AI chip startup Etched emerged from stealth on Monday with a working chip manufactured by TSMC, $800 million in funding raised, a $5 billion valuation, and over $1 billion in signed customer contracts for its Sohu inference processor.
Founded by two Harvard dropouts and based in San Jose, Etched is betting that the future of AI hardware belongs to application-specific integrated circuits, not general-purpose GPUs. Its Sohu chip is an ASIC built exclusively for transformer model inference — doing one thing, extremely well — rather than chasing Nvidia's everything-to-everyone approach.
The $1 billion in forward contracts covers full-stack inference systems that handle both prefill and decode workloads. The company claims its rack-scale systems deliver orders-of-magnitude better performance-per-dollar on transformer inference compared to general-purpose GPU clusters.
The timing is pointed. Nvidia's data center revenue continues to face headwinds in China, where Huawei's Ascend chips have taken the lead, and the broader market is hungry for alternatives to Nvidia's pricing power. Etched is not the only challenger — Groq, Cerebras, and d-Matrix are all circling — but the $1 billion in signed deals before a public product launch is unprecedented.
Analysis
A $5 billion valuation for a company with a single ASIC and no public product is either brilliant timing or a top-tick signal. The transformer-only bet is high-risk: if a new architecture displaces the transformer within the next three years, Sohu is a paperweight. But with $800 million in the bank and TSMC-manufactured silicon in hand, Etched has more running room than most hardware startups get. The real question is execution — tape-outs to volume shipments is where chip startups die.
