Etched, a startup building application-specific chips (ASICs) hardwired for transformer inference, announced it has secured $1 billion in contracted sales and achieved a $5 billion private valuation. The company is pitching its Sohu chip as a purpose-built inference accelerator that can outperform Nvidia's H100/H200 on transformer workloads at lower cost — a credible niche given that inference is increasingly the dominant AI compute workload post-training.

Etched is not publicly traded, so the headline is really a NVDA read. Nvidia posted $215.9B in revenue for FY2026 (ending Jan 2026), up 65.5% YoY, with a 71.1% gross margin and 55.6% net margin — numbers that reflect near-monopoly pricing power in AI compute. The $1B Etched contract figure is meaningful as proof of concept but represents a rounding error against Nvidia's current revenue base.

The second-order question is whether ASIC challengers (Etched, Groq, Cerebras, plus hyperscaler-custom silicon from Google TPUs, AWS Trainium, and Microsoft Maia) collectively represent a structural ceiling on Nvidia's inference TAM. The bear case on NVDA centers on margin compression and share erosion as inference workloads — which don't require Nvidia's full training flexibility — shift to cheaper alternatives. The bull case is that Nvidia's CUDA ecosystem lock-in, software stack, and networking (NVLink, InfiniBand) create switching costs that purpose-built ASICs can't easily replicate.

What to watch: Etched's ability to land hyperscaler contracts (AWS, Azure, Google) at scale would be the real signal. Until a major cloud provider publicly shifts inference capacity away from Nvidia at volume, this remains a long-tail competitive threat rather than an imminent share-shift catalyst. NVDA's next earnings print and data center segment guidance will be the cleanest tell on whether pricing or unit demand is softening.