Etched's AI Chip Startup Hits $21 Billion Valuation Days After Jane Street Puts a Rack in Production

Jane Street doesn't buy things it hasn't stress-tested. That's what makes this deal different from the usual AI hype cycle.

The quantitative trading firm took delivery of an inference cluster from chip startup Etched last month, ran live workloads through it, and liked the results enough to lead a $700 million funding round announced Tuesday, August 18, according to TechCrunch and a company release distributed via GlobeNewswire. The round values Etched at $21 billion. Jane Street also confirmed it now has its own rack running in its data center.

Etched was worth $5 billion in December 2025. It raised a $300 million Series C at $10.3 billion on July 23, led by Sequoia Capital, according to TechCrunch and startupfortune. Now, roughly 26 days later, the valuation has doubled to $21 billion, an increase of nearly $11 billion in under a month. Total capital raised across the company's history sits at approximately $1.9 billion, according to Tech Times.

Why Jane Street's Order Matters More Than the Number

Plenty of AI hardware startups raise money on roadmaps, renderings and benchmark claims. Etched now has something rarer: a sophisticated buyer that actually ran its own workloads on the silicon and wrote a check afterward.

Jane Street said in the funding announcement, distributed via GlobeNewswire, "We tested the chip and are pleased with the early results. Etched's unique approach to inference delivers the precision we will need to support our most demanding workloads. We're excited to now have our own rack running in our data center."

Jane Street trades on speed. Microseconds of latency in its systems translate directly into profit or loss, according to Tech Times. If Etched's chips couldn't hold up under that kind of pressure, Jane Street wouldn't have deployed them, let alone led a nine-figure round.

The Technical Bet: One Job, Done Faster

Etched's chip, called Sohu, is an application-specific chip built only to run transformer-based AI inference, not training, and not general computing. Co-founder and COO Robert Wachen told TechCrunch that inference breaks into two phases: "prefill," where the system parses a prompt and its context, and "decode," where it generates the answer token by token.

Etched built a low-voltage prefill chip that packs in more transistors without overheating, according to TechCrunch and Yahoo Finance, plus a new memory architecture it calls "cluster-scale memory" that lets many chips share a single fast memory pool. The pitch is straightforward: general-purpose GPUs from Nvidia typically use only 30 to 40 percent of their theoretical compute capacity on inference tasks because the hardware wasn't built for that specific job, according to Tech Times.

Etched's most aggressive claims, including a stated 20x inference speed advantage over Nvidia's H100 and the assertion that one Sohu server can replace 160 H100 GPUs, come from the company itself and haven't been independently verified in market conditions beyond the Jane Street deployment, as startupfortune noted. A single customer's satisfaction is a real signal, but it's not the same as proof the architecture holds up across every workload type Etched hopes to serve, including the massive mixture-of-experts models and non-transformer designs it says are already running on its clusters, according to GlobeNewswire.

Etched also says it's moved past its original pitch of "etching" one specific model into silicon. TechCrunch and Yahoo Finance both note the company is still shedding that early perception; its systems now claim to run any frontier AI model, not a single fixed one.

Who's Funding the Bet

The round drew participation from Kleiner Perkins, Sequoia Capital, Andreessen Horowitz, Tiger Global, Bain Capital Ventures, Neo, Stripes, Primary, Positive Sum, Diffusion, Argo and Blackstone, according to GlobeNewswire and TechCrunch. Peter Thiel is also listed among investors; Etched's founders, three Harvard dropouts, started the company under a Thiel Fellowship, according to startupfortune.

Kleiner Perkins Managing Partner Mamoon Hamid, who has also backed Anthropic, Databricks and Waymo, framed the bet in broader terms in the GlobeNewswire release, pointing to AI infrastructure spending estimates reaching $7 trillion by 2030.

What Happens Next

Etched says it has secured more than $1 billion in customer contracts beyond Jane Street, spanning frontier AI labs and cloud providers, and is now building three generations of hardware chips in parallel, according to GlobeNewswire. CEO Gavin Uberti said in the same release that the first rack took three years to deliver from scratch, adding, "Our next one will be much faster."

A single validated deployment at Jane Street doesn't prove Etched's chips outperform Nvidia's GPUs across the wide range of AI workloads the company is chasing, from frontier model providers to cloud operators. The next real test isn't the valuation. It's whether the additional billion-dollar customer contracts convert into working, deployed racks the way Jane Street's did.