Etched Doubles to $21B as Jane Street Bets on Transformer-Only Silicon

Etched, the two-year-old startup building a chip that runs only transformer models, announced on August 18 that it has raised $700 million at a $21 billion valuation — more than doubling its worth in less than a month. The round was led by Jane Street, the storied quantitative trading firm, which also happens to be the first customer for Etched’s Sohu inference chip.

The speed of the re-rating is the story’s headline, but the mechanism behind it matters more. Etched closed a $300 million Series C at a $10.3 billion valuation in late July. Three weeks later, the company was worth double that. In a market where AI infrastructure valuations are supposed to be cooling, Etched went the other way — and it did so on the strength of real purchase orders rather than benchmark hype.

What Etched Actually Builds

Etched’s core bet is radical specialization. Sohu is an application-specific integrated circuit (ASIC) that hard-codes the transformer architecture — the “T” in ChatGPT — directly into silicon. It cannot run convolutional networks, legacy recurrent models, or anything that isn’t a transformer. That constraint is the feature: by stripping out generality, the chip dedicates nearly all of its transistors, memory bandwidth, and power budget to transformer inference.

The technical profile is aggressive. Sohu is fabricated on TSMC’s 4-nanometer N4P process and carries 144 GB of HBM3E memory per chip, with roughly 1.8 times the memory bandwidth of comparable GPU offerings, according to industry reporting on the company’s published materials. Etched has claimed throughputs north of 500,000 tokens per second on a 70-billion-parameter model in an eight-chip server — a figure that, if it holds up in production, would represent a step change in inference economics.

The company says it has now raised $1.9 billion in total funding and booked more than $1 billion in customer contracts, with shipments of Sohu planned for the second half of 2026. Its roughly 400-person engineering team is drawn heavily from NVIDIA, Google’s TPU group, Broadcom, SK Hynix, and TSMC — a hiring pattern that signals the company is competing for exactly the same talent pool as the incumbent it hopes to undercut.

Why Jane Street Led — and Bought

The most telling detail in the round is that Jane Street is both lead investor and first customer. The quant firm tested Sohu hardware and committed to deploying it before committing capital. In a market flooded with AI chip startups making grand throughput claims, a sophisticated buyer putting its own trading infrastructure on the line is a stronger signal than any benchmark chart.

Jane Street’s use case is instructive. Quantitative trading is a domain where inference latency translates directly into money — models that parse news, order flow, and market microstructure a few milliseconds faster than competitors capture real edge. A chip optimized purely for transformer inference, promising lower latency and better cost-per-token than general-purpose GPUs, maps neatly onto that need. If Sohu delivers, Jane Street gains an infrastructure advantage its rivals can’t easily replicate; the investment hedges the strategic bet.

WSJ reporting on the round dubbed Etched a “$21 billion ‘kids in chips’ startup,” noting its aggressive poaching of NVIDIA engineering talent. The skepticism embedded in that phrasing is fair: Etched has shipped silicon to partners but has yet to prove volume production, and its entire thesis rests on the transformer architecture remaining dominant for years.

The Risk: Architecture Lock-In

Every ASIC bet is a wager on architectural stability. Etched’s chip is valuable precisely because it does one thing — but if the frontier moves to a fundamentally different architecture (state-space models, hybrid designs, or whatever succeeds transformers), Sohu becomes an expensive doorstop. The counterargument: transformers have been the backbone of every major frontier model since 2017, and the entire ecosystem — training frameworks, optimized kernels, inference providers — is built around them. Inertia of that magnitude doesn’t reverse quickly.

There’s also the NVIDIA question. The GPU giant’s dominance rests on generality and CUDA’s software moat, but it has been moving toward domain-specific acceleration with each generation. Etched’s pitch is that NVIDIA can’t go far enough: a company serving every workload can’t burn its architecture into silicon the way a specialist can.

The Broader Shift Toward Inference Silicon

Etched’s re-rating is also a data point in a larger industry migration. As AI spending shifts from training behemoths to serving tokens at scale, inference is becoming the dominant workload — and inference is where specialization pays off fastest, because the model architecture is frozen at serving time. That logic has already minted winners in adjacent niches: Groq built a business around deterministic low-latency inference, Cerebras around wafer-scale serving, and the hyperscalers all run their own inference accelerators. What distinguishes Etched is the purity of its bet and the velocity of its valuation curve — from a $5 billion private mark in December 2025 to $10.3 billion in July to $21 billion now.

Notably, much of the new money is coming from financial markets rather than traditional venture funds. Jane Street leading a chip round is the clearest sign yet that buy-side firms now see AI inference infrastructure as a strategic asset class — closer in spirit to building a proprietary exchange than to backing a software startup. Expect more of this: the firms with the shortest feedback loops between model latency and profit are the natural first customers for inference-first silicon.

The market’s verdict so far: investors are paying $21 billion for that specialization thesis. Whether the second half of 2026 vindicates them depends on Sohu shipping on time, at claimed throughput, into production workloads — starting on Jane Street’s trading floor.