2026 analysis, Convequity places converted Bitcoin miners in the Colocation subsegment or, where they own GPUs, in Neo Cloud, noting that IREN (IREN) owns the GPUs it operates while Bitdeer (BTDR), Cipher (CIFR) and Core Scientific (CORZ) owned roughly 5 to 15% of the GPUs they operated and act primarily as power and facility hosts, with converted sites reaching usable AI capacity in months against the two to four years a traditional greenfield data centre typically takes." ], "cite_as": "\"AI Value Chain Bottlenecks: Mapping Pricing Power and Investment Opportunity (Pt.1),\" Convequity, April 2, 2026, https://www.convequity.com/ai-value-chain-bottlenecks-mapping-pricing-power-and-investment-opportunity-pt-1/", "paywall_hook": null }, "themes-sysmoore-energy-pt-3-3-the-nuclear-investment-thesis": { "slug": "themes-sysmoore-energy-pt-3-3-the-nuclear-investment-thesis", "title": "SysMoore & Energy (Pt.3.3) - The Nuclear Investment Thesis", "url": "https://www.convequity.com/themes-sysmoore-energy-pt-3-3-the-nuclear-investment-thesis/", "published": "2026-03-18", "visibility": "tiers", "cohort": "opus", "model": "opus", "status": "corpus_draft_theme_json", "author": "Equity Research", "key_claims": [ "In an 18 March 2026 analysis, Convequity records that NuScale Power (SMR), founded in 2007, held the first and still only NRC design certification for a US small modular reactor, granted in 2023, yet had brought no reactor to completion, and that its flagship Carbon Free Power Project with Utah Associated Municipal Power Systems was cancelled in November 2023 after the estimated price rose from USD 58 to USD 89 per MWh.", "The same 18 March 2026 analysis cites China's build times as the test of the modularity premise: CNNC's ACP100 \"Linglong One\" began cold functional testing in late 2025 and was expected to reach commercial operation in 2026, a total construction timeline of roughly five years and comparable to a full-sized conventional Chinese reactor, while the HTR-PM high-temperature gas-cooled plant at Shidao Bay took roughly eleven years from construction start to commercial operation.", "Centrus Energy (LEU) produced the first HALEU on US soil in approximately 70 years in November 2023 at its American Centrifuge Plant in Piketon, Ohio, according to the 18 March 2026 analysis, with a 16-centrifuge demonstration cascade targeting about 900 kilograms per year against an industry requirement of roughly 40 metric tonnes across 2025 to 2027 for initial demonstration reactors and potentially more than 150 metric tonnes annually by 2030; the same piece cites roughly USD 2.7 billion of DOE funding for domestic enrichment expansion and a USD 900 million programme awarding contracts to Centrus, General Matter and Orano.", "On the sodium-cooled fast reactor developers, the 18 March 2026 analysis records that TerraPower, founded in 2008, received its construction permit for the Natrium demonstration plant at Kemmerer, Wyoming in 2024 with commercial operation targeted for the early 2030s, while Oklo (OKLO), founded in 2013, had its first NRC combined licence application, submitted in 2020, denied in 2022 for insufficient technical detail before a revised application was accepted for review in 2024, with its heat-pipe Aurora design sized at up to 75 MWe per unit and fuelled for roughly 20 years without refuelling.", "The 18 March 2026 analysis attributes the uranium supply base's condition to the Megatons to Megawatts programme, formally the HEU Purchase Agreement, which ran from 1993 to 2013, down-blending approximately 500 metric tonnes of weapons-grade highly enriched uranium into roughly 24 million pounds of uranium equivalent per year and meeting about half of US reactor requirements before that supply left the market in 2013; it also notes that China brought the world's first operational thorium molten salt reactor, TMSR-LF1, to full 2 MWt power in the Gobi Desert in June 2024, and that Kairos Power's Hermes demonstration reactor in Oak Ridge, Tennessee targets initial operation around 2027." ], "cite_as": "\"SysMoore & Energy (Pt.3.3) - The Nuclear Investment Thesis,\" Convequity, March 18, 2026, https://www.convequity.com/themes-sysmoore-energy-pt-3-3-the-nuclear-investment-thesis/", "paywall_hook": { "question_answered": "across the nuclear value chain now being pulled forward by AI power demand, where does exposure sit with the clearest visibility and the least dependence on any single reactor vendor winning, and which of the competing SMR architectures has actually demonstrated the speed, cost and fuel-supply position its backers claim.", "direction": "cautious on light water SMR vendors, where the delivery record so far undercuts the promised speed and cost advantages, and cautious on high-temperature gas-cooled designs outside the industrial heat niche given their fuel supply exposure; relatively constructive on sodium-cooled fast reactors as the best current balance of technological risk and capability; most constructive on the fuel layer, uranium supply and domestic enrichment, on the argument that every reactor design depends on a narrow, geographically concentrated and structurally under-invested chain; molten salt and thorium framed as a long-dated risk to that thesis rather than a near-term one.", "tickers_covered": [ "SMR (NuScale)", "OKLO", "LEU (Centrus Energy)", "Kazatomprom (NATKY on OTC markets", "KAP in London)", "plus the SPUT physical uranium trust and the URNM uranium miners ETF", "alongside private and state players TerraPower", "X-energy", "Kairos Power", "General Matter", "Orano", "CNNC and Rosatom." ], "evidence_types": "NRC licensing history and dated design certifications, permit grants and application denials, cancelled project cost-per-MWh disclosures, Chinese reactor construction timelines, HALEU capacity against projected demand, DOE funding programmes and award lists, global uranium production concentration, and the 1993 to 2013 HEU Purchase Agreement supply history.", "date": "18 March 2026, Part 3.3 of the SysMoore and Energy series, concluding the nuclear segment." } }, "themes-sysmoore-energy-pt-3-2-reactor-design-fuel": { "slug": "themes-sysmoore-energy-pt-3-2-reactor-design-fuel", "title": "SysMoore & Energy (Pt.3.2) Reactor Design & Fuel", "url": "https://www.convequity.com/themes-sysmoore-energy-pt-3-2-reactor-design-fuel/", "published": "2026-03-03", "visibility": "tiers", "cohort": "opus", "model": "opus", "status": "corpus_draft_theme_json", "author": "Equity Research", "key_claims": [ "In a 3 March 2026 analysis, Convequity counted roughly eleven companies pursuing distinct small modular reactor designs, six of them publicly traded, grouped into four technology routes (light water, high-temperature gas-cooled, sodium-cooled fast and molten salt) plus a fluoride salt-cooled high-temperature category, with incumbents Rosatom, CNNC, Rolls-Royce SMR and GE Vernova-Hitachi all building on light water architecture and NuScale (SMR) the only startup on that route.", "As of the 3 March 2026 analysis, commercial HALEU enrichment capacity (5 to 20% U-235) outside Russia was essentially zero apart from Centrus Energy's small demonstration cascade at Piketon, Ohio, with US Department of Energy contracting for expansion, and adequate commercial supply judged unlikely before the late 2020s, a single constraint that delays every non-light-water design regardless of its licensing status; TRISO particle fabrication likewise had no US commercial-scale facility, with X-energy building one.", "The 3 March 2026 analysis sets out the coolant tradeoff in physical terms: pressurised water reactors run at roughly 300°C and about 155 bar with thermal-to-electric efficiency near 33%, helium-cooled HTGRs run at 700 to 950°C and roughly 60 to 70 bar for 40 to 45% efficiency plus direct industrial process heat, and sodium runs at about 550°C at atmospheric pressure with a boiling point near 883°C but reacts violently with air and water, while TRISO particles hold structural integrity to approximately 1,600°C.", "The same 3 March 2026 analysis notes that sodium-cooled fast reactor operating experience is real rather than theoretical, citing Russia's BN-600 and BN-800 and the US Experimental Breeder Reactor II at Idaho National Laboratory from 1964 to 1994, and that a fast spectrum can convert U-238 into fissile Plutonium-239, in principle extracting close to 100% of uranium's energy content against roughly 3 to 5% in conventional light water reactors while fissioning long-lived transuranics.", "On thorium, the 3 March 2026 analysis records that Th-232 is fertile rather than fissile and converts through Th-233 and Protactinium-233 (half-life about 27 days) into fissile U-233, which yields roughly 2.3 to 2.5 neutrons per thermal fission against 2.0 to 2.1 for U-235 and therefore permits thermal breeding, offset by co-produced U-232 whose Thallium-208 daughter forces remote shielded fuel fabrication; the only significant molten salt precedent cited is Oak Ridge's Molten Salt Reactor Experiment of 1965 to 1969." ], "cite_as": "\"SysMoore & Energy (Pt.3.2) Reactor Design & Fuel,\" Convequity, March 3, 2026, https://www.convequity.com/themes-sysmoore-energy-pt-3-2-reactor-design-fuel/", "paywall_hook": { "question_answered": "which small modular reactor architecture actually earns the modularisation premium once coolant, fuel, neutron spectrum and operating pressure choices are traced through to cost and deployment timing, and which constraint binds first if the US tries to close the AI-driven power gap.", "direction": "cautious on light water SMRs, where small scale sacrifices the economies of scale of a 1+ GW plant without removing legacy containment, redundancy and instrumentation cost; more constructive on SMR-native advanced architectures, with sodium-cooled fast reactors the nearest-term of them; molten salt and thorium treated as the theoretical optimum but the furthest from commercial reality; fuel supply, specifically HALEU enrichment and TRISO fabrication, identified as the bottleneck that shifts ahead of licensing.", "tickers_covered": [ "SMR (NuScale)", "OKLO", "BWXT", "GEV", "LEU (Centrus Energy)", "alongside private and state developers X-energy", "TerraPower", "Kairos Power", "Copenhagen Atomics", "Rosatom", "CNNC and Rolls-Royce SMR." ], "evidence_types": "reactor physics and engineering tradeoff tables across coolant, fuel, spectrum and pressure; operating histories of prior sodium and molten salt programmes; enrichment and fuel fabrication capacity positions; stated temperature, pressure and efficiency parameters by coolant; developer-by-developer technology route mapping.", "date": "3 March 2026, Part 3.2 of the SysMoore and Energy series." } }, "themes-sysmoore-energy-pt-3-1": { "slug": "themes-sysmoore-energy-pt-3-1", "title": "SysMoore & Energy (Pt.3.1) Nuclear Landscape", "url": "https://www.convequity.com/themes-sysmoore-energy-pt-3-1/", "published": "2026-02-21", "visibility": "tiers", "cohort": "opus", "model": "opus", "status": "corpus_draft_theme_json", "author": "Equity Research", "key_claims": [ "Convequity's 21 February 2026 nuclear landscape analysis records that Toshiba acquired Westinghouse Electric Company in 2006 for USD 5.4 billion, more than double what most analysts then considered fair value, on the thesis that nuclear was the most rational long-term path to scaled baseload energy and that Japanese industrial execution made Toshiba its natural steward.", "The same 21 February 2026 analysis traces the unwind of that bet: an accounting scandal disclosed in 2015 in which independent investigators found Toshiba had overstated operating profits by ¥151.8 billion, about USD 1.2 billion, over seven years, multi-billion-dollar write-downs on the Vogtle and V.C. Summer projects recognised by late 2016, Westinghouse's Chapter 11 filing in March 2017 and sale to Brookfield Business Partners (BBU) in 2018, the forced sale of the medical systems division to Canon in 2016 and of a majority stake in the NAND business now known as Kioxia to a Bain Capital-led consortium, and Toshiba's take-private in 2023 by a consortium led by Japan Industrial Partners after 74 years as a listed company.", "As set out in the 21 February 2026 analysis, the installed light water reactor fleet runs on uranium enriched from a natural 0.7% U-235 to roughly 3% to 5%, and extracts only about 3% to 5% of the fuel's potential energy, leaving spent fuel that is still approximately 96% uranium by mass alongside roughly 3% fission products and about 1% plutonium and minor actinides. The analysis identifies the absence of a commercial back-end fuel cycle in the United States as the structural bottleneck, with no large-scale reprocessing infrastructure unlike France and Russia, no operating permanent geological repository, and the Yucca Mountain project politically stalled for decades.", "The 21 February 2026 analysis reports that as of late 2025 China operated more than 55 commercial reactors with approximately 38 or more under active construction, more than the rest of the world combined, having imported the AP1000 from Westinghouse and the EPR from Framatome and localised both into the domestic Hualong One design now exported to Pakistan and Argentina, at estimated construction costs of roughly USD 2,500 to 3,000 per kilowatt against USD 8,000 to 12,000 or more for recent Western projects and timelines of five to six years against ten to fifteen in the United States and Europe.", "The 21 February 2026 analysis dates the shift in sentiment to nuclear power agreements signed by Microsoft (MSFT), Alphabet (GOOGL) and Amazon (AMZN) across 2024 and 2025 to supply their data centres, including Microsoft's deal to restart Three Mile Island Unit 1, which had operated safely until its economic retirement in 2019 and is separate from the Unit 2 reactor damaged in 1979, set against Germany's closure of its last three nuclear plants in April 2023 and the resulting lean on lignite coal and imported gas." ], "cite_as": "\"SysMoore & Energy (Pt.3.1) Nuclear Landscape,\" Convequity, February 21, 2026, https://www.convequity.com/themes-sysmoore-energy-pt-3-1/", "paywall_hook": { "question_answered": "can the Western nuclear industry actually execute a renaissance at the scale AI infrastructure requires, given the state of its fuel cycle, its vendor base, its regulatory process and public sentiment, and what does China's parallel buildout imply for who supplies 21st century nuclear capacity?", "direction": "cautious on Western nuclear execution capability, treating cost overruns and schedule slippage as symptoms of lost industrial learning and one-off megaproject delivery rather than a science problem, negative on the current light water reactor fuel cycle in the United States because the unresolved back end caps utilisation and lifecycle economics, constructive on the physics and safety case for Generation III+ and Generation IV designs and on standardised serial or factory-built deployment, and structurally constructive on China's position, read as a repeat of the solar manufacturing playbook.", "tickers_covered": [ "Microsoft (MSFT)", "Alphabet (GOOGL)", "Amazon (AMZN)", "Brookfield Business Partners (BBU)", "with reference to Westinghouse", "Framatome and EDF", "Toshiba", "Canon", "Rolls-Royce SMR and NuScale (SMR)." ], "evidence_types": "reactor physics and fuel cycle composition, levelized cost of electricity decomposition under a load-cost dispatch scenario including ramping costs, corporate transaction and bankruptcy history, accident causation analysis for Three Mile Island, Chernobyl and Fukushima Daiichi, comparative construction cost and schedule data for China, South Korea and the West, epidemiological comparisons of mortality per terawatt-hour, and announced hyperscaler power agreements.", "date": "21 February 2026, part 3.1 of the SysMoore and Energy series." } }, "themes-sysmoore-energy-pt-3": { "slug": "themes-sysmoore-energy-pt-3", "title": "SysMoore & Energy (Pt.3)", "url": "https://www.convequity.com/themes-sysmoore-energy-pt-3/", "published": "2026-02-11", "visibility": "tiers", "cohort": "opus", "model": "opus", "status": "corpus_draft_theme_json", "author": "Equity Research", "key_claims": [ "Convequity's 11 February 2026 analysis puts total United States electricity generation capacity at approximately 1,270 GW of utility-scale summer capacity as of late 2025, or roughly 1,330 GW including small-scale and distributed solar photovoltaics, a figure it describes as essentially flat for the previous two decades while single frontier AI data centre campuses moved past 1 GW toward roadmaps of 5 to 10 GW.", "The same 11 February 2026 analysis records that TSMC (TSM) produced approximately 17 million twelve-inch equivalent wafers in 2025, up roughly 30% from about 13 million in 2024, with around 30% of output serving high performance computing and AI, and that HPC rose to 58% of TSMC revenue in 2025 with 48% year-over-year growth after historically representing 30% to 40%.", "As of the 11 February 2026 analysis, lead times across the electricity supply chain ran two to three years for gas turbines, high-voltage transformers, switchgear and grid interconnection, and the three large-frame gas turbine suppliers, GE Vernova (GEV), Mitsubishi Power and Siemens Energy, together carried order books exceeding 200 GW with delivery timelines stretching beyond five years.", "The 11 February 2026 analysis states that China plans to spend 4 trillion RMB, approximately USD 600 billion, between 2026 and 2030 on grid upgrades to move long-distance solar and wind power, averaging roughly 200 GW of new transmission capacity per year alongside storage and EV charging, because transmission rather than generation is the binding constraint on capacity concentrated in Xinjiang.", "On the nuclear side, the 11 February 2026 analysis notes that of the two AP1000 projects started in the United States in the early 2000s, V.C. Summer in South Carolina was abandoned in 2017 with Westinghouse entering Chapter 11 the same year, while Vogtle in Georgia rose from an initial estimate of roughly USD 14 billion to more than USD 35 billion and missed its original 2016 to 2017 commercial operation dates by seven years, against a single AP1000 unit producing about 1.1 GW and taking at minimum five to seven years to build." ], "cite_as": "\"SysMoore & Energy (Pt.3),\" Convequity, February 11, 2026, https://www.convequity.com/themes-sysmoore-energy-pt-3/", "paywall_hook": { "question_answered": "if AI turns data centre demand into a continuous, inference-driven load, can the United States power system supply it, and where does the binding constraint actually sit across gas turbines, nuclear, renewables and transmission relative to China?", "direction": "cautious on the near-term United States supply position, framing the country as having a supply problem while China has a demand problem, sceptical that terrestrial nuclear or gas unit scale can match multi-gigawatt campus roadmaps on the required timeline, and structurally constructive on solar as the only resource with the physical headroom to serve AI-scale demand over the long run, extending to space-based collection as launch costs fall.", "tickers_covered": [ "TSMC (TSM)", "NVIDIA (NVDA)", "AMD (AMD)", "Alphabet (GOOGL)", "GE Vernova (GEV)", "Siemens Energy", "Mitsubishi Heavy Industries", "plus private references to xAI and SpaceX." ], "evidence_types": "national installed capacity and gross capacity addition data for the United States and China, TSMC wafer output and segment revenue mix, a first-principles power estimate derived from wafer starts, die size and per-die power draw, supply chain lead times and gas turbine order books, nuclear project cost and schedule history, Chinese grid investment plans, launch cost trajectories, and physics-based bounds on terrestrial fission and fusion energy against solar output.", "date": "11 February 2026, part three of the SysMoore and Energy series." } }, "notes-asm-and-the-rising-complexity-of-chips-why-ald-matters-more-than-ever": { "slug": "notes-asm-and-the-rising-complexity-of-chips-why-ald-matters-more-than-ever", "title": "Notes: ASM and the Rising Complexity of Chips & Why ALD Matters More Than Ever", "url": "https://www.convequity.com/notes-asm-and-the-rising-complexity-of-chips-why-ald-matters-more-than-ever/", "published": "2026-02-08", "visibility": "tiers", "cohort": "opus", "model": "opus", "status": "corpus_draft_theme_json", "author": "Equity Research", "key_claims": [ "As set out in Convequity's 8 February 2026 note on ASM International (ASM), TSMC (TSM) began volume production of its 2nm-class N2 node in the fourth quarter of calendar 2025, a ramp that raises atomic layer deposition and selective epitaxy step counts per wafer as it progresses through 2026.", "The same 8 February 2026 note records that Intel's (INTC) 18A is the company's first node to introduce gate-all-around transistors, branded RibbonFET, alongside backside power delivery, branded PowerVia, and is generally discussed as sitting in the same 2nm-class bracket as TSMC's N2 despite the differing naming conventions.", "Convequity's 8 February 2026 note states that industry pilot work for 1.4nm is expected to begin in late 2026 with production targeted for 2027, with each node transition adding further ALD and selective epitaxy steps per wafer, so the growth driver is process intensity rather than wafer volume alone.", "In the 8 February 2026 note, Convequity identifies two memory-side shifts as of that date: DRAM's move from 6F² to 4F² cell architectures, which raises aspect ratios and gap-fill and interface demands, and the expectation that from HBM4e the base die is manufactured in leading-edge logic fabs rather than memory fabs as controller and interface functions migrate into it.", "The 8 February 2026 note observes that Applied Materials (AMAT) was active in solar deposition before exiting roughly 15 years earlier under Chinese price competition, and that ASM International (ASM) has not named solar as a target market, so any contribution from heterojunction, perovskite or tandem cell manufacturing would be many years out." ], "cite_as": "\"Notes: ASM and the Rising Complexity of Chips & Why ALD Matters More Than Ever,\" Convequity, February 8, 2026, https://www.convequity.com/notes-asm-and-the-rising-complexity-of-chips-why-ald-matters-more-than-ever/", "paywall_hook": { "question_answered": "is ASM International the pure-play beneficiary of semiconductor scaling shifting from lithography toward materials precision and 3D architecture, and does that structural position justify where the shares trade going into the 2026 wafer fab equipment spending cycle?", "direction": "constructive on the multi-year structural driver, since gate-all-around logic, 4F² DRAM, HBM base-die complexity and advanced packaging all raise ALD intensity per wafer, and neutral on the shares at the time of writing, with the re-rating case conditional on ALD being recognised as a scaling bottleneck rather than one deposition tool category among several. The risk identified is timing and spending mix rather than end demand.", "tickers_covered": [ "ASM International (ASM)", "with reference points across TSMC (TSM)", "Intel (INTC)", "Samsung", "Apple (AAPL)", "Qualcomm (QCOM)", "ASML (ASML)", "KLA (KLAC) and Applied Materials (AMAT)." ], "evidence_types": "node roadmap and volume production dates, transistor architecture and backside power delivery facts, ALD and selective epitaxy step-count logic, DRAM and HBM architecture transitions, advanced packaging and hybrid bonding process requirements, management commentary on 2026 spending mix and gross margin, export control exposure, plus a discounted cash flow valuation with terminal margin sensitivity and a peer multiple comparison.", "date": "8 February 2026.", "related_product": "Convequity's DCF valuation work on ASM International, including base-case growth path, terminal free cash flow margin sensitivity and peer positioning on a Rule of X basis (product facts provisional, per Gate 4)." } }, "notes-sitime-and-the-rising-importance-of-precision-timing-in-ai-systems": { "slug": "notes-sitime-and-the-rising-importance-of-precision-timing-in-ai-systems", "title": "Notes: SiTime and the Rising Importance of Precision Timing in AI Systems", "url": "https://www.convequity.com/notes-sitime-and-the-rising-importance-of-precision-timing-in-ai-systems/", "published": "2026-01-30", "visibility": "tiers", "cohort": "opus", "model": "opus", "status": "corpus_draft_theme_json", "author": "Equity Research", "key_claims": [ "In a note dated 30 January 2026, Convequity anchors SiTime (SITM) to the company's three reported end markets, communications, enterprise and data center (CED), mobile, IoT and consumer, and automotive, industrial and defense, with CED identified as the near-term growth engine as MEMS timing content expands across GPUs, NICs, switches, optical modules and advanced packages.", "Third-party teardown analysis cited in the 30 January 2026 note shows the iPhone 16e using SiTime (SITM) MEMS-based timing in its RF, modem and connectivity platform where prior Apple (AAPL) iPhone generations relied predominantly on quartz, and Qualcomm's (QCOM) chief executive has publicly stated that Qualcomm's share of iPhone modems is expected to fall to roughly 20% in a future iPhone generation.", "SiTime (SITM) disclosed that in the three months ended 30 June 2025 its top three customers by revenue, all distributors, accounted for approximately 66% of revenue, and approximately 64% of revenue was attributable to its ten largest end customers.", "Stock-based compensation at SiTime (SITM) rose from roughly 1.7% of revenue in fiscal 2019 and 13.8% in fiscal 2020 to approximately 45% in fiscal 2024, moderating to about 33% of revenue in the most recent reported quarter as of the 30 January 2026 note, following inducement grants for newly hired strategic personnel and a substantial follow-on offering in mid-2025. Gross margin ran in the mid-50s through the two to three years to January 2026 while free cash flow margin remained negative, and management guidance implied roughly 55% revenue growth to close fiscal 2025.", "The 30 January 2026 note maps SiTime's (SITM) stack from the Titan MEMS resonator platform, which is expected to become material only later in the decade, through oscillators as the current core revenue engine to clock generators that carry higher ASPs, are chosen early in the design cycle and pull oscillator revenue through on a 12 to 18 month lag. It also notes that the 800G to 1.6T optical module transition splits links into more lanes at higher baud rates, adding clock domains for DSPs, SerDes groups and laser drivers rather than simply doubling bandwidth, which as of 2026 raises both timing content per module and the performance required of each component." ], "cite_as": "\"Notes: SiTime and the Rising Importance of Precision Timing in AI Systems,\" Convequity, January 30, 2026, https://www.convequity.com/notes-sitime-and-the-rising-importance-of-precision-timing-in-ai-systems/", "paywall_hook": { "question_answered": "is precision timing becoming a system-level constraint in AI infrastructure rather than commodity plumbing, does SiTime's MEMS position convert that into durable content growth across compute, networking and optics, and is the mobile timing shift at Apple a single-SKU event or an architectural inflection with a second, larger wave through Qualcomm?", "direction": "constructive on the structural durability of MEMS timing content growth as clock domains multiply inside advanced packages, switches, NICs and optical modules, and constructive on mobile as a genuine architectural break rather than a one-off design win; cautious on near-term multiple expansion, with the return profile framed as long-term compounding and a phased accumulation approach, and with customer concentration and historical revenue volatility flagged as live risks. The argued bottleneck shift runs from raw compute toward coordination, where synchronization error, not silicon throughput, idles expensive GPUs.", "tickers_covered": [ "SITM", "NVDA", "QCOM", "AAPL." ], "evidence_types": "company end-market reporting structure and product roadmap (Titan, Symphonic), disclosed customer concentration from the quarter ended 30 June 2025, disclosed stock-based compensation and gross margin history, management revenue guidance and consensus growth estimates, third-party smartphone teardown analysis, public statements by Qualcomm management, optical module and packaging architecture facts, and a discounted cash flow model.", "date": "30 January 2026.", "related_product": "Convequity's SITM DCF valuation model, covering the long-term growth path and terminal free cash flow margin assumption (product facts provisional, per Gate 4)." } }, "netskope-context-rich-security-for-the-ai-era-pt-2": { "slug": "netskope-context-rich-security-for-the-ai-era-pt-2", "title": "Updates: Netskope - Context-Rich Security for the AI Era (Pt.2)", "url": "https://www.convequity.com/netskope-context-rich-security-for-the-ai-era-pt-2/", "published": "2026-01-23", "visibility": "tiers", "cohort": "opus", "model": "opus", "status": "corpus_draft_theme_json", "author": "Equity Research", "key_claims": [ "Netskope (NTSK) guided to 27% year-over-year revenue growth for fiscal 4Q26, the quarter ending 31 January 2026, a sharp step down from 44% year-over-year growth in fiscal 3Q26, on a trailing twelve month revenue base of USD 661 million.", "In its 23 January 2026 analysis, Convequity describes Netskope's (NTSK) NewEdge as a multi-tenant, software-defined global network built on x86 under a single ASN and purpose-built for enterprise security rather than adapted from CDN or hyperscaler primitives, with SD-WAN acquired through Infiot running at the customer edge while SSE enforcement runs in the PoP, the same logical rather than physical convergence used by Palo Alto Networks (PANW), Cato and Versa.", "The same 23 January 2026 analysis reads Netskope's (NTSK) acquisition of Dasera as supplying the DSPM substrate of discovery, inventory and posture workflows, powered by the CASB-era data engine Netskope already had, namely a DLP and classification layer with thousands of classifiers and a Train Your Own Classifiers workflow that lets customers teach the system proprietary document types.", "As of January 2026, Palo Alto Networks (PANW) positions AI-SPM as a distinct SKU, implying that posture management for models, agents, pipelines and toolchains is becoming its own budget line, while PANW's SASE runs largely on hyperscaler infrastructure with CloudGenix SD-WAN on-premise and Fortinet (FTNT) runs SD-WAN, NGFW and SSE in the same ASIC-accelerated FortiGate device; Gartner formalized the SASE category in 2019.", "The 23 January 2026 analysis cites November 2025 reporting that a nation-state backed group used Anthropic's Claude to orchestrate an attack that was roughly 90% autonomous across about 30 enterprises, and treats it as evidence that enterprises reallocating incremental IT budget from security toward AI is unsustainable as agentic workflows spread." ], "cite_as": "\"Updates: Netskope - Context-Rich Security for the AI Era (Pt.2),\" Convequity, January 23, 2026, https://www.convequity.com/netskope-context-rich-security-for-the-ai-era-pt-2/", "paywall_hook": { "question_answered": "as AI shifts the unit of security from human access to agent execution, where does Netskope's CASB and data-centric heritage still win, what does it not yet own, and how should investors weigh that against a monetization model still tied to human seats?", "direction": "constructive on Netskope's data-centric context and on SkopeAI's inline, preventative position in the control path relative to post-event SecOps copilots, and constructive on the difficulty of replicating a real-time, data-aware control fabric; cautious on the scope expansion from workforce control into workload, agent and AI posture management, and on the timing of the transition from human seats to machine seats, where Palo Alto Networks is read as better insulated by its workload security breadth. The argued bottleneck shift runs from governing access at the point of interaction to governing chained agent actions at machine speed, with data and model foundations determining what is possible at all.", "tickers_covered": [ "NTSK", "PANW", "FTNT", "ZS", "NET", "plus private comparators Cato and Versa." ], "evidence_types": "company guidance and reported revenue base, disclosed acquisitions (Dasera, Infiot, CloudGenix), platform and network architecture facts including single-ASN versus multi-ASN design and PoP compute placement, product capability detail across SSPM, DSPM, DLP and ZTNA, a competitive framework organised by each vendor's origin (CASB-first, NGFW-first, network-first), publicly reported security incidents, and a discounted cash flow model.", "date": "23 January 2026, Part 2 of a two-part series.", "related_product": "Convequity's Netskope DCF valuation model, covering growth, terminal free cash flow margin and scenario paths (product facts provisional, per Gate 4)." } }, "notes-convequitys-lens-on-software-3-0-investing": { "slug": "notes-convequitys-lens-on-software-3-0-investing", "title": "Notes: Convequity’s Lens on Software 3.0 Investing", "url": "https://www.convequity.com/notes-convequitys-lens-on-software-3-0-investing/", "published": "2026-01-14", "visibility": "tiers", "cohort": "opus", "model": "opus", "status": "corpus_draft_theme_json", "author": "Equity Research", "key_claims": [ "In a note published 14 January 2026, Convequity records that Confluent (CFLT) agreed to sell itself to IBM (IBM) after leadership concluded it could not reaccelerate revenue growth that had decayed below 20 percent while stock-based compensation and operating expenses remained elevated.", "The same 14 January 2026 note uses SolarWinds as its reference case for legacy software inertia: after the 2020 supply-chain breach that affected thousands of enterprises and government agencies, the share price collapsed and confidence in the franchise was damaged, yet revenue kept growing at low single digits as customers stayed put, which is what later made the business attractive to private equity.", "As of the 14 January 2026 note, Cloudflare's (NET) inference offering Workers AI is described as adding one or two low-power PCIe GPUs to existing CPU servers, an architecture constrained in bandwidth, thermal capacity and physical space relative to H100, H200 and B200 class accelerators, while R2 object storage is used mainly to hold training data and move it between venues such as CoreWeave (CRWV), Oracle (ORCL) and the hyperscalers. The note also records that as of January 2026 the Cloudflare environment had no copilot at all and no basic SQL code generation for its D1 database.", "The 14 January 2026 note attributes Palantir's (PLTR) AI position to more than two decades of building Gotham and Foundry with little revenue, with those engineering toolchains becoming the backbone of AIP, and places Snowflake (SNOW), Rubrik (RBRK) and Palo Alto Networks (PANW) in a second tier on the same logic of pre-existing foundational assets rather than well-executed pivots.", "Writing on 14 January 2026, Convequity cites the Wiz founders' cloud expertise built inside Microsoft (MSFT) Azure before they created their own platform as the template for a talent diffusion cycle out of frontier labs including OpenAI, Anthropic, Google's (GOOGL) Gemini team and xAI, which it expects to productize tooling, orchestration, evaluation and deployment knowhow into a standardized PaaS layer over the following three to seven years." ], "cite_as": "\"Notes: Convequity’s Lens on Software 3.0 Investing,\" Convequity, January 14, 2026, https://www.convequity.com/notes-convequitys-lens-on-software-3-0-investing/", "paywall_hook": { "question_answered": "during an infrastructure-first phase of an AI platform transition, where across the stack does value actually accrue over the next three, seven and fifteen years, and which parts of the public SaaS universe are structurally impaired rather than cyclically cheap?", "direction": "cautious to negative on seat-priced legacy SaaS carrying high stock-based compensation, high operating expense and decelerating growth, and on AI initiatives that are peripheral features rather than rearchitecture; constructive on AI infrastructure and the semiconductor, power and networking layers near term, on the PaaS and tooling layer intermediate term, and on consumption, workload and machine-seat monetization models; the argued bottleneck shift runs from application software toward compute, energy and interconnect first, then back to an AI-native application layer that mostly does not yet exist.", "tickers_covered": [ "CFLT", "GTLB", "OKTA", "JAMF", "NET", "PANW", "FTNT", "NTSK", "PLTR", "DDOG", "SNOW", "RBRK", "MNDY", "MDB", "NVDA", "AVGO", "MRVL", "TSM", "INTC", "CDNS", "SNPS", "AMAT", "LRCX", "KLAC", "LITE", "COHR", "MTSI", "TSEM", "PWR." ], "evidence_types": "a review of more than 100 public consumer tech, enterprise tech, semiconductor and IT infrastructure companies; screened valuation, growth, margin and stock-based compensation dispersion across Convequity's software watchlists and the Bessemer Emerging Cloud Index; company product and architecture facts; disclosed acquisitions and dated events; and a structural analogy to the 1990s infrastructure build-out.", "date": "14 January 2026." } }, "updates-netskope-context-rich-security-for-the-ai-era-pt-1": { "slug": "updates-netskope-context-rich-security-for-the-ai-era-pt-1", "title": "Updates: Netskope - Context-Rich Security for the AI Era (Pt.1)", "url": "https://www.convequity.com/updates-netskope-context-rich-security-for-the-ai-era-pt-1/", "published": "2026-01-08", "visibility": "tiers", "cohort": "opus", "model": "opus", "status": "corpus_draft_theme_json", "author": "Equity Research", "key_claims": [ "Netskope (NTSK) completed its IPO in September 2025, and Convequity's 8 January 2026 update is Part 1 of a two-part series covering technology, moat and market structure, with Part 2 reserved for AI security, competitive comparisons and valuation.", "Netskope (NTSK) unveiled NewEdge, its private cloud network for carrying and inspecting enterprise security traffic, in 2019, having hired Joe De Palo in 2018 from Amazon (AMZN), where they previously led internet services, with the network engineering team assembled and the network delivered inside roughly 18 months.", "As described in the 8 January 2026 analysis, Netskope (NTSK) runs NewEdge on a single global autonomous system number, as does Cloudflare (NET), while Zscaler (ZS) operates multiple ASNs, and Palo Alto Networks (PANW) from 2019 accelerated its SASE footprint by building on Alphabet's (GOOGL) Google Cloud backbone rather than a dedicated private network.", "Netskope's (NTSK) publicly disclosed acquisitions between 2018 and 2024 are Sift Security (2018, cloud infrastructure security), New Edge Labs (2019), Trace Data (2021, data observability), WootCloud (2022, IoT and OT device zero trust), Infiot (2022, SD-WAN and cloud networking), Kadiska (2023, digital experience monitoring) and Dasera (2024, data security posture management), none of which were mega-deals by size or price.", "In the 8 January 2026 analysis, Netskope's (NTSK) data-centric enforcement layer is documented as roughly 3,000 data classifiers with support for more than 1,500 file types, plus file and binary fingerprinting, metadata analysis, exact data matching and interpretation of screenshots and images, and chief executive Sanjay Beri used the \"Layer 8\" framing on the company's most recent earnings call ahead of that publication." ], "cite_as": "\"Updates: Netskope - Context-Rich Security for the AI Era (Pt.1),\" Convequity, January 8, 2026, https://www.convequity.com/updates-netskope-context-rich-security-for-the-ai-era-pt-1/", "paywall_hook": { "question_answered": "does Netskope's CASB-first, DLP-obsessed origin plus a purpose-built private network constitute a durable moat in SASE as GenAI and agentic workflows move the unit of risk from the application to the individual transaction, and where does that model run out of road?", "direction": "constructive on Netskope's transaction-level context and data-classification depth as a compounding advantage over inline network security incumbents moving the other way, constructive on NewEdge's single-ASN and peering design as the enabler that keeps deep inline inspection usable, and cautious on two gaps: the absence of an in-house next-generation SIEM or SOC correlation layer at a time when agentic attacks are coordinated chains rather than single acts, and a control model still centred on workforce traffic rather than workload and machine-to-machine paths.", "tickers_covered": [ "NTSK", "PANW", "ZS", "NET", "FTNT", "GOOGL", "AMZN", "MNDY." ], "evidence_types": "product and architecture facts (platform structure, enforcement grades, classifier and file-type coverage), disclosed acquisition history, earnings call language, network design comparisons across single-ASN and multi-ASN operators, peering and middle-mile reasoning, customer feedback from Convequity's channel sources, and analyst positioning in SSE and SASE.", "date": "8 January 2026, Part 1 of a two-part Netskope update, following Convequity coverage that began in 2021 while the company was private." } }, "notes-light-is-the-future-pt-2": { "slug": "notes-light-is-the-future-pt-2", "title": "Notes: Light is the Future (Pt.2)", "url": "https://www.convequity.com/notes-light-is-the-future-pt-2/", "published": "2025-12-28", "visibility": "tiers", "cohort": "opus", "model": "opus", "status": "corpus_draft_theme_json", "author": "Equity Research", "key_claims": [ "As described in the 28 December 2025 analysis, a standard 800G optical transceiver spreads traffic across eight 100G channels and a 1.6T transceiver across eight 200G channels, with each lane requiring an electro-absorption modulated laser, which makes 200G EML availability the binding physical constraint on 1.6T module output into 2026 and keeps 200G supply materially tighter than 100G.", "Deployed 1.6T bandwidth depends on the switching layer as much as on the optics, because 1.6T Ethernet ports require 200G-per-lane electrical I/O and SerDes, which is why Broadcom's (AVGO) Tomahawk 6 class of switch silicon is the gating platform in 2026 and why switch system availability and qualification can cap how many 1.6T modules the market absorbs even if laser supply improves.", "Silicon photonics relieves EML tightness by moving modulators, couplers and waveguides onto a photonic integrated circuit and shifting the light source to an external continuous-wave laser rather than an EML per lane, but as of the 28 December 2025 analysis it remained a partial substitute on the hardest 800G and 1.6T AI links because of higher insertion loss, tighter packaging and thermal tolerances, and greater burden on optical power budgets, DSP and manufacturing yield.", "In co-packaged optics, Broadcom (AVGO) uses Mach-Zehnder modulators with SPIL packaging while NVIDIA (NVDA) uses micro-ring modulators with TSMC (TSM) COUPE packaging, and as of the 28 December 2025 analysis Meta (META) data centres had passed Broadcom's CPO tests with stable operation exceeding one million hours, though a single failed lane in a co-packaged assembly can force replacement of the whole unit rather than the hot-swap of one pluggable module." ], "cite_as": "\"Notes: Light is the Future (Pt.2),\" Convequity, December 28, 2025, https://www.convequity.com/notes-light-is-the-future-pt-2/", "paywall_hook": { "question_answered": "where exactly does the 2026 optical bottleneck sit once you separate laser capacity from switch-silicon availability, how far can silicon photonics act as a release valve, and when does co-packaged optics stop being a validation exercise?", "direction": "constructive on upstream laser and photonic integrated circuit suppliers, where scarcity and long equipment lead times keep allocations and pricing elevated through 2026; cautious on co-packaged optics as a near-term displacement of pluggable modules, with linear pluggable optics framed as the more credible interim route; 800G treated as transitional rather than durable as buyers migrate to 1.6T wherever supply and switch infrastructure allow.", "tickers_covered": [ "AVGO", "NVDA", "LITE", "TSEM", "COHR", "META", "GOOGL", "MSFT", "AMZN", "ORCL", "TSM." ], "evidence_types": "supplier capacity plans and expansion cycles for EML and continuous-wave lasers, die and module price points, order lead times and prepayment structures, per-customer 800G and 1.6T demand splits, penetration forecasts for silicon photonics and co-packaged optics, and packaging and modulator architecture comparisons.", "date": "28 December 2025, Part 2 of a two-part note." } }, "notes-light-is-the-future-pt-1": { "slug": "notes-light-is-the-future-pt-1", "title": "Notes: Light is the Future (Pt.1)", "url": "https://www.convequity.com/notes-light-is-the-future-pt-1/", "published": "2025-12-24", "visibility": "tiers", "cohort": "opus", "model": "opus", "status": "corpus_draft_theme_json", "author": "Equity Research", "key_claims": [ "In the first half of 2025, DeepSeek open-sourced its prefill/decode disaggregated mixture-of-experts inference architecture, built on export-compliant NVIDIA (NVDA) H800s and dynamically load-balancing experts across hundreds of GPUs, on the order of 320, which demonstrated that high-performance MoE inference can run across a scale-out Ethernet or InfiniBand fabric rather than being capped by the memory of a single NVLink domain.", "Huawei put its 384-chip CM384 \"superpod\" into production in May 2025 and has roadmapped an 8,192-chip superpod for 4Q26 (calendar 2026) and a 15,488-chip system for 4Q27 (calendar 2027), using an all-optical fabric and an n-dimensional topology to spread a single scale-up domain across multiple racks of roughly 64 kW each rather than concentrating it in one copper-bound rack.", "NVIDIA's (NVDA) GB200 NVL72 rack, as described in the 24 December 2025 analysis, comprises 36 one-rack-unit compute trays each holding one Grace CPU and two Blackwell GPUs for 72 GPUs interconnected over active-copper NVLink at roughly 120 kW per rack, against a DGX H100 rack of about four eight-GPU servers at roughly 64 kW where servers interconnect over optical Ethernet or InfiniBand.", "NVIDIA (NVDA) announced the NVL256 external NVLink Switch System at the Hopper and DGX H100 launch in March 2022, intended to link 32 nodes and 256 GH100 packages and merge scale-up with scale-out, but it never shipped meaningfully at scale, and in September 2025 NVIDIA addressed the same gap through an acqui-hire and technology-licensing deal with accelerator-networking startup Enfabrica.", "Google (GOOGL) first deployed optical circuit switches in its TPUv4 network in 2022 and reports interconnecting 9,216 TPUs using 48 OCS switches, raising mean time between failures from 2,000 hours to 100,000 hours while lifting throughput 30%, cutting power 40% and reducing capex 30%; in the TPUv7 generation described in the 24 December 2025 analysis, a rack holds 16 trays of four TPUs each, 144 racks link through OCS into a 9,216-chip pod, and up to four regions of four pods give a maximum cluster of 147,456 chips." ], "cite_as": "\"Notes: Light is the Future (Pt.1),\" Convequity, December 24, 2025, https://www.convequity.com/notes-light-is-the-future-pt-1/", "paywall_hook": { "question_answered": "if DeepSeek's cluster inference, Huawei's optics-first superpods and Broadcom's ASIC and TPU wins have broken the copper-first consensus, where does the value migrate across the optical stack, and how large is the 800G and 1.6T module requirement implied by 2026 accelerator output?", "direction": "constructive on the optical interconnect chain across modules, DSPs, EML lasers, silicon photonics, optical circuit switches and switch silicon, on the view that investors anchored to NVIDIA's copper-heavy narrative have under-modelled optics; cautious on NVIDIA's leverage from ever-larger NVLink scale-up domains, which is argued to be eroding as cluster-native inference, open Ethernet scale-out and pooled memory reduce the rationale for keeping compute inside one copper-bound node.", "tickers_covered": [ "NVDA", "AVGO", "GOOGL", "META", "LITE", "TSEM", "COHR", "MRVL", "CLS", "SMCI", "INTC", "TSM", "ORCL", "MSFT", "AMZN", "AAPL." ], "evidence_types": "cluster and rack architecture teardowns for NVL72, CM384 and TPUv7; published Google performance figures for optical circuit switching; accelerator packaging capacity as the anchor for module demand; chip-to-transceiver ratio modelling by platform; supply-chain checks on module, laser and silicon photonics capacity, allocations and pricing; and vendor share estimates across the optical value chain.", "date": "24 December 2025, Part 1 of a two-part note." } }, "updates-palo-alto-networks-from-platformization-to-ai-native-security-observability-pt-2": { "slug": "updates-palo-alto-networks-from-platformization-to-ai-native-security-observability-pt-2", "title": "Updates: Palo Alto Networks -From Platformization To AI-Native Security & Observability (Pt.2)", "url": "https://www.convequity.com/updates-palo-alto-networks-from-platformization-to-ai-native-security-observability-pt-2/", "published": "2025-12-16", "visibility": "tiers", "cohort": "opus", "model": "opus", "status": "corpus_draft_theme_json", "author": "Equity Research", "key_claims": [ "Palo Alto Networks (PANW) acquired enterprise browser vendor Talon Cyber Security in November 2023 and has folded it into Prisma SASE, where it enforces in-session policy on managed and unmanaged devices, covering which SaaS and internal apps open, what can be downloaded or uploaded, whether copy/paste and screenshots are permitted, and whether sessions are watermarked or recorded.", "As of the 16 December 2025 analysis, Palo Alto Networks (PANW) delivers Prisma SASE from a global point-of-presence fabric that rides largely on Google's backbone rather than PANW-operated data centres, and runs VM-Series and CN-Series software firewalls in cloud and Kubernetes estates including virtual firewalls on NVIDIA (NVDA) BlueField DPUs for offloaded east-west inspection, whereas Fortinet (FTNT) concentrates enforcement in FortiGate and FortiSwitch appliances built on in-house NP-class security ASICs.", "Chronosphere, which Palo Alto Networks (PANW) is bringing in-house as of December 2025, is built on the M3 open-source time-series engine originally developed at Uber and had itself acquired Calyptia in early 2024 to absorb the team behind Fluentd and Fluent Bit, giving PANW both the in-flight telemetry pipeline and the Kubernetes-centric metrics store that feeds it.", "Palo Alto Networks (PANW) has a quantum-safe readiness partnership with IBM (IBM) that pairs IBM's post-quantum cryptography and crypto-inventory tooling with PANW's inline control points and the QRadar customer base and telemetry acquired in its SIEM asset purchase and now migrating into Cortex/XSIAM, and as of the 16 December 2025 analysis PANW was already shipping quantum-safe algorithms across hardware, software and SASE products under the \"Orion\" banner.", "Palo Alto Networks (PANW) management raised its next-generation security ARR target for fiscal 2030 from USD 15 billion to USD 20 billion following the CyberArk and Chronosphere deals and has guided to a 40% free cash flow margin by fiscal 2028; as of the 16 December 2025 analysis NGS ARR stood at roughly USD 6 billion, trailing twelve-month revenue was approaching USD 9.5 billion and trailing free cash flow margin was about 38%." ], "cite_as": "\"Updates: Palo Alto Networks -From Platformization To AI-Native Security & Observability (Pt.2),\" Convequity, December 16, 2025, https://www.convequity.com/updates-palo-alto-networks-from-platformization-to-ai-native-security-observability-pt-2/", "paywall_hook": { "question_answered": "what do CyberArk, Talon, Chronosphere and Protect AI together do to Palo Alto Networks' position as connectivity shifts from user-to-resource toward agent-to-resource, and how does a software-first control plane compare with Fortinet's ASIC-first inline enforcement model as AI traffic scales?", "direction": "constructive on Palo Alto Networks as the software control plane for identity, browser, cloud, SecOps and now observability, with observability treated as a security-plus-utility motion rather than a bolt-on; cautious on integration execution and on PANW's structural per-bit cost disadvantage versus Fortinet in high-density inline inspection; reads the market as under-crediting PANW's role in governing human and agent activity rather than as fully pricing it.", "tickers_covered": [ "PANW", "FTNT", "NVDA", "IBM", "GOOGL", "DDOG", "CFLT", "CSCO", "MSFT", "PLTR", "SNOW." ], "evidence_types": "product and architecture facts across Prisma SASE, Strata, Cortex/XSIAM, Talon, Chronosphere with M3 and Fluent Bit, and the CN-Series and VM-Series software firewalls; acquisition history and dated vendor partnerships; management TAM framing, ARR targets and margin guidance; peer comparison of hardware-accelerated versus software-delivered enforcement economics; and multi-scenario discounted cash flow work.", "date": "16 December 2025, Part 2 of a two-part update on Palo Alto Networks.", "related_product": "Convequity's PANW discounted cash flow model, which subscribers can adjust for their own growth path and terminal free cash flow margin assumptions (product facts provisional, per Gate 4)." } }, "updates-palo-alto-networks-from-platformization-to-ai-native-security-observability-pt-1": { "slug": "updates-palo-alto-networks-from-platformization-to-ai-native-security-observability-pt-1", "title": "Updates: Palo Alto Networks -From Platformization To AI-Native Security & Observability (Pt.1)", "url": "https://www.convequity.com/updates-palo-alto-networks-from-platformization-to-ai-native-security-observability-pt-1/", "published": "2025-12-13", "visibility": "tiers", "cohort": "opus", "model": "opus", "status": "corpus_draft_theme_json", "author": "Equity Research", "key_claims": [ "In its fiscal first quarter of 2026, reported ahead of Convequity's 13 December 2025 update, Palo Alto Networks (PANW) grew next-generation security annual recurring revenue to close to USD 6 billion at just under 30% year over year, and chief executive Nikesh Arora raised the fiscal 2030 NGS ARR target from USD 15 billion to USD 20 billion.", "Palo Alto Networks (PANW) disclosed for its fiscal first quarter of 2026 a SASE win worth USD 33 million and an XSIAM deal of around USD 100 million, with average contract duration stretching to roughly three years from a little over two, total remaining performance obligations growing in the mid-20s percent against reported revenue growth of 16%, a back-end loaded profile the 13 December 2025 analysis reads as evidence that platform contracts are replacing point-product purchases.", "As of the fiscal first quarter of 2026, almost half of Palo Alto Networks' (PANW) trailing twelve month product revenue came from software form factors, up from the high 30s percent roughly a year earlier, driven by software firewalls and PAN-OS based SD-WAN sold heavily into SaaS and AI-native startups.", "Palo Alto Networks (PANW) agreed to acquire CyberArk (CYBR) for roughly USD 25 billion, mostly in stock, the largest acquisition in its history and an order of magnitude above prior deals, alongside Protect AI, since folded into Prisma AIRS, and Chronosphere for Kubernetes-scale observability and telemetry pipelining built on M3 and Fluent Bit; CyberArk had itself paid roughly 10 times annual recurring revenue for machine-identity specialist Venafi in 2024.", "Anthropic disclosed in November 2025 a largely autonomous cyber-espionage campaign run with a jailbroken version of Claude Code and Model Context Protocol tooling, in which the agent performed an estimated 80% to 90% of the kill chain across reconnaissance, exploitation, credential harvesting, lateral movement, exfiltration and log cleanup, an event the 13 December 2025 analysis treats alongside the Google (GOOGL) Antigravity integrated development environment incident, where a Gemini 3 Pro agent asked to clear a project cache recursively deleted the contents of a drive, as the forcing function for identity-gated agent governance." ], "cite_as": "\"Updates: Palo Alto Networks -From Platformization To AI-Native Security & Observability (Pt.1),\" Convequity, December 13, 2025, https://www.convequity.com/updates-palo-alto-networks-from-platformization-to-ai-native-security-observability-pt-1/", "paywall_hook": { "question_answered": "is the serial-acquirer discount on Palo Alto Networks justified by integration risk, or is the company assembling the identity, observability and AI-security control plane that agentic workloads will require, and what would have to show up in the numbers for that discount to flip into an AI rerating?", "direction": "constructive on the strategic architecture, in particular the bet that privileged and machine identity becomes the choke point for agentic AI and that security converges with operational utility; cautious on execution, integration debt and the near-term overhang from the size of the CyberArk transaction; the shares are read as mispriced against AI-rerated software peers because the market is crediting the consolidation story and discounting the AI enablement story.", "tickers_covered": [ "PANW", "CYBR", "OKTA", "CRWD", "FTNT", "SNOW", "PLTR", "GOOGL." ], "evidence_types": "quarterly results and annual recurring revenue disclosures, remaining performance obligation and contract duration metrics, product revenue mix by form factor, announced acquisition terms and precedent identity transaction multiples, vendor product releases and rebrands including Cortex Cloud, Darwin, XSIAM for Cloud, AgentiX and Prisma AIRS, published incident disclosures from Anthropic and reporting on the Google Antigravity failure, and Convequity's reading of customer and channel feedback on cloud security visibility.", "date": "13 December 2025, Part 1 of a two-part update." } }, "updates-fortinet-the-convergence-of-network-security-and-the-physical-world": { "slug": "updates-fortinet-the-convergence-of-network-security-and-the-physical-world", "title": "Updates: Fortinet - The Convergence of Network, Security, and the Physical World", "url": "https://www.convequity.com/updates-fortinet-the-convergence-of-network-security-and-the-physical-world/", "published": "2025-11-25", "visibility": "tiers", "cohort": "opus", "model": "opus", "status": "corpus_draft_theme_json", "author": "Equity Research", "key_claims": [ "In a 25 November 2025 update, Convequity argued that Fortinet's (FTNT) differentiation rests on a single operating system, FortiOS, running across next-generation firewalls, SD-WAN, SASE, switches, access points and endpoints with one telemetry model, policy language and control plane, an organic build that contrasts with the multi-codebase estates left by acquisition-led peers such as Palo Alto Networks (PANW), Cisco (CSCO) and Check Point (CHKP).", "Gartner's 2025 Magic Quadrant for SASE Platforms requires branch appliances with performance-based traffic steering and explicitly incorporates private backbones and hybrid topologies, a redefinition that Convequity's 25 November 2025 update reads as confirming the hybrid convergence view it published in 2023 rather than the cloud-only SASE model promoted by vendors such as Zscaler (ZS).", "As of the 25 November 2025 update, Fortinet's (FTNT) operational technology security business exceeded USD 1 billion in revenue and was growing around 30% year over year, a franchise Convequity positions as the enforcement layer beneath discovery-led OT specialists such as Claroty, Armis and Dragos, because agentless industrial devices cannot host endpoint software and must be constrained at the network instead.", "Fortinet (FTNT) plans to ship FortiGate appliances powered by its new NP8 security processor during 2026, succeeding the NP7 generation, and Convequity's 25 November 2025 update expects the performance-per-watt step up to be the hardware trigger for an AI-driven customer premises equipment refresh comparable to the SD-WAN and Zero Trust upgrade wave of 2021 to 2023.", "As disclosed ahead of the 25 November 2025 update, Fortinet (FTNT) expanded its buyback authorisation by USD 1 billion after deploying roughly USD 1.8 billion on repurchases in the latest quarter, its largest quarterly buyback to date, taking cumulative repurchases since IPO to nearly USD 10 billion and retiring about 40% of shares outstanding." ], "cite_as": "\"Updates: Fortinet - The Convergence of Network, Security, and the Physical World,\" Convequity, November 25, 2025, https://www.convequity.com/updates-fortinet-the-convergence-of-network-security-and-the-physical-world/", "paywall_hook": { "question_answered": "does Fortinet's single-OS, ASIC-accelerated architecture still confer a structural edge as networking, security and compute converge across cloud, edge and data centre, and does the growth trajectory embedded in the share price as of 25 November 2025 match the evidence from SASE, operational technology and the emerging physical AI footprint?", "direction": "constructive on the durability of Fortinet's architectural advantage and on its position in both securing AI workloads and AI-enabling security operations, with the prevailing market assumption of permanent low-gear growth read as mispriced against the SASE, OT and edge refresh evidence; cautious on the cyclicality of appliance demand, which the analysis flags as the reason any product revenue acceleration should not be extrapolated as a new paradigm.", "tickers_covered": [ "FTNT", "PANW", "CSCO", "CHKP", "ZS", "CRWD", "AVGO." ], "evidence_types": "company product and architecture disclosures (FortiOS, FortiASIC security processing units, FortiSwitch, FortiSASE, the FortiAI pillars), Gartner's 2025 SASE Platforms Magic Quadrant criteria, disclosed OT revenue scale and growth, installed base and customer count disclosures, buyback authorisations and repurchase history, an OT security maturity framework, and Convequity's discounted cash flow scenario work.", "date": "25 November 2025.", "related_product": "Convequity's interactive Fortinet discounted cash flow valuation model, which allows subscribers to adjust growth, margin and terminal inputs directly (product facts provisional, per Gate 4)." } }, "notes-ai-bubble-refining-the-forward-roic": { "slug": "notes-ai-bubble-refining-the-forward-roic", "title": "Notes: AI Bubble? Refining the Forward ROIC", "url": "https://www.convequity.com/notes-ai-bubble-refining-the-forward-roic/", "published": "2025-11-17", "visibility": "public", "cohort": "opus", "model": "opus", "status": "corpus_draft_theme_json", "author": "Equity Research", "key_claims": [ "In a 17 November 2025 follow-on note, Convequity narrowed its forward return on invested capital framework to Microsoft (MSFT), Amazon (AMZN) and Alphabet (GOOGL), excluding Apple (AAPL), whose remaining performance obligations were USD 45 billion as of 2Q 2025 and tied to consumer services, and Meta (META), whose USD 33 billion backlog reflects an advertising model in which AI spend supports internal ranking and recommendation rather than external revenue commitments.", "The 17 November 2025 model defines invested capital as gross incremental capex over 2024 and 2025 of USD 455.3 billion, attributes 80% of it to AI based on third-quarter 2025 earnings call and filing commentary from the three companies, and deliberately does not net off accumulated depreciation because that cost already flows through the operating margin used to derive NOPAT.", "On the returns side, the 17 November 2025 note attributes 41% of the group's USD 749.7 billion of remaining performance obligations to AI, by treating roughly 20% of the USD 436 billion third-quarter 2023 base as AI-related and 70% of the USD 313.7 billion of subsequent growth as AI-driven, then scales the resulting USD 307.4 billion by 1.43 times to capture on-demand consumption that never enters RPO.", "Applying a blended 30% operating margin and a 21% tax rate, the 17 November 2025 model produces about USD 104 billion of two-year NOPAT on USD 364 billion of AI-specific capex, a two-year forward ROIC of 28.6% against a two-year hurdle of roughly 16% derived from an 8% annual cost of capital, down from the 33.8% annualised figure in the earlier version of the model.", "The 17 November 2025 note cites disclosed cloud profitability of about 34% operating margin at AWS in the third quarter of 2025 and about 24% at Google Cloud, with Azure estimated higher, and describes the committed versus on-demand revenue split as roughly 80/20 at Microsoft, 70% to 80% committed at Google Cloud, and historically 60/40 at AWS trending toward 70/30 as AI capacity guarantees lengthen contracts." ], "cite_as": "\"Notes: AI Bubble? Refining the Forward ROIC,\" Convequity, November 17, 2025, https://www.convequity.com/notes-ai-bubble-refining-the-forward-roic/", "paywall_hook": null }, "notes-ai-bubble-spending-risks-and-returns-a-high-level-assessment": { "slug": "notes-ai-bubble-spending-risks-and-returns-a-high-level-assessment", "title": "Notes: AI Bubble? Spending, Risks, and Returns — A High-Level Assessment", "url": "https://www.convequity.com/notes-ai-bubble-spending-risks-and-returns-a-high-level-assessment/", "published": "2025-11-14", "visibility": "tiers", "cohort": "opus", "model": "opus", "status": "corpus_draft_theme_json", "author": "Equity Research", "key_claims": [ "In a 14 November 2025 assessment, Convequity noted that AI has pushed hyperscaler capital expenditure budgets from the USD 50 billion to USD 100 billion range to well above USD 200 billion annually for the largest players, with Amazon (AMZN), Alphabet (GOOGL), Microsoft (MSFT) and Meta (META) projected to exceed USD 350 billion collectively in 2025 and companies allocating a record 60% of operating cash flow to capex.", "As of the 14 November 2025 note, Oracle (ORCL) had entered a USD 300 billion, five-year compute commitment with OpenAI, was undercutting hyperscaler GPU instance pricing by 30% to 50%, had compressed its AI infrastructure-as-a-service gross margin to about 16% near term, and expected margins to recover toward 35% to 40% or better as deals mature.", "The 14 November 2025 note records OpenAI's 2025 funding and silicon arrangements as a USD 40 billion Series F in March 2025, a USD 100 billion direct investment from NVIDIA (NVDA), USD 14 billion prepaid to Broadcom (AVGO) for two generations of custom ASICs with an initial 0.6 GW deployment of the \"Titan\" chip hosted by Oracle (ORCL), and an AMD (AMD) partnership beginning with 1 GW in its first phase.", "Chinese open-source labs cut inference unit costs materially during 2025, with DeepSeek V3.2 using sparse attention to drop its output token price to USD 0.42 per million tokens on roughly a 62% cost reduction, and Qwen3-Next combining a hybrid linear and standard attention mix with ultra-sparse mixture-of-experts routing that activates 3 billion of 80 billion parameters.", "As of the 14 November 2025 note, CoreWeave (CRWV) carried roughly USD 50 billion of remaining performance obligations against a market capitalisation near USD 38 billion and USD 4.3 billion of trailing twelve-month revenue at 74% gross margin, a backlog it can only serve by building ahead of consumption." ], "cite_as": "\"Notes: AI Bubble? Spending, Risks, and Returns — A High-Level Assessment,\" Convequity, November 14, 2025, https://www.convequity.com/notes-ai-bubble-spending-risks-and-returns-a-high-level-assessment/", "paywall_hook": { "question_answered": "is the 2025 AI capital expenditure cycle a bubble, and if a break comes, which layer of the stack breaks first, inference capacity or frontier training?", "direction": "cautiously bullish overall, with the bubble risk located in frontier training rather than inference; constructive on companies whose demand comes from disciplined enterprise and hyperscaler buyers, cautious on the neo-cloud and aggressive-buyer cohort where the same demand is financed with weaker balance sheets, and the locus of marginal value creation is read as shifting downstream from infrastructure toward application builders as token costs fall.", "tickers_covered": [ "NVDA", "PLTR", "CRWV", "ORCL", "MSFT", "AMZN", "GOOGL", "META", "AAPL", "AVGO", "AMD", "SMCI", "plus private labs OpenAI", "Anthropic", "DeepSeek and Qwen." ], "evidence_types": "hyperscaler capex and remaining performance obligation disclosures, disclosed and reported cloud operating margins, GPU rental and API price points, published model architecture and efficiency claims, announced financing and partnership structures, utilisation and grid capacity estimates, server depreciation schedule changes, and a top-down forward return on invested capital estimate built by Convequity.", "date": "14 November 2025." } }, "lumentum-holdings-the-optical-engine-behind-the-ai-data-center-revolution": { "slug": "lumentum-holdings-the-optical-engine-behind-the-ai-data-center-revolution", "title": "Lumentum Holdings: The Optical Engine Behind the AI Data-Center Revolution", "url": "https://www.convequity.com/lumentum-holdings-the-optical-engine-behind-the-ai-data-center-revolution/", "published": "2025-11-06", "visibility": "tiers", "cohort": "opus", "model": "opus", "status": "corpus_draft_theme_json", "author": "Equity Research", "key_claims": [ "Lumentum Holdings (LITE) reported record fiscal Q1 2026 revenue of USD 533.8 million, up 58% year over year, and guided the following quarter to USD 630 to 670 million, a range that would clear its stated USD 600 million per quarter target two quarters ahead of plan, with cloud and AI infrastructure accounting for more than 60% of revenue as of the 6 November 2025 report.", "Lumentum's (LITE) year-over-year revenue growth accelerated through calendar 2025, running at 10% in the December 2024 quarter, 16% in March 2025, 55% in June 2025 and 58% in September 2025, tracking the hyperscaler shift from copper scale-up interconnect to optical scale-out.", "As of November 2025, Lumentum (LITE) was the only supplier shipping 200 Gbps-per-lane electro-absorption modulated lasers (EMLs) in volume, the laser engines inside 1.6T transceivers, a process that requires seven layers of indium phosphide epitaxy and therefore turns on yield control rather than design novelty.", "Lumentum (LITE) changed its segment reporting in fiscal Q1 2026 from Cloud & Networking and Industrial to Components and Systems, following the 2018 acquisition of Oclaro, which deepened its indium phosphide platform, and the 2023 acquisition of CloudLight, which added transceiver module manufacturing; management guided to roughly 40% to 50% EML capacity expansion over the following four to five quarters and expects a 5 to 6 percentage point margin uplift once CloudLight is supplied with in-house lasers.", "The 6 November 2025 analysis dates the architectural trigger to the first half of 2025, when DeepSeek published its open-sourced prefill/decode inference cluster and Huawei demonstrated the all-optical CM384 pod, alongside Google's stated plan to make its TPU clusters fully optical within two years and NVIDIA's (NVDA) 2025 acquisition of Enfabrica, with industry estimates for calendar 2026 1.6T transceiver shipments rising from roughly 10 million units early in 2025 to about 20 million by mid-year and near 35 million by November 2025." ], "cite_as": "\"Lumentum Holdings: The Optical Engine Behind the AI Data-Center Revolution,\" Convequity, November 6, 2025, https://www.convequity.com/lumentum-holdings-the-optical-engine-behind-the-ai-data-center-revolution/", "paywall_hook": { "question_answered": "is Lumentum's reacceleration a cyclical optics upturn or a structural repricing of the physical layer, and does its move from laser components up into transceiver modules, optical circuit switching and co-packaged optics justify the growth the market is already embedding as of 6 November 2025?", "direction": "constructive on the durability of demand for 200G-per-lane EMLs and on the margin path from vertical integration, cautious on valuation because the current price already discounts a high sustained growth rate, with hyperscaler customer concentration named as the main fragility.", "tickers_covered": [ "LITE", "COHR", "AVGO", "MRVL", "NVDA", "GOOGL", "plus unlisted module makers InnoLight and Eoptolink." ], "evidence_types": "quarterly results and forward revenue guidance, segment reporting changes, acquisition history, management capacity and margin targets, investor day materials, transceiver generation roadmaps, materials science comparison of indium phosphide against silicon photonics and VCSELs, and export control history affecting China-facing sales.", "date": "6 November 2025, written immediately after the fiscal Q1 2026 print.", "related_product": "Convequity's interactive DCF model, with adjustable revenue growth and terminal free cash flow margin inputs (product facts provisional, per Gate 4)." } }, "notes-pagaya-a-fintech-we-want-to-believe-in-and-finally-might-be-able-to": { "slug": "notes-pagaya-a-fintech-we-want-to-believe-in-and-finally-might-be-able-to", "title": "Notes: Pagaya, A Fintech We Want to Believe In — and Finally Might Be Able To", "url": "https://www.convequity.com/notes-pagaya-a-fintech-we-want-to-believe-in-and-finally-might-be-able-to/", "published": "2025-10-24", "visibility": "tiers", "cohort": "opus", "model": "opus", "status": "corpus_draft_theme_json", "author": "Equity Research", "key_claims": [ "As described in a 24 October 2025 analysis, Pagaya Technologies (PGY) does not source borrowers itself but integrates into lenders' origination systems so that a declined application can be re-underwritten and routed to institutional buyers, a process the company calls \"decline monetization\", drawing on data from 31 lending partners and placing paper with more than 150 institutional ABS investors including bank and fintech originators such as SoFi (SOFI), LendingClub (LC), Klarna, Ally Financial (ALLY) and Westlake Financial.", "Pagaya (PGY) signed a USD 2.4 billion forward-flow agreement with Blue Owl Capital (OWL) in February 2025 and a further agreement with Castlelake in July 2025 to purchase up to USD 2.5 billion of consumer personal loans over 16 months, structures in which the investor buys whole loan pools rather than securitization tranches, which removes Pagaya's role as ABS sponsor and with it the Dodd-Frank 5% risk-retention obligation on that volume.", "Pagaya (PGY) retains its securitization exposure through the off-balance-sheet Opportunity Fund, which sat at roughly USD 870 million in the second quarter of 2025 and holds the horizontal first-loss unrated equity tranches at levels well above the 5% Dodd-Frank minimum; recorded impairments in that fund rose through 2024 from USD 53 million in the first quarter to USD 235 million in the fourth quarter, totalling about USD 442 million for full-year 2024, then fell to USD 24 million in the first quarter of 2025 and USD 15 million in the second quarter of 2025.", "Pagaya (PGY) reported second-quarter 2025 revenue of USD 326 million against prior guidance of USD 290 million to USD 310 million, with fee revenue of USD 318 million on USD 2.3 billion of ABS issuance across six deals, compared with full-year 2024 fee revenue of about USD 1.005 billion on roughly USD 6 billion of issuance and only USD 67 million of operating cash flow for that year.", "Iceberg Research published a critique of Pagaya (PGY) in February 2025 that took the Opportunity Fund's last-twelve-month interest income of USD 20.1 million on a USD 923 million portfolio balance as of the third quarter of 2024 and argued the retained tranches were worth a fraction of their carrying value, while also raising questions about the late-2024 acquisition of Theorem, a fund with roughly USD 2 billion of assets under management, and about redemption restrictions reported by Israeli media that limited Opportunity Fund limited partners to withdrawing 8% of principal per quarter." ], "cite_as": "\"Notes: Pagaya, A Fintech We Want to Believe In — and Finally Might Be Able To,\" Convequity, October 24, 2025, https://www.convequity.com/notes-pagaya-a-fintech-we-want-to-believe-in-and-finally-might-be-able-to/", "paywall_hook": { "question_answered": "has Pagaya's 2024 first-loss blow-up in the Opportunity Fund actually been resolved by the 2025 shift toward forward-flow whole-loan buyers and cleaner vintages, or does the improvement rest on accounting and disclosure choices that leave the same asymmetry in place?", "direction": "constructive on the structural position Pagaya occupies between lender origination and the structured-credit bid, and on the funding mix shifting away from sponsor-retained equity tranches, but cautious on governance, disclosure and the question of whether the underwriting edge is genuine model superiority or simply network scale.", "tickers_covered": [ "PGY", "UPST", "SOFI", "LC", "AFRM", "CVNA", "KMX", "ALLY", "TREE", "OWL", "KKR." ], "evidence_types": "company shareholder letters and investor presentations, quarterly Opportunity Fund balances and impairment disclosures, dated forward-flow agreement announcements, ABS issuance counts and deal sizes, operating cash flow and receivables data used to test fee quality and days sales outstanding, Dodd-Frank risk-retention rules, a published short-seller report, and Israeli press reporting on fund redemptions.", "date": "24 October 2025.", "related_product": "Convequity's discounted cash flow model for Pagaya, with adjustable growth, free cash flow margin and stock-based compensation inputs (product facts provisional, per Gate 4)." } }, "updates-rubrik-building-a-data-centric-cyber-resilience-moat-pt-2": { "slug": "updates-rubrik-building-a-data-centric-cyber-resilience-moat-pt-2", "title": "Updates: Rubrik - Building a Data-Centric Cyber Resilience Moat (Pt.2)", "url": "https://www.convequity.com/updates-rubrik-building-a-data-centric-cyber-resilience-moat-pt-2/", "published": "2025-10-19", "visibility": "tiers", "cohort": "opus", "model": "opus", "status": "corpus_draft_theme_json", "author": "Equity Research", "key_claims": [ "Rubrik (RBRK) released Annapurna in December 2024 in partnership with AWS Bedrock and took it to general availability in April 2025, exposing embeddings as a service rather than metadata alone, so that AI orchestrators and foundation models retrieve from backup copies while Rubrik handles sensitive-data suppression, access control and indexing with its own embedding model.", "For fiscal 3Q26, the quarter ending 31 October 2025, Rubrik (RBRK) guided to 32% year-over-year revenue growth against 51% year-over-year growth reported in fiscal 2Q26, attributing the step down to its transition toward a fully SaaS model and deal-timing effects among large enterprise customers, while management expected stock-based compensation to stay near 20% of revenue.", "As described in this 19 October 2025 note, Rubrik (RBRK) uses fixed-block deduplication at block sizes on the order of 4 MB, which rewrites a whole block on any change but keeps hashing and comparison cheap, whereas Commvault (CVLT) and Cohesity use content-defined variable-length chunking that stores only changed portions at the cost of higher CPU and memory during backup; the trade-off favours recovery speed over storage efficiency.", "Rubrik (RBRK) markets a USD 10 million ransomware recovery warranty, larger and earlier than Veeam's warranty of up to USD 5 million as of this 19 October 2025 note, and pairs it with deep two-vendor integrations into Zscaler (ZS) and CrowdStrike (CRWD) for double-extortion coverage rather than broad API integration across many security vendors.", "Veeam, Rubrik's (RBRK) largest private competitor in modern data protection, was acquired by Insight Partners in 2020 for approximately USD 5 billion after incorporating in the Netherlands, and in 2024 Insight led a USD 2 billion secondary transaction that brought in TPG, Temasek and Neuberger Berman at a valuation of roughly USD 15 billion, with revenue past USD 2 billion by the time of this October 2025 note." ], "cite_as": "\"Updates: Rubrik - Building a Data-Centric Cyber Resilience Moat (Pt.2),\" Convequity, October 19, 2025, https://www.convequity.com/updates-rubrik-building-a-data-centric-cyber-resilience-moat-pt-2/", "paywall_hook": { "question_answered": "is the shift from disaster recovery to hack recovery durable enough to make Rubrik an independent cybersecurity franchise in the data domain, and does its DSPM and GenAI retrieval optionality justify a re-rating relative to where the market was pricing growth in October 2025?", "direction": "constructive on Rubrik's position in cyber resilience, on the widening go-to-market lead over Cohesity and on Annapurna as call-option optionality; cautious on the DSPM product itself, where go-to-market execution and mindshare are doing more work than technical differentiation against preventive-security specialists; cautious near term on fiscal 3Q26 seasonality and macro-driven deferral of data-management projects.", "tickers_covered": [ "RBRK", "PANW", "CRWD", "ZS", "FTNT", "CVLT", "CYBR", "PLTR", "SNOW." ], "evidence_types": "company guidance and reported growth rates, product and architecture facts on deduplication, indexing and embedding services, dated product launches and general availability, private-market ownership and funding history for Veeam and RAG peers, partner-programme recognition including Microsoft Co-sell Partner of the Year, customer-cited reliability and automation feedback, and a scenario-based discounted cash flow framework.", "date": "19 October 2025, an update to earlier Rubrik coverage.", "related_product": "an interactive discounted cash flow model for RBRK with adjustable key parameters, plus RBRK financial charts (product facts provisional, per Gate 4)." } }, "updates-rubrik-building-a-data-centric-cyber-resilience-moat-pt-1": { "slug": "updates-rubrik-building-a-data-centric-cyber-resilience-moat-pt-1", "title": "Updates: Rubrik - Building a Data-Centric Cyber Resilience Moat (Pt.1)", "url": "https://www.convequity.com/updates-rubrik-building-a-data-centric-cyber-resilience-moat-pt-1/", "published": "2025-10-12", "visibility": "tiers", "cohort": "opus", "model": "opus", "status": "corpus_draft_theme_json", "author": "Equity Research", "key_claims": [ "Convequity's 12 October 2025 update traces the divergence between the two next-generation data protection vendors to their founding premises: Cohesity entered in 2013 with a broad secondary-data thesis in which backup copies become a base for analytics, search and file sharing, while Rubrik (RBRK) entered in 2014 and concentrated on security and rapid restoration, moving the category from disaster recovery toward cyber resiliency.", "As described in the 12 October 2025 update, Rubrik (RBRK) uses fixed-block deduplication with blocks on the order of 4 MB, which rewrites a whole block when any part of it changes but makes hashing and comparison straightforward and therefore favours backup speed and predictable recovery performance, whereas Commvault (CVLT) and Cohesity use variable-length, content-boundary chunking that is more storage-efficient at higher CPU and memory cost, with Cohesity's scale-out parallelism partly offsetting that penalty.", "The 12 October 2025 update notes that Rubrik's (RBRK) USD 10 million ransomware recovery warranty was the earliest, largest and most heavily marketed in the category, against Veeam's warranty of up to USD 5 million, and treats it as both a go-to-market differentiator and an entry point into security buyers.", "Convequity's 12 October 2025 update argues that Palo Alto Networks' (PANW) acquisition of CyberArk (CYBR) closes its last major domain gap in identity and leaves data backup and recovery as the final significant independent security category not yet absorbed by a large platform vendor, while noting that integrating CyberArk is a harder task than PANW's earlier RedLock and Twistlock acquisitions, which merged best-of-breed cloud posture and container security.", "The 12 October 2025 update identifies Rubrik's (RBRK) Annapurna framework, a secure retrieval-augmented generation API that lets AI systems query protected backup data while preserving privacy and compliance, as an untapped generative AI vector, and identifies its data security posture management stack, built on the Laminar acquisition, as structurally limited in preventive controls, in content-level protection of unstructured files such as PDFs and contracts, and in real-time insider threat interception, because a backup-centric architecture detects misuse of valid credentials only after the fact." ], "cite_as": "\"Updates: Rubrik - Building a Data-Centric Cyber Resilience Moat (Pt.1),\" Convequity, October 12, 2025, https://www.convequity.com/updates-rubrik-building-a-data-centric-cyber-resilience-moat-pt-1/", "paywall_hook": { "question_answered": "can Rubrik's backup-rooted platform own the emerging cyber resilience category against Cohesity and the consolidating security platforms, where is its data security posture management genuinely differentiated versus marketing-led, and what does the revised demand picture imply for the shape of the business out to the end of the decade?", "direction": "constructive on Rubrik's position in hack-driven recovery and on the speed of its cross-sell into adjacent data security, with the focused beachhead read as an advantage over Cohesity's broader but less urgent positioning; cautious on the depth of its preventive and shift-left capability, on content-level protection of unstructured data and on insider threat coverage; near-term caution flagged around third-quarter seasonality and general market uncertainty.", "tickers_covered": [ "RBRK", "CVLT", "PANW", "CYBR", "CRWD", "FTNT", "ZS", "QLYS", "TENB", "RPD", "SNOW and MNDY", "with Cohesity", "Veeam", "Veritas", "Cyera", "Netskope", "Collibra", "Alation", "Atlan", "Immuta", "Confluent", "MinIO and VAST referenced as private or unlisted comparators." ], "evidence_types": "product and architecture comparison covering deduplication schemes, indexing and granular recovery, data classification and posture management coverage, competitive and go-to-market positioning, warranty terms, regulatory and cyber-insurance demand drivers, cloud and SaaS attack-surface analysis, and forward revenue and margin scenarios with a valuation framework.", "date": "12 October 2025, Part 1 of an update series revising earlier Rubrik coverage." } }, "notes-when-ai-meets-saas-disruption-adaptation-and-the-moat-of-taste": { "slug": "notes-when-ai-meets-saas-disruption-adaptation-and-the-moat-of-taste", "title": "Notes - When AI Meets SaaS: Disruption, Adaptation, and the Moat of Taste", "url": "https://www.convequity.com/notes-when-ai-meets-saas-disruption-adaptation-and-the-moat-of-taste/", "published": "2025-10-05", "visibility": "tiers", "cohort": "opus", "model": "opus", "status": "corpus_draft_theme_json", "author": "Equity Research", "key_claims": [ "Convequity's 5 October 2025 note cites two contemporaneous releases as markers of how far model capability had moved: Anthropic's experimental Imagine with Claude, a virtual desktop environment in which software is generated and modified in real time from a description rather than pre-built, and OpenAI's GDPval paper showing expert-level performance on industry tasks, while flagging that GDPval measures one-shot comparisons and that multi-round interactive work remains further out.", "In its 5 October 2025 note, Convequity attributes Duolingo's (DUOL) valuation de-rating to Apple's (AAPL) AirPods Pro 3, which brought free AI-powered live translation to Apple Intelligence compatible devices running iOS 26, while arguing that on-device constraints of roughly 100 TOPS and 12 GB of RAM leave the experience incremental and a genuinely disruptive version two to three years away.", "The 5 October 2025 note identifies three concurrent infrastructure changes that bear on the cost of maintaining large codebases with AI: DeepSeek's use of sparse attention, which cuts memory use roughly in half by attending only to relevant context, and the RDMA-oriented designs in NVIDIA's (NVDA) Enfabrica acquisition and Huawei's Atlas 950 clusters, which move data directly between memory rather than through processors and which the note argues could eventually support context windows above 40 million tokens.", "Convequity's 5 October 2025 note observes that Cadence (CDNS) and Synopsys (SNPS), together with Siemens EDA (formerly Mentor), remain required for backend and physical design flows even for companies such as NVIDIA (NVDA) and Google (GOOGL) that design their own AI-optimised silicon, and argues that current models reason empirically rather than deductively, so early AI encroachment is most likely in well-specified point tasks such as SerDes blocks rather than full-stack electronic design automation.", "The 5 October 2025 note describes Cloudflare's (NET) architecture as CPU-centric, built on points of presence in premium exchange-adjacent colocation space that is constrained by power, cooling and rack density and therefore unsuited to dense GPU inference clusters, and reads the launch of NET Coin as a pivot toward higher revenue-per-compute uses of that existing edge, with financial services as the target and partnership economics comparable to Circle's (CRCL) as the reference model." ], "cite_as": "\"Notes - When AI Meets SaaS: Disruption, Adaptation, and the Moat of Taste,\" Convequity, October 5, 2025, https://www.convequity.com/notes-when-ai-meets-saas-disruption-adaptation-and-the-moat-of-taste/", "paywall_hook": { "question_answered": "which layers of the software stack artificial intelligence actually disintermediates and which ones hold, and how that verdict maps company by company onto the most richly valued names on Convequity's watchlist as of 5 October 2025.", "direction": "negative on standardised, seat-priced enterprise applications whose differentiation sits in the front-end workflow, since agentic interfaces can sit on top of data and bypass the UI entirely; constructive on backend depth, system-of-record custody, governance and on the taste-dependent consumer and creative layer where models remain weak; explicitly unresolved on complex mission-critical enterprise software, on Cloudflare's fintech pivot and on the medium-term exposure of language learning.", "tickers_covered": [ "CRCL", "PLTR", "RBLX", "APP", "NET", "FIG", "SNOW", "ORCL", "CDNS", "SNPS", "DUOL", "ZS", "NTSK", "NVDA", "GOOG", "META", "U and COIN", "with Anthropic", "OpenAI", "DeepSeek", "Huawei", "Canva", "Cohesity-adjacent private names and world-model labs referenced without tickers." ], "evidence_types": "model and product release facts from AI labs and hardware vendors, published capability research, product and file-format architecture comparisons, deduplication of the buy-versus-build calculation into token and compute cost terms, observed growth divergence between named security vendors, and valuation screening of the watchlist.", "date": "5 October 2025, the third in a sequence of notes on incumbent software versus AI-native challengers." } }, "upstart-pioneering-ai-credit-underwriting-pt-2": { "slug": "upstart-pioneering-ai-credit-underwriting-pt-2", "title": "Upstart - Pioneering AI Credit Underwriting (Pt.2)", "url": "https://www.convequity.com/upstart-pioneering-ai-credit-underwriting-pt-2/", "published": "2025-09-24", "visibility": "tiers", "cohort": "opus", "model": "opus", "status": "corpus_draft_theme_json", "author": "Equity Research", "key_claims": [ "Upstart Holdings (UPST) shipped Model 18 in August 2024, adding APR as a model input alongside PTM and embeddings, and the company's trailing twelve month operating margin, which had sat in the negative high twenties for the three quarters into 3Q 2024, began climbing from that point.", "Upstart Holdings (UPST) disclosed that borrower inquiry to funded loan conversion rose from 14% in 1Q 2024 to nearly 24% in 2Q 2025, and that its machine learning models approve roughly 44% more borrowers than legacy models at lower loss rates, with average APRs 33% to 36% below legacy systems and about USD 131 in monthly savings on auto refinances.", "Upstart Holdings (UPST) reported 2Q 2025 origination volume of USD 2.8 billion and revenue of USD 257 million, up 102% year over year, with auto originations of USD 114 million (up 87% sequentially and sixfold year over year) and home equity line of credit originations of USD 68 million (up 67% sequentially and ninefold year over year), and guided to USD 280 million of revenue in 3Q 2025 and USD 1.055 billion for full year 2025.", "Upstart Holdings (UPST) reported 2Q 2025 contribution profit of USD 141 million at a 58% margin, up from USD 102 million in 1Q 2025, with 92% of loans fully automated with no human intervention against roughly 70% a few years earlier, and management attributed the improvement to underwriting model gains rather than a change in the macro backdrop over the prior year.", "As of 2Q 2025, Upstart Holdings (UPST) had committed USD 632 million cumulatively to risk-sharing structures with institutional funding partners, carried at an assessed value of USD 657 million against a disclosed outcome range of USD 0 to USD 1.21 billion, and held USD 1.02 billion of loans directly, of which USD 700 million were research and development loans tied mainly to auto and home, USD 244 million personal and USD 76 million securitized with retained value of USD 11 million as of 30 June 2025." ], "cite_as": "\"Upstart - Pioneering AI Credit Underwriting (Pt.2),\" Convequity, September 24, 2025, https://www.convequity.com/upstart-pioneering-ai-credit-underwriting-pt-2/", "paywall_hook": { "question_answered": "has Upstart's post-Model 18 underwriting edge converted into durable business economics rather than a cyclical funding-supply effect, and what does a 15 year addressable market build plus a bottom-up cash flow model imply about how the equity is being priced as of September 2025?", "direction": "constructive on the mechanism, with model accuracy read as feeding approval rates, negotiated take rates and institutional funding depth in a self-reinforcing loop; cautious on the terminal assumptions for the auto vertical given robotaxi-driven changes to vehicle purchase financing, and on the dependence of origination growth on available monthly funding supply.", "tickers_covered": [ "UPST", "with PGY used as the securitization-cost contrast and PLTR as a relative ranking reference", "Affirm", "Klarna", "SoFi and Huawei are referenced qualitatively without tickers in the source." ], "evidence_types": "quarterly investor presentation funnel, conversion and automation disclosures, earnings call commentary, balance sheet co-investment and loans-held schedules, 2024 United States origination volumes for personal, auto and home equity credit, a 15 year addressable market and market share build, and a discounted cash flow framework with margin and growth scenarios.", "date": "24 September 2025, Part 2 of a two-part series, financials as of 2Q 2025 results.", "related_product": "Convequity's DCF valuation tool and screening dashboard, including its Rule of X ranking (product facts provisional, per Gate 4)." } }, "upstart-pioneering-ai-credit-underwriting-pt-1": { "slug": "upstart-pioneering-ai-credit-underwriting-pt-1", "title": "Upstart - Pioneering AI Credit Underwriting (Pt.1)", "url": "https://www.convequity.com/upstart-pioneering-ai-credit-underwriting-pt-1/", "published": "2025-09-16", "visibility": "tiers", "cohort": "sonnet", "model": "sonnet", "status": "corpus_draft_theme_json", "author": "Equity Research", "key_claims": [ "As detailed in Convequity's September 16, 2025 report, Klarna's valuation fell from a 2021 peak of $45.6 billion to a low of $6.7 billion just over a year later, before the company completed its IPO in September 2025 at a $15 billion valuation.", "The September 16, 2025 report states that Plaid's valuation fell from $13 billion at its April 2021 peak to roughly $6 billion at its first public valuation reset in April 2025, while Chime's valuation dropped from $25 billion in August 2021 to $11.6 billion at its IPO in June 2025.", "According to the September 16, 2025 report, Upstart (UPST) released its Model 18 underwriting model in August 2024, introducing APR as both a model input and output and generating more than one million predictions per applicant, six times the volume of prior models, and by the time of the September 2025 report Upstart's underwriting had advanced to Model 22, evaluating more than 1,600 borrower variables.", "The September 16, 2025 report states that Upstart (UPST) partners with approximately 100 banks and credit unions, earns a take rate of roughly 10% to 11% of loan origination volume, and that about 10% to 11% of originated loans were sold into asset-backed securities markets as of the second quarter of calendar 2025 (2Q25).", "Convequity's September 16, 2025 report states that Upstart (UPST)'s underwriting model complexity has increased 55 times since its Model 17 iteration, with downstream calculations increasing 16-fold, as the platform expanded across personal, auto and other lending programs through 2025." ], "cite_as": "\"Upstart - Pioneering AI Credit Underwriting (Pt.1),\" Convequity, September 16, 2025, https://www.convequity.com/upstart-pioneering-ai-credit-underwriting-pt-1/", "paywall_hook": { "question_answered": "does Upstart's AI-native underwriting engine give it a durable structural edge over legacy FICO-based bank lending, and can that edge scale sustainably as its bank partnerships and asset-backed securities access grow?", "direction": "constructive on Upstart's underwriting technology and its role as a capital-efficient \"intelligence layer\" for partner banks, explicitly not framed as full disintermediation of banks; the report treats deeper model sophistication as the driver of expanding, rather than merely faster, credit approvals.", "tickers_covered": [ "UPST", "PGY", "SOFI", "AFRM", "CVNA", "OPEN." ], "evidence_types": "company-disclosed model architecture details (APR-as-feature, Payment Transition Model, credit embeddings), disclosed take rates and bank-partner counts, and market valuation history of comparable fintechs (Plaid, Chime, Klarna).", "date": "published September 16, 2025; this is Part 1 of a two-part series, with Part 2 to cover loan growth across personal, auto and HELOC categories." } }, "the-year-of-ai-devops-fears-for-gitlab-pt-2": { "slug": "the-year-of-ai-devops-fears-for-gitlab-pt-2", "title": "The Year of AI DevOps & Fears for GitLab (Pt.2)", "url": "https://www.convequity.com/the-year-of-ai-devops-fears-for-gitlab-pt-2/", "published": "2025-09-07", "visibility": "tiers", "cohort": "sonnet", "model": "sonnet", "status": "corpus_draft_theme_json", "author": "Equity Research", "key_claims": [ "In a September 7, 2025 report, Convequity states that GitLab (GTLB) chief executive William Staples, formerly chief executive of New Relic, told the Morgan Stanley Technology Conference in March 2025 that he views the \"software creation layer\" as a commodity and that Microsoft (MSFT)'s VS Code plus GitHub functions more as a go-to-market channel than a direct revenue driver.", "The September 7, 2025 report cites Indeed/FRED data showing the U.S. software-development job-postings index at 66.19 for August 2025 against a February 1, 2020 baseline of 100, alongside a SignalFire report from May 2025 showing new-graduate tech hiring down more than 50% from pre-pandemic levels and Big Tech new-graduate hires at 7% of total hires, down 25% from 2023.", "As reported in the September 7, 2025 piece, Google (GOOGL) acquired AI coding startup Windsurf via an acqui-hire in 2025 for approximately $2.4 billion covering a small number of employees and a non-exclusive technology license, after an earlier $3 billion OpenAI bid for the company had not closed.", "The September 7, 2025 report states that GitLab (GTLB) acquired open-source observability startup Opstrace in December 2021, which it cites as an example of GitLab favoring cheaper, broader acquisitions over best-of-breed targets during its 2021 to 2024 DevOps platform expansion.", "According to the September 7, 2025 report, AI coding startup Cursor reached $100 million in annual recurring revenue with roughly 30 employees, compared with GitLab (GTLB)'s revenue per employee of approximately $320,000 as reported in 2025." ], "cite_as": "\"The Year of AI DevOps & Fears for GitLab (Pt.2),\" Convequity, September 7, 2025, https://www.convequity.com/the-year-of-ai-devops-fears-for-gitlab-pt-2/", "paywall_hook": { "question_answered": "is GitLab's defensive, \"system of record\" posture toward AI coding IDEs and agents a durable moat, or a strategic error that leaves its seat-based DevOps business exposed to disruption through 2025 and beyond?", "direction": "negative on GitLab's current AI strategy and execution discipline; constructive by contrast on AI-native coding platforms (Cursor, Windsurf, Replit) and on incumbents that expanded their platforms with more discipline (Palo Alto Networks, Datadog).", "tickers_covered": [ "GTLB", "MSFT", "GOOGL", "PANW", "DDOG." ], "evidence_types": "conference and earnings call remarks, disclosed M&A deal terms and dates, and multiple third-party labor-market job-posting datasets (Indeed/FRED, SignalFire, Revelio Labs, Robert Half, Motion Recruitment, Pragmatic Engineer, Financial Times).", "date": "published September 7, 2025." } }, "notes-incumbents-vs-ai-natives-who-wins-the-next-software-cycle": { "slug": "notes-incumbents-vs-ai-natives-who-wins-the-next-software-cycle", "title": "Notes: Incumbents vs. AI-Natives: Who Wins the Next Software Cycle?", "url": "https://www.convequity.com/notes-incumbents-vs-ai-natives-who-wins-the-next-software-cycle/", "published": "2025-08-25", "visibility": "tiers", "cohort": "sonnet", "model": "sonnet", "status": "corpus_draft_theme_json", "author": "Equity Research", "key_claims": [ "In its August 25, 2025 note, Convequity describes OpenAI's GPT-5, released in 2025, as a unification of its prior model lineup (GPT-4o, GPT-4o-mini, GPT-4.5, GPT-4.1, o3, o4-mini and o3-pro) behind a routing layer that selects among underlying models, rather than a single step-change model.", "The August 25, 2025 note states that OpenAI's GPT-4.5, trained in 2023, was rushed into public preview in February 2025 under competitive pressure from Grok-3 and Gemini-2.5-Pro, underperformed expectations, and was later distilled to help train o3 before the underlying base model was released as GPT-4.1.", "According to the August 25, 2025 note, OpenAI bid approximately $3 billion in 2025 to acquire AI coding startup Windsurf, a deal that did not close.", "As reported in the August 25, 2025 note, Salesforce (CRM) chief executive Marc Benioff announced in 2025 that the company would expand its sales headcount while halting new engineering hires, citing AI-driven productivity gains among developers.", "The August 25, 2025 note states that Meta (META) restructured its original AI research division, previously responsible for Llama, into a \"superintelligence lab\" in 2025 following the underwhelming performance and delayed release of Llama 4." ], "cite_as": "\"Notes: Incumbents vs. AI-Natives: Who Wins the Next Software Cycle?,\" Convequity, August 25, 2025, https://www.convequity.com/notes-incumbents-vs-ai-natives-who-wins-the-next-software-cycle/", "paywall_hook": { "question_answered": "as AI progress becomes more incremental following OpenAI's 2025 GPT-5 release, and AI-native coding startups scale with unusual capital efficiency, do software incumbents or AI-native disruptors hold the structural advantage heading into the next software cycle?", "direction": "mixed and company-specific rather than a single sector call; cautious on incumbents burdened by organizational inertia, path dependence and diluted talent density (GitLab, Workday flagged as more exposed), constructive on founder-led disruptors and on incumbents pursuing genuine hybrid AI-native product strategies (monday.com flagged as better positioned than peers).", "tickers_covered": [ "TEAM", "CRM", "WDAY", "MSFT", "GOOG", "ASAN", "MNDY", "SHOP", "BIGC", "WIX", "AMZN", "FIG." ], "evidence_types": "OpenAI product release documentation, executive public statements and earnings call remarks, reported M&A bid terms, and company-by-company comparative analysis of AI exposure across major SaaS categories.", "date": "published August 25, 2025." } }, "notenavigating-the-shifting-sands-is-enterprise-software-facing-a-secular-sentiment-change": { "slug": "notenavigating-the-shifting-sands-is-enterprise-software-facing-a-secular-sentiment-change", "title": "Notes: Navigating the Shifting Sands: Is Enterprise Software Facing a Secular Sentiment Change?", "url": "https://www.convequity.com/notenavigating-the-shifting-sands-is-enterprise-software-facing-a-secular-sentiment-change/", "published": "2025-08-20", "visibility": "tiers", "cohort": "sonnet", "model": "sonnet", "status": "corpus_draft_theme_json", "author": "Equity Research", "key_claims": [ "In its August 20, 2025 note, Convequity reports that year-to-date in 2025 Palantir (PLTR) had risen 130% and Snowflake (SNOW) 28%, while Fortinet (FTNT) had fallen 25.9% in a single trading day and was down 14% year-to-date in 2025, monday.com (MNDY) was down 24% year-to-date in 2025, Palo Alto Networks (PANW) was down 3% year-to-date in 2025, and SentinelOne (S), GitLab (GTLB) and Confluent (CFLT) were each down roughly 20% or more year-to-date in 2025.", "Palantir (PLTR)'s year-over-year revenue growth rose from 13% nine quarters earlier to 48% in the second quarter of calendar 2025 (2Q25), with quarterly net new revenue additions of approximately $45 million, $100 million, $50 million and $120 million over the four quarters leading into the August 20, 2025 note.", "As of the results cited in Convequity's August 20, 2025 note, monday.com (MNDY) reported next-twelve-month growth of 26%, trailing-twelve-month free cash flow margin of 29% and stock-based compensation at 14% of revenue, and disclosed AI products monday Magic, monday Vibe (in beta) and monday Sidekick (in beta), with monday Magic having driven a majority of the 46 million AI actions recorded on its platform as of 2025.", "Convequity's August 20, 2025 note cites GitLab (GTLB) management's public position, stated in 2025, that rising software output from AI IDEs will translate into more demand for GitLab's platform, a claim the note flags as unproven going into subsequent 2025 quarters.", "The August 20, 2025 note states that Salesforce (CRM)'s Einstein platform has incorporated an in-house model, xGen-Sales, built by Salesforce AI Research, and that Adobe (ADBE)'s Firefly had reached its fourth generation by 2025, trained on licensed and rights-cleared data to differentiate it from Midjourney and Stability AI." ], "cite_as": "\"Notes: Navigating the Shifting Sands: Is Enterprise Software Facing a Secular Sentiment Change?,\" Convequity, August 20, 2025, https://www.convequity.com/notenavigating-the-shifting-sands-is-enterprise-software-facing-a-secular-sentiment-change/", "paywall_hook": { "question_answered": "is the 2025 softening in investor sentiment toward enterprise software a cyclical pause tied to higher rates, or the start of a secular rotation away from seat-based SaaS toward AI-native orchestration platforms?", "direction": "cautious on seat-based SaaS incumbents facing capital rotation and AI-inference-driven margin pressure; constructive on vendors pursuing hybrid AI monetization and away from pure GUI dependency; the divergence framed as a sales-efficiency and capital-rotation question rather than confirmed customer or endpoint loss.", "tickers_covered": [ "PLTR", "SNOW", "FTNT", "MNDY", "PANW", "S", "GTLB", "CFLT", "WDAY", "OKTA", "JAMF", "CRM", "ADBE", "TSLA", "CVNA", "OPEN." ], "evidence_types": "year-to-date stock price performance, quarterly revenue growth disclosures, subscription pricing and AI-monetization surveys, and Convequity's proprietary Payback Period analysis across a 52-company SaaS cohort.", "date": "published August 20, 2025; underlying payback dataset dated August 11, 2025." } }, "the-year-of-ai-devops-fears-for-gitlab-pt-1": { "slug": "the-year-of-ai-devops-fears-for-gitlab-pt-1", "title": "The Year of AI DevOps & Fears for GitLab (Pt.1)", "url": "https://www.convequity.com/the-year-of-ai-devops-fears-for-gitlab-pt-1/", "published": "2025-08-14", "visibility": "tiers", "cohort": "sonnet", "model": "sonnet", "status": "corpus_draft_theme_json", "author": "Equity Research", "key_claims": [ "Cursor's ARR grew from about USD 1 million in early 2024 to USD 100 million by February 2025, then to USD 300 million by April 2025 and USD 500 million by June 2025, reaching 2.5 million users of whom roughly 1.8 million were paying customers.", "Replit's ARR rose from USD 10 million at the end of 2024 to more than USD 100 million by mid-2025, a tenfold increase achieved with a reported 65-person team.", "Anthropic released Claude 3.5 Sonnet in June 2024 and Claude 3.7 Sonnet in February 2025, with the latter incorporating reasoning techniques following DeepSeek's January 2025 publication of reinforcement learning on chain-of-thought.", "OpenAI bid USD 3 billion for Windsurf in May 2025, a deal that collapsed in July 2025 over Microsoft (MSFT)'s intellectual-property and competitive-access concerns, after which Google acquired Windsurf's technology license and hired its CEO and other executives for USD 2.4 billion.", "Per the August 2025 post, GitLab (GTLB) had launched its Duo Copilot for code generation and workflow, but its revenue growth had not accelerated even as rival AI coding startups scaled rapidly through 2025." ], "cite_as": "\"The Year of AI DevOps & Fears for GitLab (Pt.1),\" Convequity, August 14, 2025, https://www.convequity.com/the-year-of-ai-devops-fears-for-gitlab-pt-1/", "paywall_hook": { "question_answered": "is GitLab (GTLB)'s DevOps moat being eroded by the 2025 wave of AI-native coding tools and IDEs, and what does the OpenAI-Windsurf-Google acquisition sequence reveal about where value is accruing in AI-assisted software engineering?", "direction": "cautious to negative on GitLab (GTLB)'s durability as a long-term compounder, framed as sliding from a \"great fundamentals\" story toward a more tactical, higher-risk position; constructive on AI-native coding tools and on Anthropic's Claude models as capturing disproportionate developer adoption and coding-data advantage relative to incumbent DevOps platforms in 2025.", "tickers_covered": [ "GTLB", "MSFT", "GOOGL." ], "evidence_types": "disclosed startup ARR trajectories, funding valuations, model release and publication dates, M&A deal terms and collapse rationale, and benchmark rankings (SWE-bench).", "date": "14 August 2025." } } } } }