Etched Exits Stealth With Working AI Inference Chip

The scarce, expensive resource in production AI is no longer training compute but inference capacity, and almost all of it runs on Nvidia GPUs. A merchant-silicon vendor shipping a working, rack-scale inference system built specifically for serving large models is exactly the kind of supply-side shock that could bend the cost curve teams budget around. Etched is betting that purpose-built inference hardware, co-designed across chip, memory, and rack, can beat general-purpose GPUs on the metrics that decide serving economics: tokens per second, latency, and watts per token.

Etched said it achieved first-pass (A0) silicon success on TSMC's N4P process in under three years from seed funding and is now validating its first rack-scale product with customers. The company reported $800 million raised across several unannounced rounds, most recently $500 million at a $5 billion post-money valuation in December, and said it holds more than $1 billion in signed customer contracts. Its systems are described as running models including DeepSeek, Qwen, Mamba, and Llama. CEO Gavin Uberti said the infrastructure needed to serve frontier models sustainably 'simply did not exist,' framing the company's thesis around accelerated inference. Co-founder Rob Wachen said Etched built for gigawatt-scale from the start and characterized 'production' as 'the product,' pointing to a Taiwan factory plus a San Jose data center, test house, and prototyping lab, with a path to gigawatt-scale in 2027.

Two claims deserve scrutiny before teams factor Etched into capacity plans. First, breadth: early transformer-only accelerators traded flexibility for speed, so support for Mamba and 'arbitrarily large' parameter counts, if it holds across future architectures, would be the differentiator that matters. Second, the throughput, latency, and power numbers are the company's own and are not yet independently benchmarked; the meaningful comparison is against Nvidia's GB300-class racks on identical models at identical accuracy. The investor roster and a stated deep foundry partnership signal supply-chain seriousness, but the real test is whether 'first racks ship this summer' converts $1 billion in contracts into deployed, repeatable systems.

Etched joins a crowded field of inference-focused silicon, alongside custom ASIC efforts from hyperscalers and merchant challengers such as Cerebras and Groq. What separates this announcement is the combination of working A0 silicon, named large-model support, and contracts in hand rather than design wins on paper. If the shipped systems match the claims, buyers gain leverage in a market where Nvidia allocation has been the binding constraint.