A four-year-old startup is making a bold run at the most valuable company in tech. Etched — founded in 2022 by Harvard dropouts Gavin Uberti and Chris Zhu — has raised roughly $500 million at a $5 billion valuation, TechCrunch reported, and says it has about $1 billion in customer orders lined up. Backers reportedly include prominent names like Peter Thiel and AI pioneers Geoffrey Hinton and Fei-Fei Li. (Whether the ~$1 billion is firm revenue or non-binding commitments isn't fully clear; treat it as booked orders.)

The bet: one job, done faster

Etched's pitch rests on specialization. Nvidia's chips are general-purpose — they can run almost any AI workload, which makes them flexible but carries overhead. Etched's chip, "Sohu," is hard-wired for one thing: "transformer" models, the architecture behind ChatGPT, Claude and most modern AI. By baking transformer math directly into the silicon (an approach called an ASIC — an application-specific chip), Etched claims it can run these models dramatically faster and cheaper.

The company's numbers are eye-catching — it has claimed a single server of its chips can do the work of a rack of Nvidia GPUs — but those figures come from Etched's own tests, not independent benchmarks, so they should be treated with caution until verified. The chip is reportedly made on TSMC's 4-nanometer process.

Why this market is the prize

The target is inference — the act of running a trained AI model to answer a query — as opposed to training, the upfront work of building it. As AI use scales, inference becomes the bigger, recurring cost: every chatbot reply, every generated image runs on inference. Cheaper inference is therefore the economic lever for AI companies drowning in compute bills — the same enormous AI-capex strain Boursel has tracked through the BIS warning and the Mag-7 selloff.

That's why Nvidia, dominant as it is, faces a crowd of challengers specifically on inference: Groq, Cerebras, SambaNova, plus the cloud giants' own in-house chips (Google's TPUs, Amazon's Trainium). Etched is the latest, best-funded entrant.

The risks are real

Specialization cuts both ways:

  • Architecture risk. Sohu is built for transformers. If AI shifts to a different model architecture, Etched's chips could be stranded — a bet on the present winning the future.
  • Nvidia's moat is software. Nvidia's CUDA ecosystem locks in developers; rivals must convince customers to re-tool, a high switching cost.
  • Execution. A working first chip is not the same as manufacturing at scale, hitting yield, and delivering the promised performance. Plenty of well-funded chip startups have stumbled here.

Why it matters

For Nvidia, Etched is a sign that the inference market — increasingly the larger half of AI compute — will be contested, even as Nvidia's grip on training remains near-total. For the AI industry, credible cheaper-inference chips could lower costs and ease the capacity crunch that's straining budgets and power grids alike. And for Etched, a $5 billion valuation is a vote of confidence — but the company now has to ship, at scale, against the most formidable incumbent in technology. Boursel offers no view on any company's value; the takeaway is that the AI-chip race is widening from "can you get Nvidia GPUs?" to "is there a cheaper way to run the model?" — and a lot of money is betting the answer is yes.