Editor's note (July 9, 2026): After publication, Etched told us it has retired the "Sohu" name (no replacement announced yet) and now describes its product as a full rack-scale AI inference system rather than a transformer-only chip, one the company says is not limited to transformer models, with formal naming and specifications still to come. The reporting below reflects Etched's earlier, publicly documented positioning through mid-2026; we have updated the framing accordingly and will cover the new system once detailed specifications are available.
In 2023, Gavin Uberti and Robert Wachen spent the year pitching Silicon Valley's biggest investors with a 30-page memo. Its argument: AI would eventually outgrow general-purpose GPUs and need chips built for one specific architecture. Every major investor they pitched passed.
The two Harvard dropouts kept going anyway. By their own account, told later on the "Invest Like the Best" podcast, Etched was operating month-to-month in those early days, reportedly close to running out of cash.
On Tuesday, the same company announced it has booked $1 billion in contract orders for its inference systems, and revealed a December funding round that valued it at five billion dollars.
The memo did not change. The market did.
The Pitch Is a Chip That Refuses to Do Almost Everything
Etched's original chip, which it called Sohu, was an ASIC: an application-specific integrated circuit, silicon designed to do exactly one job instead of many. That chip's one job was running inference for transformer models, the architecture behind nearly every frontier LLM in production today.
That specialization was the entire bet. By Etched's early description, the chip could not run most traditional AI workloads, including deep learning recommendation models and recurrent neural networks. Strip away the flexibility a GPU carries, the argument went, and what remains runs transformers faster, cheaper, and with better power efficiency.
The company's early performance claims were aggressive. Etched said an eight-chip server could serve more than 500,000 tokens per second running Llama 70B, and Uberti told TechCrunch in 2024 that one such server would replace 160 Nvidia H100 GPUs. Those numbers are the company's own, from its original Sohu-era materials, and it has not cited independent verification for them.
What is no longer just a claim: TSMC successfully manufactured the chip earlier this year on its 4nm process, and Etched says it is now testing complete systems with customers.
The Numbers Etched Just Put on the Table
| What Etched announced | Detail |
|---|---|
| Booked contract orders | $1 billion for "frontier inference clusters" |
| Total raised to date | $800 million |
| Latest round | $500 million, closed quietly in December |
| Post-money valuation | $5 billion |
| Manufacturing | Chip produced by TSMC on its 4nm process |
| Product status | Full systems in customer testing; first product not yet generally shipping |
The product itself is not a bare chip. Etched sells what it calls frontier inference clusters: bundles of its custom chips, custom-designed racks, and software, aimed at labs and companies serving models at scale.
The Cap Table Reads Like an AI Awards Dinner
Stripes led the December round. The wider investor group includes VentureTech Alliance, Jane Street, Hudson River Trading, Two Sigma, and Ribbit Capital. Angel investors include Andrej Karpathy, Geoffrey Hinton, Fei-Fei Li, Mistral CEO Arthur Mensch, and Cognition CEO Scott Wu. Stanley Druckenmiller and Peter Thiel are also on the cap table.
Two things stand out in that list. The AI researchers who know transformer internals best put personal money behind silicon originally built to run transformers. And three of the most sophisticated quantitative trading firms in the world, Jane Street, Hudson River Trading, and Two Sigma, made the same bet. Those firms live and die on compute economics.
The reason is the line item that dominates every AI company's books. Inference, the work of actually answering prompts after a model is trained, is now the biggest cost center and the biggest bottleneck for anyone serving models at scale. Whoever makes inference meaningfully cheaper collects rent on the entire industry.
Everyone Is Suddenly Building the Same Thing
Etched's announcement lands in the middle of an inference gold rush.
Cerebras had the first breakout AI chip IPO of the year after OpenAI doubled its Cerebras order to $20 billion.
Groq confirmed a $650 million raise in June and re-staffed after Nvidia's not-acqui-hire deal. OpenAI unveiled Jalapeño, its first custom inference chip, built with Broadcom in nine months. Amazon, Google, and Microsoft all build in-house AI silicon, and Microsoft's Maia 200 just landed Anthropic as its first external customer.
Against that field, Etched built as the purist. Where most rivals hedged with some level of general-purpose capability, Etched's original design welded its architecture to the transformer. The company now says its current hardware is not limited to transformer models.
The Fine Print Nobody Should Skip
A billion dollars in bookings is not a billion dollars in revenue. The figure represents forward contracts for systems still in customer testing, and Etched's first racks were not expected to ship until this summer. Contracts can be renegotiated, delayed, or cancelled if hardware underdelivers.
The performance numbers deserve the same caution. The 500,000 tokens-per-second figure and the 160-H100 comparison come from Etched's own materials, and the H100 comparison dates to 2024, before Nvidia shipped Blackwell and Vera Rubin.
The deeper risk is the one Etched chose at the outset. A transformer-only chip is a bet that the architecture stays dominant for the useful life of the silicon. That bet looked safe for years. It looks slightly less safe every time a credible alternative emerges, and Google's DiffusionGemma, which generates 256 tokens at once using diffusion instead of autoregression, is exactly the kind of research that keeps ASIC designers up at night. Etched now says its current system is not limited to transformers, which, if borne out by the eventual specifications, would soften this risk. If the frontier moves, a GPU gets repurposed; a fixed-function rack does not.
Even the "coming out of stealth" framing in Tuesday's press release earned a raised eyebrow: as TechCrunch noted, the founders have been publicly discussing their chip plans with reporters since 2024.
The Bottom Line
Etched now has what it lacked in 2023: TSMC-manufactured silicon, a billion dollars in signed demand, and backers who understand both the math and the money. What it does not yet have is a single rack running in production at a paying customer. The next two quarters, when those frontier inference clusters either perform at claimed levels or do not, will decide whether Etched is the next Cerebras or an expensive monument to overfitting on an architecture.
In 2023, a 30-page memo arguing that transformers would eat AI compute could not buy a meeting. In 2026, the same memo is collateral for a billion dollars in contracts. The argument never changed. The believers did.
Sources
- Nvidia competitor Etched hits 5B valuation, 1B in sales for AI chip — TechCrunch, June 30, 2026
- Etched.ai raises 500m for a 5bn valuation – report — Data Center Dynamics, January 19, 2026
- AI chip startup Etched raises 500 million to take on Nvidia — Bloomberg, January 13, 2026
- Etched is building an AI chip that only runs transformer models — TechCrunch, June 25, 2024
- Etched unveils Sohu chip and first inference system, plans summer shipments — Crypto Briefing, 2026
- AI chipmaker Groq confirms 650M raise, re-staffs after Nvidia's 20B not-acqui-hire deal — TechCrunch, June 22, 2026
- OpenAI unveils its first custom chip, built by Broadcom — TechCrunch, June 24, 2026