Etched raises landmark round from Jane Street, Jump, Two Sigma and Peter Thiel to build transformer-specific AI inference chips

Etched emerges from stealth with rack-scale inference hardware and a standout cap table

Etched is building full-stack AI inference hardware — chips, boards, platforms, racks, clusters, and the production lines to manufacture them at scale. The company's founding thesis, that the transformer architecture is durable enough to warrant purpose-built silicon, has held up well enough that Gavin Uberti is now ready to recruit publicly.

The technology

Etched is announcing two proprietary technologies alongside its emergence from stealth.

Low voltage inference attacks the thermal ceiling that prevents GPUs from running at full theoretical capacity. Uberti says current GPUs can only sustain around 50% of their theoretical compute before overheating. By co-designing transistors with TSMC to operate at dramatically lower voltages, Etched can draw less power per unit of compute and fit significantly more FLOPS onto a chip. Uberti describes TSMC as essential to making this work — lower voltage manufacturing is materially harder to execute, and Etched could not have done it without a fab partner willing to co-design at the transistor level.

Cluster scale memory targets the decode side of inference, where bandwidth rather than raw compute is the bottleneck. The technology enables coherent memory pooling across chips within a cluster, effectively treating all HBM, HBM bandwidth, SRAM, and SRAM bandwidth as a single pool. The practical result, Uberti says, is faster tokens per second for end users without a corresponding cost increase.

“We are building rack scale inference hardware. That means chips, boards, platforms, racks, clusters, as well as software, and importantly, the production lines to build those things at scale. People don't realize how big of a market inference is going to be. Right now, only a couple million people have access to the frontier models, and those are going to have to get deployed to 8,000,000,000 people across the globe.”

The moat question

Uberti's answer to whether chip companies are becoming commodity products is that speed still commands a premium, in the same way a Ferrari commands one over a Toyota despite both being cars. Trade secrets and patents matter, he says, but the more durable advantage is the willingness to build the whole stack — chip through production line — rather than just the silicon. He also argues that as models grow to many trillions of parameters and token demand scales to quadrillions per month, the economics of investing in large-scale manufacturing become increasingly compelling.

Why come out of stealth now

Talent, not customers. Uberti says Etched has no shortage of demand — he references a billion-dollar pipeline — but scaling a semiconductor company across multiple product generations in parallel requires a different caliber of hire. He notes that roughly half the platform team came from Nvidia, and the VP overseeing hardware previously ran Nvidia's HGX and DGX programs, which accounted for most of Nvidia's revenue during the Hopper and Blackwell cycles.

The cap table

The investor list is unusually eclectic for a semiconductor startup. Participants include Jane Street, Jump Trading, Two Sigma, HRT, Founders Fund (Peter Thiel), Geoffrey Hinton, Scott Wu, Patrick O'Shaughnessy, Zach Dell, and Scott Belsky. The round size was not disclosed. The concentration of quantitative trading firms alongside deep learning's most prominent academic figure and a range of tech investors suggests Etched has made a credible case that inference infrastructure is a large, near-term market rather than a speculative hardware bet.

Etched's near-term customer target is all of the above — hyperscalers, frontier labs, neoclouds, and open-source deployments. The pitch in every case is the same: faster tokens, lower cost per token, full-stack ownership.