Home Companies Etched
Specialised inference silicon Private Pre-production / early shipping

Etched

Etched made the most aggressive bet in AI hardware: burn the transformer architecture into silicon and give up the flexibility to run anything else. It has since broadened that position, and the market has repriced the company from $5 billion to $10.3 billion in seven months.

Key facts

Founded
2022
Headquarters
California; 80,000 sq ft, 10 MW facility in Milpitas
Co-founder and chief executive
Gavin Uberti
Co-founder and chief operating officer
Robert Wachen
Co-founder
Chris Zhu
Inference architecture
Specialised ASIC — split prefill compute and decode memory
Lead product
Sohu, and rack-scale inference systems
Latest valuation
$10.3B, July 2026
Pre-booked orders
Over $1B, company-stated
Last reviewed
18 August 2026

Overview

Etched was founded in 2022 by Gavin Uberti, Robert Wachen and Chris Zhu, who left Harvard to build it. The founding premise was unusually stark: rather than build a flexible accelerator, hard-wire the transformer architecture directly into the chip. A processor that can only do one thing can do that thing with far less silicon spent on generality — no instruction scheduling for operations it will never run, no datapaths for models it will never serve.

The corresponding risk was equally stark, and Etched's critics named it immediately: a chip that runs only transformers is worthless if the field moves past transformers. For its first years the company was as well known for that objection as for its technology.

Etched emerged from stealth on 30 June 2026 stating that it had raised $800 million, had working silicon, and held more than $1 billion in signed customer contracts. Three weeks later, on 23 July 2026, it closed a $300 million Series C led by Sequoia Capital at a $10.3 billion valuation — double its December 2025 valuation of $5 billion, and described by Sequoia as the highest valuation at which it has led a Series C.

The architecture, and how it has changed

Etched's public description of its product has evolved in an important way, and readers relying on earlier coverage will have an out-of-date picture.

The original framing was a transformer-only ASIC: Sohu, with headline claims that a single eight-chip server could replace on the order of 160 Nvidia H100 GPUs for transformer inference. The current framing describes two distinct components addressing the two stages of inference separately:

ComponentApproachStage addressed
Prefill chip Operates at low voltage, which reduces heat per transistor and allows more transistors to be packed into the same thermal budget. Prefill — the compute-dense, highly parallel stage that processes the input prompt.
Decode memory Cluster-scale memory with an interconnect designed to present a shared memory pool at very low latency. Decode — the sequential, memory-bound generation stage where KV cache pressure dominates.

Etched now states that its systems can run any AI model, including mixture-of-experts and non-transformer designs. If that holds under independent scrutiny it substantially weakens the central objection to the company — but it also softens the original thesis. The extreme specialisation argument was the reason to believe an order-of-magnitude advantage was achievable; a more general system has to justify its advantage on other grounds.

This publication has not independently verified Etched's model-compatibility claims, and no third-party benchmark of the current systems is available at the time of review. The shift from "transformer-only" to "runs any model" is a material change in positioning and should be treated as company-stated until independently tested.

Production status

Etched has manufactured first-batch silicon through TSMC, and rack systems are in testing with limited access granted to investors and customers. The company has opened an 80,000 square foot, 10 megawatt facility in Milpitas to support production and system integration.

What does not yet exist publicly is the evidence that matters most: there is no public cloud availability, no independently verified deployment at scale, and the disclosed contracts are forward-looking commitments rather than delivered revenue. Etched has said first racks ship during summer 2026, which places the company at the point where its claims become testable.

Strengths and key questions

Strengths

What is compelling

  • Real economic headroom. If specialisation delivers even a fraction of the claimed advantage, the cost-per-token implications for high-volume serving are large.
  • Order book. More than $1 billion in pre-booked orders is a strong signal that sophisticated buyers have evaluated the technology and committed.
  • Working silicon. The company has taped out and manufactured, which eliminates the largest single risk facing a fabless startup.
  • Investor quality. Sequoia leading at this valuation, alongside Andreessen Horowitz, SK hynix, Jane Street and Diffusion Capital, with earlier backing from Peter Thiel, Andrej Karpathy and Dylan Field.
  • SK hynix on the cap table. A memory manufacturer as investor is strategically useful when memory supply is an industry-wide constraint.
Key questions

What has yet to be proven

  • Production at scale. First silicon is not volume manufacturing. Yield, packaging, thermals and system reliability are all unproven publicly.
  • Contracts versus revenue. $1 billion in forward-looking orders is not $1 billion recognised. Conversion depends on delivery.
  • Independent benchmarks. Every published performance figure originates with the company.
  • Valuation. $10.3 billion for a company at pre-production stage prices in a great deal of successful execution.
  • Architectural drift. Even in its broadened form, a specialised design carries more exposure than a GPU to changes in dominant model architecture.
  • Software. The stack required to serve arbitrary models well is a large, ongoing engineering commitment, and it is the failure mode that has caught most specialists.

Funding history

DateRoundTerms and investors
2024Series A$120M, led by Primary Venture Partners and Positive Sum
December 2025Series B$500M at a $5B valuation, led by Stripes, with Peter Thiel, Positive Sum and Ribbit Capital
30 June 2026Stealth exitCompany stated $800M raised to date, working silicon and more than $1B in customer contracts
23 July 2026Series C$300M at a $10.3B valuation, led by Sequoia Capital, with Andreessen Horowitz, SK hynix, Jane Street and Diffusion Capital
RevenueUndisclosed. Etched is private and does not report financial results. Order book figures are company-stated commitments, not revenue.

Adding the July Series C to the $800 million the company reported in June implies total disclosed funding of approximately $1.1 billion. Etched has not published a consolidated total since the Series C, so this figure is arithmetic on disclosed rounds rather than a company-confirmed number.

Leadership

Gavin Uberti is co-founder and chief executive, Robert Wachen is co-founder and chief operating officer, and Chris Zhu is a co-founder; all three left Harvard in 2022 to start the company. The founding team is notably young for a semiconductor company, which was itself a recurring source of scepticism — and which the company's ability to reach working silicon at TSMC substantially answers.

Etched has also attracted an unusual angel roster for a hardware company, including Andrej Karpathy and Dylan Field alongside Peter Thiel, which has given the company disproportionate visibility relative to its stage.

Why Etched matters to the market

Etched is the clearest test of the specialisation thesis. If a purpose-built inference system can deliver a large, durable cost-per-token advantage over general-purpose GPUs, Etched should be the company that demonstrates it, because no one has committed harder to the premise. If it cannot — if software burden, model drift and manufacturing economics erode the advantage — that is strong evidence for the incumbency case.

The Groq outcome adds a second dimension. Groq had a genuine architectural advantage and a real customer base, and the resolution was a licensing transaction that moved the technology and team to Nvidia. Anyone assessing Etched at a $10.3 billion valuation is implicitly taking a view on whether an independent path exists, or whether the realistic ceiling for a successful inference architecture is an acquisition by the incumbent.

What to watch

  • First rack deliveries. Whether systems ship to customers on the stated summer 2026 timeline, and to whom.
  • Named customers. The $1 billion order book is currently anonymous. Named references would materially strengthen the claim.
  • Independent benchmarks. Third-party measurement, particularly on mixture-of-experts models, given the broadened compatibility claims.
  • Revenue recognition. The transition from contracted orders to delivered and recognised revenue.
  • Cloud availability. Any route to buying Etched capacity without a direct systems purchase.
  • Strategic interest. Given the Groq precedent, any approach from an incumbent would be significant for the whole sector.

Sources

  1. Etched — $300M Series C at a $10.3B valuation (23 July 2026). Company announcement of round size, valuation and investors.
  2. TechCrunch — Etched hits $10.3B valuation (23 July 2026). Founder titles, architecture description, production status and Milpitas facility.
  3. Etched — Emerges from stealth with working chip, $800M raised and over $1B in customer contracts (30 June 2026). Stealth-exit disclosures.
  4. Tech Times — Etched exits stealth. Independent reporting on the June announcement.
  5. Spheron — Etched Sohu versus Nvidia: transformer ASIC and GPU inference compared. Secondary technical analysis, including caveats on pre-production status.

Etched is a private company at pre-production stage. Every performance and order-book figure on this page is company-stated. This profile identifies them as such throughout and will be updated when independently verifiable deployment data becomes available. See the editorial methodology for how unverified claims are handled.