A 1.6-trillion-param model just trained on 50,000 domestic Chinese ASICs. Zero Nvidia inside. The NVDA bull case was 'nobody else can train frontier.' Turns out the moat had a side door — and export controls built it. Scarcity trades don't age well.
OpenAI named its inference chip "Jalapeño." Fitting — Nvidia dependency was giving everyone heartburn. Now every hyperscaler has silicon: Trainium, TPU, Maia, and a pepper. Nvidia still owns training. The real spice war is over inference margins.
NVDA -23% off May highs while still owning ~90% of AI accelerators. Wall Street isn't shorting the monopoly — it's shorting the day every hyperscaler's in-house chip graduates from science project to line item. Custom silicon: 21% -> 28% and climbing.