Here's a number to start your evening with: $10.3 billion. That's the new valuation of Etched, a five-year-old chip startup founded by three Harvard dropouts — and it's more than double what the company was worth just seven months ago (TechCrunch).

The fresh $300 million Series C, led by Sequoia, is — per Reuters — the highest valuation ever for a Sequoia-led Series C (Reuters via Yahoo). Not bad for a company whose entire pitch is a chip that deliberately does less than Nvidia's.

That's the twist worth sitting with. In a market where every accelerator races to be more flexible, Etched went the other way — it hard-wired a single AI architecture into silicon and threw everything else out. And investors just paid a record price for that stubbornness.

The thesis: the AI hardware war is quietly splitting in two, and Etched is betting the entire future of chips on the half Nvidia was never built for.

🧠 Why This Matters

Every AI chip you've heard of — Nvidia's H100, its Blackwell line, AMD's MI-series — is a generalist. It can train models and run them. That flexibility is the whole point of a GPU, and it's why Nvidia is worth trillions.

Etched's argument is that flexibility is now a tax. The world has settled on one dominant AI architecture — the transformer, the "T" in ChatGPT — and if you're only ever going to run transformers, why pay for silicon that can do anything else? Its flagship, Sohu, is billed as "the world's first transformer ASIC": a chip that can't train models, can't run yesterday's architectures, and does exactly one job — inference — at full tilt (Tech Startups).

"The infrastructure required to serve frontier AI sustainably and economically was never going to come from incremental improvements to existing hardware."

— Gavin Uberti, Etched co-founder and CEO (SiliconANGLE)

It's a genuinely contrarian bet, and the money following it is the news. When Sequoia writes its largest-ever Series C check for a chip that intentionally narrows its own market, that's a signal about where the smart money thinks AI spending is headed: not toward training the next model, but toward serving the ones we already have, billions of times a day, as cheaply as physics allows.

📊 Deep Dive

Etched was started in 2022 by Gavin Uberti (CEO), Robert Wachen (COO) and Chris Zhu (CTO), all of whom walked out of Harvard to build it. The company now has roughly 400 employees and just opened an 80,000-square-foot facility near its San Jose headquarters to build and test hardware (Reuters via Yahoo).

The pitch isn't just the transformer bet — it's two pieces of engineering aimed squarely at the economics of inference:

  • Low-voltage inference (LVI): the chip runs at a lower voltage than a typical GPU, generating less heat, which lets Etched push higher clock speeds and pack transistors more densely (SiliconANGLE).
  • Cluster Scale Memory: every accelerator in a rack shares one pool of memory instead of each chip hoarding its own copy — killing the data duplication that bloats big-model serving.
  • Inference-only by design: no training silicon, no legacy-architecture overhead — the two-stage inference flow (prefill, then token-by-token decode) is the only thing it's built to do.
  • Real orders, not vaporware: Etched says it has already booked more than $1 billion in signed customer contracts (TechCrunch).

The backer list reads like a who's-who. Alongside Sequoia sit Andreessen Horowitz, SK Hynix, Jane Street and Diffusion Capital, with angel checks from Peter Thiel, Andrej Karpathy, Dylan Field and Amjad Masad (TechCrunch). SK Hynix's presence is the tell — a memory giant does not casually invest in a chip company whose whole thesis rests on how memory is shared.

⚠️ The Catch

Betting everything on the transformer is brilliant right up until the transformer stops being the only game in town. Etched's entire moat is architectural loyalty — and AI research is not famous for standing still. State-space models like Mamba, diffusion-based language models, and whatever comes next could erode the assumption the whole company is soldered onto. A general-purpose GPU can pivot; an ASIC is, by definition, etched.

Then there's the valuation itself. Doubling to $10.3 billion in seven months is a stunning number — but Etched is still shipping its first production racks, and a $1 billion order book is a promise, not revenue in the bank. The company also has to fight Nvidia's real weapon, which was never just the chip: it's CUDA, the software ecosystem millions of engineers already live inside. Faster silicon means little if it's a pain to actually deploy.

🎯 What Happens Next

Watch the racks. Etched's first inference systems — the ones built on that TSMC-manufactured silicon — are slated to start shipping in summer 2026, which is right now (SiliconANGLE). The gap between "$1 billion in signed contracts" and "$1 billion in delivered, humming hardware" is exactly where chip startups have historically gone to die. Clear it, and the $10.3 billion looks cheap. Stumble on yields or heat, and it looks like 2021 all over again.

The other thing to watch is Nvidia's shadow. If a transformer-only ASIC really does serve tokens dramatically cheaper, expect the incumbents to answer — AMD is already courting inference customers, and Nvidia has every incentive to make its own inference story airtight before a startup writes the narrative for it.

🧩 Bigger Picture

Step back and Etched is a bet on a single, sweeping idea: that AI's center of gravity has moved from building models to running them. Training grabs the headlines, but inference is the workload that repeats forever — every chat, every image, every agent call — and that's where the real, recurring compute bill lives. Etched's investors are effectively saying that the trillion-dollar prize isn't teaching AI anything new; it's answering the questions we're already asking, a billion times a second, for a fraction of a cent.

If they're right, the age of the do-everything GPU may give way to an age of specialists — chips carved for one job and terrifyingly good at it. If they're wrong, they've built a very expensive monument to a single architecture. Either way, three former college students just convinced the world's sharpest investors to pay a record price to find out.

In a market obsessed with chips that can do anything, Etched just got rich betting on one that does exactly one thing — and dares the rest of the world to keep up.

Sources