Etched $20B Valuation: The AI Chip Bet on Nvidia

The market’s appetite for anything that could loosen Nvidia’s grip on AI compute just repriced a startup by a factor of four in a matter of weeks. According to reporting on July 20, 2026, Etched, a semiconductor company building a chip designed to run — not train — AI models, is in talks to raise new capital at a valuation of roughly $20 billion. The company is separately raising money at about a $10 billion valuation in a distinct round. Neither deal has closed, and the terms could still shift.

The jump is steep by any measure. As recently as June 30, 2026, Etched exited stealth at a $5 billion valuation, disclosing about $800 million raised to date and roughly $1 billion in customer contracts for its chip. A $20 billion mark would be a fourfold step-up in under a month.

The two rounds

The larger financing — the one that would set the $20 billion valuation — is reportedly being led by Jane Street, the quantitative trading firm and an existing Etched investor. A separate round at approximately $10 billion is said to be led by Sequoia Capital. Running two rounds at two different prices in quick succession is unusual, but it has become a signature move of the current AI cycle: a company sells a stake at one valuation and almost immediately raises more at a far higher one, as demand from investors outruns the company’s need for cash.

Etched has told investors it is still testing its initial chip design and working to validate its first product. That is worth underlining. The valuation is being set on a chip that, by the company’s own account, has not yet shipped in volume — a bet on the roadmap rather than the revenue.

What Etched actually makes

Etched was founded in 2022 by three Harvard dropouts — Gavin Uberti, Chris Zhu, and Robert Wachen. Its product is a chip called Sohu, and the pitch rests on a single architectural wager: rather than build a general-purpose accelerator like a GPU, Etched hardwired the transformer — the architecture behind virtually every large language model — directly into the silicon.

The tradeoff is deliberate. A GPU can run almost any workload, which makes it flexible but leaves performance on the table for any single one. An application-specific chip like Sohu can run only transformers, but in exchange it can pack far more of the exact math those models need onto the die. Sohu is fabricated on TSMC’s 4-nanometer process and carries 144GB of HBM3E high-bandwidth memory per chip — the same memory class used on Nvidia’s top data-center parts.

Etched’s performance claims are aggressive. The company says a single 8-chip Sohu server can produce more than 500,000 tokens per second running Llama-70B, versus roughly 23,000 tokens per second for an equivalent 8-GPU Nvidia H100 system — and that one Sohu server could stand in for as many as 160 H100s on transformer inference. Those are the company’s own figures. As of the funding reports, there is no independent benchmark of Sohu against Nvidia’s current-generation hardware, no named launch customer, and no disclosed volume for the summer shipments Etched has promised.

Why the inference angle matters

The logic behind the raise is about where the money in AI is moving. Training a frontier model is a one-time, capital-intensive event. Inference — actually serving the model to users, over and over, every time someone sends a prompt — is a recurring cost that scales with usage. As AI products reach hundreds of millions of users, inference is becoming the larger and stickier share of the compute bill.

That is the seam Etched is trying to pry open. Nvidia dominates training and is deeply entrenched in inference too, but inference is also where specialized silicon has the clearest shot: the workloads are more predictable, the models are converging on the transformer, and buyers care intensely about cost-per-token and power draw. Etched is not alone in seeing this. Qualcomm has pushed data-center inference parts, AMD is pressing its MI-series accelerators against Nvidia, and China’s DeepSeek is reportedly designing its own inference chip to cut its dependence on foreign silicon. The hyperscalers, meanwhile, keep building in-house.

The backers tell a story

The investor list is as much a part of the pitch as the chip. Jane Street leading the marquee round signals that sophisticated, quantitatively minded capital is willing to underwrite a hardware bet at a venture-stage price. Sequoia leading the smaller round adds a blue-chip venture imprimatur. Etched’s earlier backers reportedly included a TSMC-linked venture firm, which — if the manufacturing relationship holds — matters enormously for a fabless startup that needs guaranteed capacity at a leading-edge foundry.

Capacity is the quiet constraint behind all of this. HBM supply remains tight amid the broader memory supercycle, and advanced-node wafer allocation at TSMC is spoken for years out. A chip startup’s valuation is only as real as its ability to actually build and ship at volume — and both of those inputs are scarce.

What it means

The Etched round is a clean read on where AI-hardware money is flowing and how fast. Investors are paying up — a 4x markup in under a month — for even a credible shot at Nvidia’s inference franchise, before the challenger has shipped a product in volume or posted an independent benchmark. That tells you two things: the market believes inference is the bigger long-term prize, and it believes specialized silicon is the most plausible route to contesting it.

For Nvidia, one pre-revenue startup at $20 billion is not a threat to the quarter. But the pattern is the story. Etched, Qualcomm, AMD, DeepSeek, and every hyperscaler’s in-house team are all aiming at the same target, and each new mega-round makes it easier for the next to raise. Nvidia’s moat has always rested as much on CUDA and software lock-in as on raw silicon; a transformer-only ASIC that only ever needs to run transformers is precisely the kind of design that tries to route around that moat.

The risks cut hard the other way. Etched’s entire thesis is that the transformer stays dominant long enough to justify freezing it into hardware. If model architectures shift — toward new attention variants, mixture-of-experts at extreme scale, or something not yet invented — a chip that can only do transformers ages fast, while a GPU adapts. There is execution risk on every axis: tape-out, yield, foundry allocation, HBM supply, software tooling, and landing customers willing to bet production traffic on unproven silicon. Back-to-back rounds at climbing valuations also raise the stakes of a stumble; a company priced for a fourfold move has little room to disappoint.

What to watch next: whether either round actually closes at the reported marks, the first independent third-party benchmark of Sohu against current Nvidia parts, a named launch customer with real inference volume, and any signal on how much TSMC capacity and HBM Etched has actually secured. Valuations can be printed in a press cycle. Shipping silicon that beats Nvidia on cost-per-token — at volume — is the part that will take longer to prove.