Notes: The Bottleneck Heuristic — Blackwell's Gauntlet, the ASIC Breakout, and the Optical Chokepoint (Pt.2)
Summary
* Following Part 1's investment heuristics, Part 2 recaps the pivotal developments of 2024–25, starting with NVDA's Blackwell — delayed ~9–12 months by packaging respins, liquid-cooling redesigns, NVL72 reliability issues, and an on-chip SFU bottleneck.
* Google's TPUv7 (AVGO-designed) shattered the "ASICs can't train" consensus by powering Gemini 3 Pro past GPT-5 and cutting Anthropic's token costs ~3x — the custom ASIC breakout moment investors can no longer dismiss.
* MSFT's Maia
