Etched, a startup building chips purpose-built to run a single AI model architecture, has raised $800 million in fresh funding that values the company at roughly $5 billion. The round, led by trading giant Jane Street and a venture arm affiliated with Taiwan Semiconductor Manufacturing Company, is one of the largest dedicated AI chip financings of the year and signals that a contrarian bet on specialized silicon is gaining serious institutional support.
The company, founded two years ago by a small group of engineers who previously worked on transformer accelerators at major labs, has staked its entire product strategy on what it calls a “locked architecture” approach. Rather than building a general-purpose GPU competitor, Etched’s flagship Sohu chip is designed from the silicon up to execute one specific model architecture at maximum throughput. The trade-off is narrower applicability, but the company claims performance gains that are difficult for general-purpose hardware to match.
According to Bloomberg, the latest funding brings Etched’s reported forward-order book to roughly $1 billion in committed customer purchases, a figure that would have been unthinkable for a chip startup at this stage just two years ago. The company says it expects first silicon to sample with select customers in the fourth quarter of this year, with volume shipments beginning in early 2027.
Why a locked architecture matters
The pitch behind Etched’s bet is that the AI workload has converged. Most of the meaningful progress in model capability over the past three years has come from variations on the transformer architecture, and that convergence has created an opening for silicon that optimizes for one workload rather than trying to be flexible across many. General-purpose GPUs excel at flexibility, but they pay a steep efficiency tax when workloads are predictable.
Etched claims that by stripping out the flexibility, its chip can run the target architecture at multiple times the tokens-per-second of comparable GPUs while drawing a fraction of the power. The numbers are aggressive enough that several large customers have reportedly placed non-binding reservations just to lock in early access, even before production silicon has been demonstrated to external evaluators.
The investor lineup
The composition of the round is itself a signal. Jane Street, the algorithmic trading firm, has emerged as one of the most active strategic investors in AI infrastructure, with stakes in several chip companies and model labs. Its presence in the Etched round suggests a view that the economics of AI compute are about to bifurcate, with general-purpose providers competing on flexibility and specialized providers competing on unit economics.
The TSMC-affiliated venture arm’s participation is perhaps more telling. TSMC manufactures the vast majority of advanced AI chips today, and its venture investments tend to track where the company believes future wafer demand will go. A TSMC-linked bet on a locked-architecture startup is a quiet signal that the foundry sees room in its roadmap for high-volume specialty designs alongside the dominant GPU business.
“Specialty silicon is no longer a curiosity. It is a category. The same way that ARM processors coexisted with x86 in the data center, we expect to see a tier of purpose-built AI chips that don’t try to be GPUs at all.”
Competitive landscape
Etched is not the only company pursuing a specialized-silicon strategy. Cerebras, Groq, SambaNova, and a long tail of well-funded startups have all made similar arguments, and several have already shipped production hardware. The difference Etched is betting on is the depth of the lock-in: most competitors retain the ability to run multiple model architectures with varying degrees of efficiency, while Etched has explicitly given up that flexibility in exchange for extreme performance on the target workload.
The bull case is that this trade-off becomes more attractive as model architectures stabilize. If the transformer family remains dominant for the foreseeable future, customers have an economic incentive to buy the chip that runs transformers best, even if it cannot run anything else. The bear case is that the next architectural shift, whatever form it takes, will leave locked-architecture chips stranded with the wrong workload hardwired in.
What customers are signing up for
Etched’s reported customer pipeline includes a mix of large model labs, inference providers, and at least one major cloud platform. The company has been deliberately quiet about specific names, in part because some of its largest reservations are still conditional on silicon performance meeting published targets. The funding round gives the company enough runway to absorb the cost of any early silicon revisions without renegotiating customer contracts.
For the broader AI infrastructure market, the Etched round is the latest data point in a story that has been building for two years: general-purpose GPUs are not the only game in town, and the economics of running frontier models at scale are starting to favor purpose-built hardware for specific deployment profiles. Whether Etched becomes the leading example of that trend or merely an early one will depend on how the next twelve months of silicon and software development unfold.
For now, the company has the capital, the manufacturing partner, and a customer pipeline large enough to validate the business model in principle. The hard part, as always with chip companies, is execution.