A Chip Startup Doubled Its Value in a Month by Being Deliberately Inflexible

Etched, a chip company in San Jose, announced on Tuesday that it has raised 700 million dollars at a valuation of 21 billion. The round was led by Jane Street, the trading firm that is also Etched’s first customer and that took delivery of its first rack of hardware last month. Kleiner Perkins, Sequoia, a16z, Peter Thiel, BCV and Blackstone all participated. The company says it has signed more than a billion dollars in customer contracts across AI companies and cloud providers.

The numbers move fast enough to be worth laying out in order. Etched was valued at 5 billion dollars in December. It raised a 300 million dollar round at 10.3 billion in July. Four weeks later it is at 21 billion. Two pieces of technology carry the pitch: Low Voltage Inference, which the company says packs more computing into the same power budget, and Cluster Scale Memory, a hybrid memory design that pools memory across an entire cluster of chips instead of keeping each chip to its own.

Here is the idea underneath, in plain terms. A graphics processor of the sort Nvidia sells is a general-purpose machine: it will run any model you throw at it, this year’s architecture and next year’s. Etched builds an ASIC, an application-specific chip, and it has hardwired one specific design, the transformer, directly into the silicon. Every model you have heard of, from GPT to Claude to Gemini, is a transformer. By giving up the ability to run anything else, you can strip out enormous amounts of general-purpose machinery and go much faster and much cheaper on the one thing that matters. The bet is not subtle: it is that transformers will still be what everybody runs in five years. If a genuinely different architecture wins, the chips become expensive paperweights.

Worth noting what the investors are actually buying. Inference, meaning the cost of running a model to answer your question, has quietly become the dominant expense in this industry, larger than training for anyone with real users. Every cent shaved off a million tokens compounds across billions of requests. That is why capital is piling into inference silicon specifically, and why a lead investor who is also the first customer is a meaningful signal rather than a conflict: Jane Street put the rack in production before writing the cheque.

What this means for you: you will never buy one of these, and nothing about your tools changes this week. What it explains is a trend you have already felt, which is that the price of using AI keeps falling while the models keep getting better. That is not generosity, it is hardware competition working its way through to the bill. Over the next year or two, expect the cheap end of the model market to get noticeably faster and cheaper again. Expect the marketing around it to be considerably louder than the actual delivered chips.

Sources

• TechCrunch: Etched’s valuation doubles to 21 billion dollars in a month
• GlobeNewswire: Etched raises 700M at a 21B valuation and completes first customer delivery to Jane Street
• Dealroom: Etched hits 21B valuation with 700M Series D led by Jane Street