Skip to content

Etched's Valuation Quadrupled to $20 Billion. Its Chip Still Hasn't Shipped.

DS
LDS Team
Let's Data Science
7 min
The AI chip startup is negotiating two venture rounds at once, one led by Jane Street at $20 billion and a separate one led by Sequoia Capital at half that, according to the Wall Street Journal. Etched's own website says its first chip is still being validated. Customers have committed a billion dollars anyway.

In 2023, Gavin Uberti and Robert Wachen carried a 30-page memo around Silicon Valley and got turned down by every investor who read it. The pitch: general-purpose GPUs would eventually run out of road, and someone needed to build a chip that did one thing, running transformer models, better than a GPU could do a hundred things. Nobody bit. By their own account, the company was running month to month.

Three years later, two of the world's most selective investors are competing to hand Etched a fortune for a chip that has not shipped a single unit to a paying customer.

According to Wall Street Journal reporters Kate Clark, Anissa Gardizy, and Robbie Whelan, Etched is simultaneously negotiating two venture rounds: one led by existing backer Jane Street that would value the company at $20 billion, and a separate round led by Sequoia Capital at roughly half that figure. Neither deal had closed as of July 17. The higher number would quadruple the valuation investors assigned Etched just months ago.

Two Term Sheets, Two Prices, One Company

The structure itself is the story. Back-to-back financings, where a startup raises money at one price and returns for a second round at a sharply higher price within the same year, have become routine in the current AI investment cycle. Etched is doing something more unusual: running two of them at once, with two different lead investors arriving at two different numbers for the same asset in the same month.

Jane Street, the quantitative trading firm that already backed Etched's prior raise in June alongside a TSMC-linked fund, is leading the richer of the two rounds. Sequoia Capital, one of the most prominent venture firms in Silicon Valley, is leading the other at roughly $10 billion. Both are still being negotiated, and the Journal reported that terms on each could still change before signing.

That dynamic tells its own story about who holds the advantage in this negotiation. Investors are not haggling Etched down. They are competing to get into whichever round closes first, at whichever price gets them an allocation before the next one goes even higher.

The Product Behind the Price Tag Is Still Being Tested

What makes the number remarkable is what it is being attached to. By Etched's own account on its website, cited by the Journal, the company is still testing its initial chip design and working to validate its first product. It has not shipped.

Despite that, Etched says customers have already committed $1 billion in demand for systems built around the chip, called Sohu. Sohu is an application-specific integrated circuit: silicon hardwired to do one job, running inference for transformer-based models, instead of the general-purpose flexibility a GPU offers. Etched has claimed an eight-chip Sohu server can replace up to 160 Nvidia H100 GPUs, a figure founder Gavin Uberti gave TechCrunch in 2024. Etched has not published independent benchmark verification of that claim, and none of the coverage of this financing round cites third-party testing of a finished unit.

TSMC did manufacture Sohu on its 4-nanometer process earlier this year, and Etched has said it is now testing complete systems with early customers, a milestone this outlet reported in July alongside the company's prior $5 billion valuation. The gap between "manufactured and being tested" and "commercially validated" is exactly where this new round is being priced.

The Field Etched Is Trying to Break Into Is Already Crowded

Nvidia still supplies the overwhelming majority of chips used to train frontier AI models. But training and inference are different jobs, and a growing list of challengers is betting that inference, running a model that is already built, is where purpose-built silicon can beat a general-purpose GPU on cost and speed.

Cerebras Systems and Groq have already shown commercial traction in that market. Newer entrants, including the UK's Fractile and SambaNova, are chasing the same customers. Etched's bet is narrower than any of them: Sohu cannot run the deep learning recommendation models or recurrent neural networks that still make up a meaningful share of production AI workloads. It is built to do exactly one thing.

CompanyFocusStatus as of July 2026
NvidiaTraining and general-purpose inferenceMarket incumbent, dominant share
Cerebras SystemsWafer-scale training and inferenceCommercial deployments live
GroqInference-optimized chipsCommercial deployments live
EtchedTransformer-only inference (Sohu)Chip manufactured, still in validation
Fractile, SambaNovaInference-focused alternativesCompeting for the same customers

Why Investors Are Paying Ahead of Proof

The bull case is straightforward: a billion dollars in customer commitments for a chip that has not shipped is a stronger signal of real demand than a chip that ships to no buyers. Enterprises reserving Sohu capacity now are betting that general-purpose GPU supply will stay constrained and expensive well into 2027, and they would rather lock in inference capacity early than wait for a fully proven product and find the line already full.

The risk case deserves equal weight, since no source in this financing round states it directly: a $20 billion price tag assumes Sohu performs in production the way Etched's own marketing describes it, and that assumption has not yet been tested outside Etched's contracts with its own early customers. Etched's TSMC manufacturing run and its billion-dollar backlog are real and verifiable. Its performance claims, including the 160-GPU replacement figure, remain the company's own, unverified by any independent lab cited in this round of reporting.

What This Means for Teams Planning Inference Capacity

For engineering teams deciding how to build out inference infrastructure, Etched's backlog is a signal worth tracking, not a benchmark to build a procurement plan around yet. The company's own admission that it is still validating its first product is the more decisive fact for anyone evaluating vendor risk today.

The broader pattern is one worth watching across AI infrastructure generally. Chip financing rounds are increasingly being priced on projected demand and investor conviction rather than shipped, independently benchmarked hardware, a dynamic also visible in Qualcomm and Tenstorrent's recent RISC-V chip partnership and in Nvidia's own GTC keynote, where Jensen Huang used the company's dominant position to preview chips years ahead of shipping.

The Bottom Line

Etched went from a rejected pitch deck to a possible $20 billion valuation without shipping a chip to a single customer. That is not necessarily proof of a bubble. Inference demand is real, GPU supply is tight, and Jane Street and Sequoia do not write checks this large casually.

But the honest version of this story is that two of the most sophisticated investors in the world are pricing a company on a product that does not yet exist in commercial form, and the last time this specific bet gets tested is not in a term sheet. It is in production, on someone else's workload, months from now.

Sources

Practice interview problems based on real data

1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.

Try 250 free problems