Etched Doubles to a $21 Billion Valuation in a Month as Jane Street Puts Its Sohu Racks to Work

Etched, the San Jose startup building transformer-only AI inference chips, has pulled off one of the fastest valuation jumps in the current AI cycle: a $700 million round led by Jane Street that values the company at $21 billion — roughly double the $10.3 billion it was worth when it closed its $300 million Series C just weeks earlier, and up from $5 billion as recently as December 2025.

The round, announced August 18, 2026, came with proof that the company’s bet is starting to pay off in the real world. Etched shipped its first rack-scale inference system to Jane Street last month, and the quantitative trading firm — now also its lead investor — has confirmed the rack is fully operational in its own data center, running live production workloads on the company’s Sohu chips.

From stealth to $21 billion in eight months

The pace is disorienting even by AI-boom standards. Etched — founded by Harvard dropouts Gavin Uberti and Robert Tsiang — re-emerged from stealth in July 2026 having raised roughly $800 million to date and booked more than $1 billion in signed customer contracts for Sohu, its application-specific chip (ASIC) that runs exactly one kind of model: transformers. By July 23, the company had closed a $300 million Series C at a $10.3 billion valuation, led by Sequoia Capital with participation from a16z and others.

Twenty-six days later, the valuation doubled again.

The catalyst was not a slide deck but silicon in a rack. Jane Street ran demanding, high-precision workloads on the delivered system, liked what it saw, and led the new $700 million round. “Etched shipped its first rack last month to Jane Street, and the quantitative trading firm is actively deploying the technology into its [production infrastructure],” Business Insider reported. Customer contracts now top $1 billion.

What makes Sohu different

Etched’s core wager is that the era of general-purpose AI silicon is ending. GPUs like Nvidia’s H100 and B200 are designed to run anything — convolutional nets, recurrent models, whatever a researcher dreams up. Sohu throws that flexibility away. It is a transformer-only ASIC, hardwired for autoregressive LLM inference and nothing else.

That specialism buys throughput. Etched claims an eight-chip Sohu server can push more than 500,000 tokens per second on Llama 70B — versus roughly 23,000 tokens per second for a comparable eight-GPU H100 cluster. Each chip carries 144GB of HBM3 memory, and the company says its co-design of chips, racks, software, and manufacturing delivers best-in-class throughput, latency, cost, and power efficiency for frontier models.

The trade-off is existential: if the industry moves past transformers, Sohu becomes an expensive doorstop. Etched is betting years of silicon roadmap and more than a billion dollars in customer commitments on the transformer architecture remaining the dominant paradigm for AI inference — a bet that looks considerably safer in 2026 than it did when the company first pitched it in 2024.

Why a quant fund is leading

Jane Street leading the round is the detail worth pausing on. Quantitative trading firms are among the most demanding compute buyers on the planet: latency-sensitive, throughput-hungry, and allergic to marketing. A quant shop doesn’t invest $700 million in a chip startup out of ecosystem goodwill — it does so because it has run the workloads itself and sees an economic edge.

That pattern — customer-turned-lead-investor — is becoming a signature of this AI infrastructure cycle. When the buyer of your product is also the buyer of your equity, due diligence and deployment are the same act. For Etched, it converts a marquee reference customer into a balance-sheet validator, and it gives Jane Street both a stake in the upside and privileged access to supply in a market where AI inference capacity is scarce.

The inference land grab

Etched now competes in a crowded, capital-intensive lane. Cerebras is pushing wafer-scale inference systems; Groq champions deterministic low-latency inference; the hyperscalers build their own accelerators; and Nvidia — whose Blackwell and Rubin roadmaps dominate training — keeps expanding its own inference story. Etched differentiates by delivering full rack-scale systems, not just chips, and by specializing so narrowly that it can undercut general-purpose hardware on cost-per-token for transformer workloads.

The $21 billion question is volume. A $1 billion contract book is impressive for a company that shipped its first rack a month ago, but it is small next to the tens of billions the hyperscalers spend annually on AI infrastructure. Etched must now prove it can manufacture, deliver, and support rack-scale systems at a rate that satisfies Wall Street-grade expectations — with a customer base that now includes the most demanding quant firm in the market watching every token.

What the doubling signals, more broadly, is where late-stage capital believes the AI story is heading. Training gets the headlines, but inference is where the recurring revenue lives. Every agent loop, every chat, every video-generation call is an inference bill — and investors are paying ever-steeper multiples for credible alternatives to Nvidia at that layer. Etched’s $5B → $10.3B → $21B trajectory in eight months is the market pricing that shift in real time.

For now, the transformer-only bet has its first paying rack, its first $21 billion price tag, and the unusual distinction of being validated not by benchmark charts but by a customer who wired the money itself.