Etched Doubles to $21 Billion: Jane Street Leads $700M Round and Buys the Racks

The fastest valuation doubling in this cycle’s chip boom just happened in three weeks. On August 18, 2026, the Wall Street Journal reported that Etched, the San Jose startup building a transformer-only inference chip, has raised $700 million at a $21 billion valuation — roughly double the $10.3 billion mark it set with a $300 million Series C in late July. The lead investor is Jane Street, the quantitative trading giant. And in a twist that defines this moment in AI infrastructure, Jane Street is not just backing the company: it is also a customer, buying server racks built around Etched’s specialized silicon.

The Deal

The numbers alone tell a startling story:

• $5B valuation — December 2025/January 2026, on a $500 million round led by Stripes, after crossing $1 billion in signed customer contracts
• $10.3B valuation — July 23, 2026, on a $300 million Series C led by Sequoia Capital, with Andreessen Horowitz, Jane Street, Diffusion, Argo, and SK Hynix participating
• $21B valuation — August 18, 2026, on a $700 million round led by Jane Street

That is a 4x valuation jump in roughly eight months, and a 2x jump in under a month. Etched has now raised well over $1.5 billion across its lifetime — a remarkable pile for a company founded in 2022 by three Harvard dropouts: CEO Gavin Uberti, Chris Zhu, and Robert Wachen, who started the firm when they were barely out of their teens. The WSJ’s headline called them a “‘kids in chips’ startup,” and the label sticks precisely because the contrast is so extreme: twenty-somethings commanding a valuation that would have placed them among the most valuable semiconductor companies in the world a decade ago.

The dual role of Jane Street is the most consequential detail. The trading firm is described as “a key early customer and lead investor” — it buys server racks assembled around Etched’s chips and has now anchored two consecutive rounds. This is the customer-as-investor pattern that has become a signature of the 2026 AI financing cycle: the same circular structure seen in OpenAI’s arrangements with Nvidia and SoftBank, and Anthropic’s share-based acquisition of Decart. When your biggest customer funds your expansion to deliver more product to itself, valuation and demand become difficult to disentangle.

What Etched Actually Builds

Etched’s flagship product is Sohu, an application-specific integrated circuit (ASIC) built on TSMC’s 4nm process that does exactly one thing: run transformer neural networks. Where an Nvidia GPU is a general-purpose parallel processor that can, in principle, execute any architecture you compile for it, Sohu hard-wires the computational patterns specific to transformers — the attention mechanisms and matrix multiplications that power GPT, Claude, Gemini, Llama, and nearly every mainstream large language model — directly into silicon.

The company’s claims are aggressive. Etched has said a Sohu server can serve Llama-70B at over 500,000 tokens per second, a more-than-10x improvement over Nvidia’s flagship Blackwell GPUs, and earlier materials claimed up to 20x inference throughput versus the H100. The logic is straightforward: if you delete the flexibility a GPU needs to support arbitrary workloads, you can spend every transistor and every watt on the one workload that dominates the market. Eight Sohu chips, the company has argued, could replace 160 GPUs for transformer inference.

The bet has an obvious flip side. If the industry’s frontier models move meaningfully beyond transformers — toward state-space models, diffusion-based language models, or hybrid architectures — a chip that only runs transformers becomes an expensive mistake. Etched is wagering that the transformer’s six-year dominance of the model landscape persists through the chip’s useful life. So far, that wager has looked reasonable: even the newest frontier reasoning models remain transformer-based.

Poaching Silicon Valley’s Best

The WSJ report emphasizes another dimension of the story: Etched is “scooping up Nvidia talent.” The company now employs roughly 400 people, drawing engineers from Nvidia, Broadcom, Google’s TPU team, and SK Hynix. For a firm whose founders had never shipped a commercial chip when they started, importing veterans from the companies that built the modern accelerator industry was the only credible path — and the money to do it arrived remarkably fast.

Notably, SK Hynix — one of the world’s three memory giants and a critical supplier of the high-bandwidth memory (HBM) that every AI accelerator depends on — is both an investor and, implicitly, a strategic partner. In a market where HBM allocation is a genuine bottleneck, a chip startup with a memory giant on its cap table holds a card that most rivals do not.

The Inference-Silicon Shuffle

Etched’s round lands amid a broader reshuffling of the inference-chip landscape. Groq, the LPU maker that once commanded a $6.9 billion valuation, recently closed at $3.5 billion — roughly half its prior mark — after striking a supply arrangement with Nvidia. Cerebras went public and has traded on earnings beats. AMD bought Taalas to hard-wire model-specific silicon. And the hyperscalers — Google’s TPU, Amazon’s Trainium, Microsoft’s Maia — keep expanding their internal silicon programs to blunt Nvidia’s pricing power.

Against that backdrop, Etched’s trajectory is the exception: valuations rising, not compressing, in a window where investors have grown distinctly choosier about AI infrastructure. The difference is the specialization argument. General-purpose inference plays are being squeezed from both sides — by Nvidia above and by hyperscaler in-house chips below — but a chip that is architecturally unmatched for the single workload that consumes the most tokens has a clearer, if narrower, path to economics that beat both.

The Open Questions

Skepticism remains, and it is worth stating plainly:

Customer concentration. Etched says it holds more than $1 billion in signed customer contracts, but neither the company nor the WSJ has named a customer other than Jane Street. A $21 billion valuation resting on one visible reference customer — however deep-pocketed — is a structure that deserves scrutiny. As AI Weekly’s analysis put it, the next tell is whether a second production customer is named before this round officially closes.

Verification pending. Sohu’s performance claims are company numbers. Until independent benchmarks or customer-deployed systems are publicly measured, the 10x claims should be treated as directional rather than proven.

Valuation math. Even if all $1 billion-plus in contracts converts to revenue on schedule, $21 billion represents a multiple that assumes near-monopoly economics in a segment Nvidia has every incentive to defend — Blackwell’s successor roadmaps, aggressive pricing, and CUDA’s ecosystem lock-in are not static targets.

Why It Matters

Strip away the eye-popping numbers and the real signal is this: transformer-only ASICs have moved from research curiosity to a live procurement question. Any organization planning inference capacity for 2027 now has to weigh a future that is not CUDA-only. When a quantitative trading firm — an institution whose entire business is calculating expected value with brutal discipline — both funds the company at $21 billion and writes purchase orders for its racks, that is a stronger demand signal than a pitch deck.

Etched’s founders are betting their careers that the future of AI compute looks like a chip that does one thing, perfectly. Jane Street is betting billions that they’re right. Within weeks, the market will get its next data point on whether the doubling continues — or whether the kids in chips finally meet physics, competition, and the limits of a very specific architectural wager.