Notes: Huawei’s 910C And CM384: A Strategic Shock To The NVIDIA Thesis (Pt.2)

Summary

• Huawei's AI-native 910C and CM384 systems now rival NVIDIA’s Blackwell in real-world performance, despite using older 7nm and HBM2e components.
• By vertically integrating chip, system, cloud, and power infrastructure, Huawei is scaling compute in ways general-purpose GPGPU systems like NVIDIA’s Rubin cannot.
• While Huawei remains uninvestable, its rise threatens NVIDIA’s pricing power, ecosystem dominance, and long-term competitiveness in China and beyond.

Ascend’s Adoption Gap and System-Level Trade-Offs

Although Huawei’s Ascend platform is proving far more capable than many initially expected, it still faces key limitations. Most notably, bottom-up adoption of Ascend remains weak compared to NVIDIA. This is partly due to Ascend’s advanced NPU architecture, which requires deeper optimization work to unlock its full potential — something that deters many smaller engineering teams and independent developers.

Despite the H20 offering roughly half the BF16 compute of the 910B, it shipped with larger and faster HBM3 memory, and sold at a similar price point. As a result, the H20 significantly outsold the 910B throughout 2024. Following the release of DeepSeek V3 and R1, demand for H20 surged in China. Leading AI customers — including DeepSeek, ByteDance, Tencent, and Alibaba — each placed H20 orders exceeding $1bn, even as U.S. sanctions intensified and Chinese authorities advised companies to reduce reliance on NVIDIA. This demand spike was largely due to a shortage of 910B supply, which constrained Chinese hyperscalers' ability to meet growing inference needs. ByteDance and Tencent, for instance, have integrated LLMs natively into core applications like Douyin (TikTok China) and WeChat — platforms with over 700 million MAUs—driving enormous, urgent demand for inference compute.

Another structural issue is the state of Huawei’s AI software stack. While Huawei has made significant progress with its Ascend ecosystem and improved CANN's (Compute Architecture for Neural Networks) usability, the platform remains optimized primarily for large enterprises and high-volume inference workloads like DeepSeek. For developers working outside these common use cases — or for those building from the ground up using open tooling — the experience is more fragmented. In these scenarios, engineers may encounter issues quickly, and unless they’re a large enterprise with direct Huawei support, resolving them often means waiting on technical service or working around limitations independently.

Finally, although using all optical modules allows Huawei to deliver larger scale up clusters with higher MFU and linearity whilst lowering down power density per rack, Huawei still needs to figure out ways to improve the cost-effectiveness further in the future via either cutting down optical connection costs further or increasing rack density and use copper cables for inner-rack communication to reduce the cost further.

910C: Made in China?

Contrary to widespread belief, we believe the Ascend 910C is — if not already — now being mass-produced entirely within mainland China. Specifically, it is being fabricated by SMIC (Semiconductor Manufacturing International Corporation), China’s leading foundry, based in the People’s Republic of China (PRC). This contrasts with TSMC, which, of course, is located in Taiwan, officially known as the Republic of China (ROC).

The misconception likely stems from early reporting by TechInsights, a Canada-based semiconductor reverse engineering firm known for its industry-standard chip teardowns. When TechInsights examined the re-released 910B in 2023, their analysis confirmed SMIC fabrication, but did not cover the newer 910C. As of now, neither TechInsights nor other Western labs have published teardown results for the 910C. However, by 2024, 910C chips were already circulating widely in China — and all samples we've seen point to SMIC as the foundry, including recent batches of the 910B.

The prevailing confusion seems to be rooted in outdated reports and the opaque nature of China’s domestic supply chain. Below, we outline our reasoning in support of the view that the 910C is domestically manufactured:

1. TechInsights’ teardown analysis
2. The progress of SMIC South, Sicarrier, CXMT, and Swaysure

TechInsights’ Analysis May Be Misguided

In October 2024, TechInsights — one of the most respected semiconductor reverse engineering labs — reportedly shared findings with both TSMC and the U.S. Bureau of Industry and Security (BIS), claiming that it had detected TSMC-manufactured components in Huawei’s Ascend 910B chip. This raised concerns due to the U.S. export restrictions that forced TSMC to cease all shipments to Huawei after September 15, 2020. Both TSMC and Huawei denied any continued business ties beyond that date, and the BIS declined to comment.

This situation mirrors a separate case involving TSMC and Sophgo, a Chinese shell company tied to Bitmain. Sophgo had used TSMC’s foundry services for an AI ASIC that violated U.S. sanctions. Once this was uncovered, TSMC was fined $1 billion. Despite the similarities, no penalties have been issued against TSMC or Huawei in the context of the 910B.

The most likely explanation is that TechInsights analyzed a 910B unit from an early batch, one assembled using pre-ban TSMC logic dies that had been manufactured but not yet packaged at the time of the U.S. cutoff. These were later assembled using a CoWoS-like 2.5D packaging process by SJ Semiconductor, a SMIC subsidiary. Back in September 2020, as the ban deadline approached, TSMC prioritized Huawei’s orders for mobile SoCs, NPUs, and base station chips — at one point even bumping Apple’s orders to the back of the queue. On the final day, September 15, Huawei dispatched a cargo aircraft to Taiwan’s Taoyuan International Airport to retrieve all remaining inventory: packaged and unpackaged chips, untested wafers, and everything in between. This included over 10 million 5nm Kirin 9000 SoCs, millions of 7nm server CPUs, fewer than 1 million Ascend NPUs, and more than 1 million 7nm base station controllers.

Because demand for smartphones and base stations surged in the following years, those inventories depleted quickly. However, demand for NPUs was initially lower, allowing Huawei to draw from this stockpile of older 910B dies well into 2023. That’s likely the source of the sample TechInsights analyzed in 2024 — not a sign of recent TSMC involvement, but a reflection of Huawei’s strategic inventory buffer.

Meanwhile, SJ Semiconductor had developed its own advanced 2.5D packaging capabilities, designed to be compatible with TSMC-style CoWoS. This allowed Huawei to continue shipping 910Bs without waiting for SMIC’s 7nm process to fully ramp. Using the leftover TSMC dies, Huawei likely shipped around 200,000–250,000 910Bs before transitioning to chips fully fabricated by SMIC.

TechInsights’ teardown found TSV patterns and interconnect layouts that closely resembled TSMC’s packaging, sparking speculation that Huawei had continued sourcing from TSMC through shell companies. However, we see that theory as unlikely. Unlike EUV-dependent 5nm nodes, 7nm and 2.5D packaging are well within SMIC’s capabilities. Given the three-year gap between the 2020 ban and 2023, it’s entirely plausible that SMIC caught up by replicating TSMC’s legacy packaging standards.

We estimate cumulative shipments of the 910B to be around 350,000 units, with the majority based on pre-ban TSMC dies. SMIC-produced batches remain a minority due to earlier yield challenges. Looking ahead, Huawei plans to ship roughly 320,000 units of the newer 910C in 2025. This target is feasible from a logic die perspective, though scaling further may constrain production of Kirin mobile SoCs. According to one source, Huawei has already postponed its next-generation Mate-series smartphone to preserve capacity for 910C production. Another source notes that the primary bottleneck may now be HBM supply, not logic. While CXMT and Huawei have made progress on HBM2e domestication, their pace trails that of YMTC, which is already leading globally in 3D NAND layer advancement.

China's semiconductor progress

Entering 2025, both logic and memory bottlenecks appear largely resolved for Huawei. On the logic side, Huawei’s reborn Kirin chips — manufactured by SMIC on the 7nm process — have been in mass production since late 2023, with cumulative shipments surpassing 50 million units. These are highly complex system-on-chip (SoC) dies, sensitive to manufacturing defects due to their intricate, non-repetitive internal layouts. Notably, the latest Kirin 9020, released in late 2024, features a die size 15% larger (now 136.6mm²), signaling growing confidence in SMIC South’s yield performance. Larger die sizes typically reflect better yields because manufacturers would not risk increasing area unless the process was stable enough to deliver a high percentage of usable chips.

Although the 910B is more than five times larger than the Kirin SoC, much of its die consists of highly repetitive structures — specifically, arrays of identical tensor or matrix computation units. This structural uniformity makes it more tolerant to manufacturing defects. It also explains why NVIDIA’s A100, H100, and B200 all push close to the 800mm² reticle limit — the maximum die size allowed by chip lithography tools. Even at such large sizes, these chips maintain acceptable yields because any defective blocks can be selectively disabled, allowing the chip to function correctly with only minimal performance loss. This defect-masking strategy is commonly used in large AI accelerators and GPUs to balance performance and manufacturability.

China is estimated to possess at least 100+ ASML 1980di and 25+ 2000i DUV lithography machines, suitable for custom patterning — a critical factor for producing 7nm, and even 5nm chips, without EUV. For now, these advanced DUV machines suffice for 910C production, but scaling to 1 million 910C units per year remains a significant challenge. Either domestic advanced DUV development must accelerate, or SMIC must achieve further yield improvements.

Huawei’s confidence in the domestic supply chain is evident: it publicly announced the 910C at a major event without apparent concern for further US sanctions. The release date, April 10th, was deliberately chosen to coincide with new US tariffs on China, and about three weeks after SiCarrier — Huawei’s WFE spin-out — unveiled 14 product lines covering all aspects of advanced node production except lithography machines. SiCarrier’s lineup includes advanced wafer fab equipment for CoWoS packaging and HBM production, and they claim fully domestic sub-component production, leaving little room for further sanctions. Though SiCarrier has operated for years, Huawei’s public disclosure now signals strong confidence in the technology.

Similarly, Huawei has never publicly marketed its SMIC-produced chips. For the reborn Kirin series since late 2023, Huawei has avoided mentioning chip names or technical details at launch events; any teardown or deep analysis videos have been swiftly removed by authorities. This discretion also extended to Huawei’s ADAS chips, shipped in over 500,000 “Huawei Inside” EVs last year. Thus, the public release and promotion of the 910C is a clear sign of renewed confidence.

The 910C tape-out was completed in Q1 2024 at SMIC, with logic die shipments from SMIC starting in February and March. By April 2025, packaging and testing were finalized. The 910C underwent its first HVM batch and core customer verification — including ByteDance — before release. During this period, tech enthusiasts dismantled recent 910B and 910C samples and found zero TSMC footprints.

On the memory front, as TechInsights aptly summarized in a recent article titled “China Enters 2025 with Big Memory Breakthroughs,” Chinese NAND and DRAM manufacturers have largely overcome US sanctions and are advancing rapidly. YMTC, the NAND producer, is now state-of-the-art and has compelled Samsung to license YMTC patents; Samsung admitted it is impossible to bypass YMTC’s patents for 300-layer NAND and beyond. CXMT, the DRAM producer, had lagged behind — producing dies two generations behind DDR4 for the past three years. However, by late 2024, CXMT succeeded in mass producing DDR5 and HBM2e as well.

CXMT appears to have a looser relationship with Huawei. Instead, Huawei’s closest memory supplier is likely Swaysure, based in Shenzhen, which has received significant funding and has attracted top talent, including Japanese memory chip veteran Yukio Sakamoto as Chief Strategy Officer. Sakamoto, the former CEO of Elpida (Japan’s last DRAM giant), was known for high-performance products that struggled to compete on cost with Korean and Taiwanese rivals. Elpida was eventually acquired by Micron, but many ex-Elpida engineers, like Sakamoto, see renewed opportunity in China’s combination of capital and engineering strength to make advanced technology affordable and scalable.

Like advanced packaging, HBM production does not require EUV, but rather advanced wafer fab equipment for high aspect ratio processing. With SiCarrier’s recent product releases, it’s no surprise that China has already achieved HBM2e production, and may follow Micron’s path to start producing HBM3e (bypassing HBM3), or even HBM4 in the future.

Future roadmap

Huawei’s Ascend AI Cloud service generated 50 billion RMB in revenue in 2024, representing 700% year-over-year growth. In anticipation of continued demand, Huawei began development of the 910C and CM384 systems in 2023 with a focus on supporting advanced AI workloads, including:

• Ultra-long sequence processing
• 10 trillion+ parameter models
• Multi-modality
• Enhanced sparse Mixture of Experts (MoE) architectures

The naming of “CM384” is particularly notable. It directly reflects DeepSeek V3’s inference cluster design: 256 active experts, 64 redundant experts, and 64 dedicated for prefill. This represents a deliberate architectural departure from the conventional 8-card node paradigm, further evidence of Huawei’s close coordination with leading AI workloads and frameworks.

Looking ahead, Huawei plans to scale across all four core system dimensions — compute, networking, memory, and power. The next-generation 920-series will complete technical design by Q2 2025, enter production later that year, and begin shipping in 2026. Huawei expects a 30–40% performance gain, enabled by a transition to a 6nm node offering 30% greater density and 15% lower power consumption. The 920 will feature HBM3 with 4 TB/s bandwidth, and, as with the 910-series, will come in two versions: a single-die inference chip and a dual-die training chip. Targeted single-die performance exceeds 900 TOPS (BF16) — bringing it close to NVIDIA’s Blackwell.

Huawei’s roadmap also envisions scale-up architectures that exceed 8,000 NPUs per node, enabled by its vertically integrated optical interconnects and distributed system design. In contrast, NVIDIA’s upcoming NVL576 Rubin system supports 144 “GPUs” per node — but this figure reflects a marketing convention rather than true chip count, as each Rubin chip contains four GPU dies that NVIDIA now counts individually. This shift illustrates the physical scaling limits of general-purpose GPU systems and the challenges NVIDIA faces in pushing beyond single-rack configurations. Huawei, by contrast, is embedding distributed micro-switches directly within each NPU, allowing it to scale horizontally without incurring the latency and cost overheads of traditional all-to-all interconnects. These optimizations are only possible in purpose-built AI architectures, underscoring how Huawei’s AI-native approach is diverging sharply from legacy GPGPU models constrained by package complexity and software abstraction layers.

Looking further ahead, the pace of Huawei’s silicon roadmap will depend on when China can field a fully domestic immersion-DUV tool that matches the “2000 i” class scanners SMIC now runs for 7nm. Today there are two parallel tracks:

• Imported ASML immersion-DUV – already installed at SMIC and stretched (via multi-patterning) to make Huawei’s current 7 nm Kirin and Ascend dies.
• Huawei’s own DUV programme – a home-grown 193 nm scanner that has reached volume capability at the 28 nm node and is now in pilot trials for immersion mode; its first production target would be something in the mid-teens rather than full 7 nm.

Beyond DUV, Huawei has assembled a first-generation DPP EUV light source, but a production-grade system will require an LPP architecture comparable to ASML’s. The most optimistic window for that step is 2028-2030. Encouragingly, on 28 Apr 2025 a team led by a former ASML engineer demonstrated a solid-state LPP source with 3 % conversion efficiency (roughly half of ASML’s 5.5 %) at one-tenth the power, and project a path to 6% efficiency and 10 kW output. Until domestic EUV is industrialised, Huawei will keep extracting performance by enlarging die area and relying on SMIC’s imported immersion-DUV for 7 nm, while its own DUV toolset underpins less-advanced nodes such as 28 nm.

Huawei Cloud & Security Stack & Electricity Dominance

Beyond AI, it’s important to recognize that Huawei Cloud itself is expanding rapidly. Huawei Cloud now operates nearly 100 Availability Zones across 33 regions, with 2,800 CDN PoPs and total network capacity reaching 150 Tbps. In many respects, it is poised to become one of the largest cloud providers outside the US. Huawei emphasizes its reliability advantage, rooted in its telco heritage and robust service arm. From January 1, 2024, to April 10, 2025, Huawei claims there were 0 major Availability Zone-level outages, compared to 23 for Google Cloud Platform, 19 for Azure, and 8 for AWS.

Leveraging its telco background, Huawei can automate and accelerate troubleshooting — aiming to spot errors within seconds, identify issues within 1 minute, and resolve them within 5 minutes. Even in colocation data centers not directly managed by Huawei, it delivers superior reliability. In one notable instance, while AWS and Azure experienced outages in the same facility due to air conditioning failure, Huawei Cloud remained unaffected: they dispatched several trucks loaded with dry ice to cool the server rooms for two consecutive days until repairs were completed, ensuring uninterrupted service.

For Ascend AI services, Huawei does not sell individual chips. Instead, it delivers AI capabilities via public cloud or private, on-premises cloud deployments. Private cloud options enable customers to retain ownership while still enjoying a cloud-like “as-a-service” experience, with Huawei service engineers handling everything from hardware maintenance to software optimization.

On the infrastructure side, Huawei differentiates itself from NVIDIA by delivering the Ascend Cloud service primarily through its own cloud platform, ensuring customers can fully leverage CM384’s capabilities. While NVIDIA has simplified customer deployment with pre-built CUDA operators, NVL72 rack systems, and DGX SuperPod reference designs, organizations with cutting-edge needs often require additional expertise. Each AI lab typically develops unique methods to maximize MFU, quickly recover from frequent system failures, and customize frameworks, operators, and schedulers as needed.

Given the inherent complexity of managing CM384 clusters, Huawei has taken a proactive approach — offering direct cloud services or on-premises Ascend servers, with most maintenance managed by Huawei experts. The company claims superior MFU, MTBF (Mean Time Between Failures), and overall cluster reliability, achieved through:

• Full-stack troubleshooting: issue detection in 1 minute, scope identification in 3 minutes, and recovery in 10 minutes
• Increasing automated issue detection from 40% to 90%
• Over 1,000 issue-fixing templates
• More than 100 automated remediation solutions

Through collaboration with one AI lab, Huawei successfully reduced the failure rate from 2.5 machine failures per day to just 0.15 per day.

To further streamline the experience for customers, Huawei is also automating inference optimization software. This approach is very similar to DeepEP, open-sourced by DeepSeek and later adopted by NVIDIA’s Dynamo. However, Huawei takes it a step further by offering these optimizations as a managed service, rather than as a framework for customers to deploy and tweak themselves. Key features include:

• Elastic Inference Service: Delivers a 50% performance boost for MoE (Mixture of Experts) + CoT (Chain of Thought) workloads
• Containerized, serverless resource allocation: Improves resource utilization by 50%
• Dynamic allocation of prefill and decoding tasks within the cluster: Delivers 50% higher throughput
• Dynamic expert allocation based on usage popularity: Achieves 20% greater resource utilization

This service model means Huawei must dispatch teams of engineers to customers’ private cloud data centers, working closely with them on-site. In many respects, these engineers function similarly to Palantir’s FDE (Forward Deployed Engineers), solving customer issues on-premises and then synthesizing their expertise to build operational software that automates and reduces repetitive tasks.

As with Palantir, however, Huawei’s approach has a limitation: it primarily serves large customers with spending above 40 million RMB (about $5.5 million). Smaller clients are limited to cloud-delivered solutions and cannot deploy Ascend infrastructure on-premises.

It’s also noteworthy that Huawei’s core focus extends beyond just chips or catching up with NVIDIA. The company has concentrated on providing vertical solutions for its customers, forming numerous partnerships in industrial and manufacturing sectors to deliver tangible AI ROI. This strategy closely resembles Palantir’s approach, focusing on real-world applications and vertical integration, rather than simply emulating NVIDIA or other competitors.

Security

After reviewing Huawei’s entire product line we truly find it highly impressive. We’ve often wondered how Huawei manages to stay secure despite persistent attempts by entities like the US government and the NSA, which possess vast resources, advanced hacking tools, and access to zero-day vulnerabilities. The answer appears to be that Huawei has developed a robust security product line — and has been “dogfooding” for quite some time. In many ways, Huawei is akin to a Chinese version of Fortinet (FTNT) plus Palo Alto Networks (PANW): it might adopt the latest security paradigms a bit slower, but it never falls behind so far that it can't catch up — and it has the engineering muscle to integrate everything under one roof.

Huawei has trained its own security LLM, developed over 400 specialized ML models, and created more than 100 automated playbooks, unifying these capabilities into a single security agent capable of:

• Sensitive data autodetection
• Threat detection
• Payload identification
• Auto-generated response
• Auto-generated root tracing
• Delivering results where 1 security operator + agent > 10 security experts
• Reducing average time to threat identification by 83%
• Automatically handling 99% of threats
• Cutting threat root tracing time by 95%

This technology reduces the number of required SOC operators by a factor of ten. Huawei Cloud customers can easily create and deploy custom-defined security agents in just a few clicks. This could be the first security agent of its kind in production — potentially even a step ahead of what PANW’s XSIAM platform is currently doing.

On the cloud security front, Huawei Cloud offers shift-left solutions such as SAST, SCA, SBOM, and SDLC, as well as CSPM and CIEM for configuration and access. Notably, it also provides Open Source Security, with a secured repository and curated packages for customers — mirroring the cutting edge of open source security being pioneered by startups like Chainguard.

Huawei’s shift-left solution helps application developers on Huawei Cloud comply with data sovereignty rules, sanctions, and regulations, while also ensuring software security by enforcing the use of only secured open-source packages, thereby minimizing attack surfaces wherever possible.

It is remarkable to see that Huawei maintains such a rapid pace of innovation across all fronts, staying current with the latest industry trends and cybersecurity startup advancements — all while being a $120 billion revenue giant, twice the size of Cisco.

Electricity Dominance

Lastly, we took a closer look into Huawei’s electricity supply status. Once again, it appears that Huawei is extremely well-positioned. Huawei Cloud operates three central hubs in China: Inner Mongolia (North), Wuhu City, Anhui (East), and Guizhou (Southwest). Inner Mongolia is China’s wind farm center, generating massive amounts of low-cost renewable energy during winter. Guizhou, by contrast, is located in the southwest and produces significant hydroelectric power during summer. The Three Gorges Dam alone, in this region, generates 7% of China’s total electricity.

In recent years, China has doubled down on investment in solar, wind, and hydro power in both Inner Mongolia and the southwest. In Inner Mongolia, total solar capacity reached 48 GW in 2024, up 51% year-over-year, while wind capacity hit 86 GW, a 23% annual increase. Meanwhile, Southwest China’s hydro capacity now exceeds 210 GW, growing at low single digits.

Unlike Inner Mongolia, which is close to northern cities, the southwest is geographically isolated from major industrial and population centers. As a result, during summer, a huge amount of electricity has historically been wasted. Previously, blockchain miners would migrate their data centers from Inner Mongolia in winter to the southwest in summer to utilize this excess power. In recent years, however, much of this hydro capacity has remained idle. The ongoing buildout of cloud data centers in the region is now helping to address this underutilization and reduce operating costs.

Wuhu city serves as the central hub for Huawei Cloud. Located in China’s most populated and energy-consuming region, Wuhu faces higher electricity costs and more limited access to renewable energy — except for nuclear. The planned nuclear capacity for East China is projected to reach about 70 GW by 2035.

Huawei is also the market leader in several critical infrastructure segments, including data center air conditioning, power supply, and solar energy transformers. Huawei’s air conditioner business began over a decade ago, driven by the need for reliable, digitized AC systems for telco base stations — demanding standards unmet by existing market offerings. Similarly, Huawei leads China’s solar energy supply chain, holding a 22.9% market share in photovoltaic inverters within China’s 223 GW market. The company’s power supply and other data center infrastructure units originated from the same need: existing market solutions could not meet Huawei’s demanding requirements.

From what we’ve found, the data center infrastructure market remains dominated by decades- or even century-old incumbents, with little disruption. Huawei and Tesla (TSLA) stand out as two digital-native tech innovators with significant potential to capture more share in this space going forward.