On Tuesday a four-year-old chip company announced it had raised another $700 million and was now valued at $21 billion. Etched had been worth $5 billion in December, $10.3 billion in late July, and roughly double that by 18 August. In venture terms, that is less a step up than a launch.
The round was led by Jane Street, the famously secretive quantitative trading firm, which did something more telling than sign a cheque. It tested Etched’s hardware, liked what it saw, bought a rack, installed the thing in its own data centre, and then led the funding round. Which leaves a $21 billion valuation resting substantially on a single named, paying customer.
What Etched actually builds
Etched makes silicon tuned specifically for AI inference, the number-crunching that happens after you hit enter on a prompt, and sells it as complete systems it calls frontier inference clusters. Nvidia, the incumbent it is aiming at, calls its own versions AI factories. Etched’s pitch is speed and cost. Its co-founder and chief operating officer, Robert Wachen, says the firm designed two components from scratch: a low-voltage chip for the compute-heavy “prefill” stage that reads your prompt, and a shared, low-latency memory pool for the “decode” stage that writes the answer.
The selling point that got reversed
Etched’s original idea, and the source of its name, was to bake a single AI model permanently into the chip. That premise has now been abandoned. Wachen says the systems run any frontier model, which is commercially sensible in a world where the leading model changes every few weeks, though it does leave the company with a memorable name for a product it no longer sells. Founded in 2022 by the Harvard dropouts Gavin Uberti, Chris Zhu and Wachen, Etched claims more than $1 billion in signed customer contracts and counts Kleiner Perkins, Sequoia, Andreessen Horowitz, Peter Thiel, Tiger Global and Blackstone among its backers.
Reasons to keep a hand on your wallet
The enthusiasm is not hard to understand. Inference is where the money in AI is shifting, away from the enormous one-off cost of training a model and towards the relentless per-query cost of running it. If Etched can genuinely undercut Nvidia on that, $21 billion could look cheap.
The scepticism is not hard to understand either. The valuation nearly doubled in a month on the strength of one delivered rack and one marquee customer. Inference-specialist chips are a crowded field, with Groq, Cerebras and SambaNova all chasing the same promise, and Nvidia sells not only chips but an entire software ecosystem that customers are loath to leave. A single rack humming away in Jane Street’s data centre is a fine proof point. It is a long way from mass production, and $21 billion is an enormous amount of faith to place on the distance in between. (Not investment advice.)
Did you know: Etched’s $21 billion valuation now tops that of several long-listed semiconductor companies, despite the startup having delivered its first full system to a paying customer only this month.
Sources
- TechCrunch: Etched’s valuation doubles to $21B in a month
- SiliconANGLE: Etched raises another $700M
- Quartz: Etched raises $700m at $21b valuation
Related on Top Tool Stack: Walmart Beat Its Own Numbers and the Market Sold It Anyway · A Different Gas Might Make Quantum Chips Far Easier to Build