3 min read
Etched, a four-year-old chip startup founded by Harvard dropouts, is reportedly in talks to roughly quadruple its valuation to around $20 billion, while simultaneously raising a separate round led by Sequoia at a mere $10 billion. Two rounds, two prices, one very big bet. And the bet is genuinely audacious: Etched has hard-wired the transformer, the specific architecture behind basically every large language model, directly into its silicon.
Why that is clever
Nvidia’s GPUs are generalists. They can run more or less any kind of AI maths, which is exactly why they cost the earth and burn power like a small town. Etched’s chip, called Sohu, throws away that flexibility and does precisely one thing, run transformers, spectacularly fast and cheap. For the inference workloads that now dominate real-world AI usage, a specialist that only speaks transformer could in theory run circles around a general-purpose GPU on both speed and cost. The company says it is already sitting on around $1 billion in customer demand.
Why that is terrifying
Because it is a bet that the transformer stays king forever, and forever is a very long time in AI. Bake an architecture into hardware and you win enormously if the field stands still, and you are left holding a very expensive paperweight if it moves. The entire industry is, right now, actively hunting for the transformer’s successor. If one of those bets lands in 2028, Etched’s silicon does not get a software update, it gets a display case in a museum. In fairness, that risk is the whole reason the upside is so large: nobody pays you $20 billion for a safe bet.
The market-watcher’s note (not investment advice)
The back-to-back-rounds trick, raising at one valuation from one set of investors while raising at a very different one from another, is a very 2026 move and a small tell in itself. It lets a company bank the cash and the headline number at the same time. Etched also has to survive not just Nvidia’s hardware but Nvidia’s software ecosystem, plus custom chips from Google, Amazon, Meta and Microsoft. It is a spicy wager on a single idea in a market stuffed with trillion-dollar incumbents. Thrilling if it pays off, brutal if it does not, and either way a useful barometer of just how much money is now willing to bet against Nvidia.
The AI tool stack actually worth paying for
One email a week. The tools, models and moves that matter, minus the hype and the horseshit filter set to maximum. Free.
Did you know: the transformer architecture that Etched has physically carved into silicon comes from a single 2017 Google research paper titled “Attention Is All You Need”. A nine-page academic PDF now underpins a multi-trillion-dollar industry and, apparently, a $20 billion chip company. Not a bad return on a weekend’s writing.