Etched burned the transformer into silicon.

01 · The Room Right NowThe bet on a single architecture.

Walk into Etched's San Jose office in 2026 and you will not find a research lab. You will find a manufacturing schedule. Wafers booked at TSMC. Memory contracts with Rambus. A floorplan with one chip drawn on it - and only one. The company has roughly 340 people, a reported $5 billion valuation, and a product that, as of this writing, has not shipped a single unit to a paying customer.

That is the entire story in one paragraph. Everything else - the dorm room, the Thiel Fellowship, the press release that called Sohu "the fastest AI chip of all time" - is decoration around a single, slightly unfashionable conviction. Transformers won. The hardware should admit it.

02 · The ProblemGenerality has a price, and someone is paying it.

The modern GPU is a marvel of indecision. It can train a vision model on Monday, route traffic for a recommendation system on Tuesday, render a video game on Wednesday, and serve a 70-billion-parameter language model on Thursday. The cost of that flexibility is a chip mostly idle, mostly waiting, and mostly expensive.

This is fine when you do not know what workload will win. It becomes very expensive once you do. By 2022, it was reasonably clear: nearly every model anyone cared about - GPT, Claude, Gemini, DALL-E, Stable Diffusion - was a transformer. The hardware industry, for entirely sensible commercial reasons, kept selling a Swiss Army knife to people who only ever needed the corkscrew.

Etched's founders looked at the FLOPs utilization numbers on transformer inference - the embarrassing percentage of a GPU that actually does work at any given moment - and decided the gap between possible and actual was a business. A large one.

03 · The Founders' BetThree dropouts with one slide.

Gavin Uberti, Chris Zhu, and Robert Wachen met at Harvard, where they did the thing Harvard alumni rarely admit to doing in writing: they left. Their pitch, by accounts of investors who heard it in 2022, contained one slide that mattered. It argued Moore's Law was finished, that general-purpose performance gains had run out of road, and that the only honest path forward was specialization. Pick one architecture. Build for it. Win or lose.

This is, of course, the kind of argument that sounds reckless until it sounds inevitable. The Thiel Foundation took the risk first - all three founders were named Thiel Fellows in the same class, an unusual concentration for the program. Then Primary Venture Partners and Positive Sum led a $120 million Series A in June 2024. Then, in January 2026, Bloomberg reported a $500 million round at a $5 billion valuation.

The bet has a name: Sohu. It is a chip that does not run convolutional networks, recurrent networks, or anything else. If transformers go out of fashion, Etched goes with them. The founders are aware. They appear unconcerned.

04 · The ProductOne chip. One job.

Sohu is a transformer-only ASIC, fabricated on TSMC's 4-nanometer process, with 144 gigabytes of HBM3 memory per chip. The matrix multiplication patterns specific to transformer inference are hard-wired into the silicon - the math that a GPU would orchestrate in software is, on Sohu, the floorplan.

The performance claim is the headline: an 8-chip Sohu server, Etched says, can produce more than 500,000 tokens per second running Llama-70B. An equivalent 8-GPU Nvidia H100 system runs the same workload at roughly 23,000 tokens per second. The math implies a single Sohu server replaces about 160 H100s.

Sohu ships with an open-source software stack. The pitch to developers is that you keep your model code - the FlashAttention kernels, the TensorRT-LLM tricks, the careful batching - and Sohu just runs it faster. The pitch to CFOs is simpler: fewer servers, less power, lower bills.

05 · The ProofWhat is actually there, and what isn't.

Here is what Etched has: TSMC capacity, a partnership with Rambus for the memory interface, a team of around 340 engineers, a Series B round that closed in a market where most chip startups could not get a meeting, and a chip design that has been taped out. The investors include Peter Thiel, Stanley Druckenmiller, Two Sigma Ventures, Primary Venture Partners, and Positive Sum Ventures.

Here is what Etched does not yet have: a single customer benchmark, an independent third-party number, or a Sohu server in a production data center anyone has confirmed. As of March 2026, the chip has not shipped. The valuation is essentially an expression of belief - in the founders, in the architecture, in the math of specialization - rather than a record of revenue.

This is the fork in the road. If Sohu hits its specs, Etched will be the most consequential semiconductor company since the founding of TSMC. If it misses by 30%, it is still likely the most cost-effective transformer hardware on the market. If it misses by more than that - well, the chip industry is a graveyard of companies that promised the same chart with different bars.

06 · The MissionHardware for things that don't exist yet.

"Building hardware for superintelligence" reads like a marketing line until you look at what it actually implies about the product roadmap. Real-time voice agents. Interactive video that renders frame-by-frame from a prompt. Models that batch through tokens fast enough that the latency disappears as a category. None of these are GPU-friendly workloads at scale. All of them are, in the Etched telling, Sohu's natural habitat.

The mission is therefore not "make GPUs cheaper" - it is "make inference so cheap that applications that are currently impossible become trivial." Whether you find this thrilling or terrifying probably tells you something about your relationship with AI in general. Etched does not appear to be ambivalent.

07 · Why It Matters TomorrowThe shape of an industry, one chip at a time.

If Etched is right, the next decade of AI compute will not be a single curve of bigger general-purpose GPUs. It will be a Cambrian explosion of specialized chips - one for transformers, one for diffusion, one for state-space models, one for whatever comes next - and the companies that pick the right architecture early will eat the margins of the ones that didn't.

If Etched is wrong, the cautionary tale writes itself. A new architecture emerges. The world moves on. A very expensive chip sits in a warehouse, exquisitely designed for a problem no one has anymore. That is the nature of the bet, and Etched has the unusual virtue of not pretending otherwise.

Back to the room.

Return to the San Jose office. The floorplan is still on the wall. The wafer schedule is still booked. The 340 people are still working on one chip. What has changed is the room around them: the funding round has closed, the valuation has cleared $5 billion, and the next question is no longer whether anyone believes the pitch. The next question is whether the chip works. Etched has spent three years preparing to be measured. Soon, finally, it will be.

Where to find Etched

Further reading + watching