Every AI chip pitch for the last decade has bragged about flexibility. Ours can run anything, they say. Vision, speech, recommendation engines, whatever comes next. Etched built the opposite of that — a chip that runs exactly one kind of model and falls over on everything else — and investors just handed the company $300 million for it.

Let's get the number on the table first: $300 million, at a $10.3 billion valuation, in a Series C led by Sequoia with Andreessen Horowitz, SK Hynix, Jane Street and Diffusion piling in alongside ([TechCrunch](https://techcrunch.com/2026/07/23/ai-chip-startup-etched-defies-skeptics-hits-10-3b-valuation-from-big-name-investors/)). That roughly doubles the company's valuation from its previous round near $5 billion.

The bet is narrow on purpose. Etched's Sohu chip does transformers — the architecture underneath basically every large language model you've used — and nothing else. The thesis: if the whole world settles on one design, stop building silicon that hedges against the alternatives.

🧠 Why This Matters

Nvidia's GPUs are magnificent generalists. That's the point and, Etched argues, the weakness. A general-purpose chip spends transistors, power and die area staying ready for workloads it may never run. When 90-plus percent of your compute is one repeated operation — the matrix math inside a transformer — all that flexibility is overhead you're paying for and cooling.

An ASIC, a chip hard-wired for a single job, throws the generality out and spends everything on the one thing. Etched's claim is that this trade is now worth making because the architecture stopped moving. GPT, Claude, Llama, Gemini — different labs, same transformer backbone. If that holds, a transformer-only chip isn't a gamble on the future. It's a bet the present is done changing.

And demand is showing up before the hardware does. Etched says it is validating rack-scale systems to fulfill more than $1 billion in customer commitments ([Business News Today](https://business-news-today.com/etched-raises-300m-as-10-3bn-valuation-tests-ai-inference-chip-economics/)) — orders on the books for a product that hasn't shipped at scale.

📊 Deep Dive

Sohu is built on TSMC's 4nm process with 144GB of HBM3E memory per chip, and Etched frames it purely as an inference part — the chip that runs a trained model, not the one that trains it ([MLQ](https://mlq.ai/news/etched-raises-300m-series-c-at-103b-valuation-to-scale-gpu-free-inference-chips/)). The headline performance claim is the whole sales pitch:

  • Etched's number: roughly 500,000 tokens per second on Llama-70B for a single Sohu server
  • Its Nvidia comparison: about 25,000 tokens per second for an equivalent H100 setup — a claimed ~20x gap
  • The catch on that number: it's Etched's own benchmark, not an independent one
  • The moat it's attacking: Nvidia's CUDA software stack, which locks most of the industry in regardless of raw silicon speed

The people behind it are almost a caricature of the genre: Gavin Uberti, Chris Zhu and Robert Wachen, three Harvard dropouts who started Etched in 2022. Sequoia had reportedly passed on them early; it's now leading their round. The chips are fabricated at TSMC, and the roughly 400-person company runs facilities in California and Taiwan. The backer list reads like a group text of Silicon Valley — Peter Thiel, Andrej Karpathy, Figma's Dylan Field and Replit's Amjad Masad are all in ([TechCrunch](https://techcrunch.com/2026/07/23/ai-chip-startup-etched-defies-skeptics-hits-10-3b-valuation-from-big-name-investors/)).

"When you really think something's possible, and you just work at it for a long time, you can do it." — Robert Wachen, COO, Etched

⚠️ The Catch

The entire company is one architectural assumption. If the field drifts away from transformers — toward state-space models, diffusion-style language models, or something not yet named — a chip that can only do transformers becomes very expensive scrap. A GPU would just run the new thing slower. Sohu wouldn't run it at all.

Then there's the gap between slides and racks. That 20x figure is Etched's own benchmark, and no outside lab has published numbers to confirm it. Sohu was first unveiled back in 2024; shipping specialized silicon at volume is exactly where hardware startups tend to stall. And $1 billion in commitments is not $1 billion in revenue — pre-orders convert only if the chips arrive, land at the promised performance, and clear the CUDA switching cost that has sunk faster silicon before. Talks are reportedly already underway for a follow-on round near a $20 billion valuation, which is a lot of expectation stacked on a product still in testing.

🎯 What Happens Next

Watch three things. First, whether the first commercial racks post independently measured throughput anywhere near the claim — that's the moment the story becomes real or doesn't. Second, whether a marquee customer publicly names Sohu in production, because signed commitments stay abstract until someone runs their live traffic on it. Third, Nvidia's counter: its Rubin generation is aimed squarely at the inference market Etched is targeting, and the incumbent has never been shy about pricing to keep a lane closed.

🧩 Bigger Picture

Etched is the sharpest version of a bet spreading across the industry: that AI has standardized enough to justify hard-wiring it into silicon. Google has ridden its transformer-friendly TPUs for years; Amazon has Trainium; a wave of inference-chip startups is arguing the same thesis from different angles. What makes Etched the loud one is the refusal to hedge. Most challengers still build something a little flexible, a little future-proof. Etched deleted the escape hatch and asked investors to fund the conviction — and at $10.3 billion, enough of them said yes.

The whole company is a wager that the transformer is permanent. If it's right, Etched owns the fastest lane on the busiest road in computing. If it's wrong, it built a very expensive chip for a world that moved on. There's no partial credit on a bet this pure — the transformer stays king, or the chip becomes a museum piece.

Sources