Etched Raised $300 Million to Carve One Idea Into Silicon. If That Idea's Wrong, the Chip Is a Paperweight.
The purest bet in AI hardware just doubled its price in seven months. The upside and the risk are the same sentence.
Nvidia sells a chip that can run almost anything. Etched built one that runs almost nothing, and just raised $300 million on it.
Here are the numbers. On July 23, Etched told TechCrunch it closed a $300 million Series C at a $10.3 billion post-money valuation, led by Sequoia with Andreessen Horowitz, Jane Street, SK hynix, and Diffusion participating. The company says that’s the highest valuation ever for a Sequoia-led Series C. Etched was valued at $5 billion in December, so it roughly doubled in about seven months, and total funding now sits above $1 billion. Founded in 2022 by three Harvard dropouts, the company says it holds more than $1 billion in signed customer contracts, with first racks scheduled to ship this summer.
The product is the whole story. Etched’s chip, Sohu, hard-wires the transformer straight into the silicon. Nvidia’s GPUs stay general-purpose, so they can run whatever architecture comes next. Etched threw that flexibility out on purpose to buy density and speed on the one workload it cares about: transformer inference, the step where a trained model actually answers you.
The model: the Hedgehog and the Fox.
The fox knows many things. The hedgehog knows one big thing. Nvidia is the fox, staying adaptable so it never gets stranded by a shift in the field. Etched is the hedgehog, and it’s the purest hedgehog in AI hardware right now. It has bet the entire company on a single conviction: the transformer stays king long enough for silicon built only for transformers to pay off. If that’s right, a chip stripped of everything the transformer doesn’t need should beat a general-purpose GPU on cost and speed, and $10.3 billion will look cheap. Hedgehogs win biggest when their one big thing is true.
The contrarian close.
They also lose hardest when it isn’t, and here’s the part the valuation doesn’t tell you: as of the company’s stealth exit, no independent third-party organization had published benchmarks from physical Sohu hardware under production conditions. The throughput claims, the contract figures, the $10.3 billion, those are the company’s numbers and the syndicate’s conviction, not verified results. So the bet has two blades. One is technical, will Sohu actually beat a GPU when someone outside the building measures it. The other is architectural, does the transformer stay dominant long enough to matter. You can win the first and still lose the second. A hard-wired chip can’t be patched into a new paradigm. Watch for one thing before the next up-round: an independent benchmark. Until that lands, everything here is a story about conviction, and conviction isn’t a spec sheet.
If this was useful, share it with someone who builds things. And if you want the full toolkit of 50 mental models, you can grab my book, Mental Models: How to Think, Act, and Win, on Amazon right now.
If you want the mental models behind breakdowns like this, my book, Mental Models: How to Think, Act, and Win, is on Amazon now.
This post is for informational and educational purposes only. It is not investment advice, a recommendation, or a solicitation to buy or sell any security. Funding figures, revenue, and valuation are as reported by the company, regulatory filings, and named outlets; the ~$470M valuation is a reported, time-sensitive snapshot and not independently verified. Dhoni's individual investment amount was not disclosed. The /mkt reference is a structural illustration of building in regulated markets and is not an offer or solicitation. Past performance and third-party investment decisions do not indicate future results.
