$1 billion in pre-orders for a chip that runs exactly one kind of model. Etched just raised $300 million at a $10.3 billion valuation to bet that the transformer, and only the transformer, is worth etching into silicon.
The chip, Sohu, hard-wires transformer attention into fixed circuits. It trades flexibility for speed on the one workload that now dominates production AI.
The $1B order book is genuine. No third party has published independent benchmarks yet. That gap between commercial conviction and public evidence is the whole story.
TIMELINE: Etched — from three dropouts to a $10.3B chip
─────────────────────────────────────────────────────────────
2022 ────── Dec 2025 ────── Jun 2026 ────── Jul 2026
🎓 💰 🏭 ◉ NOW
Founded $500M @ $5B First silicon $300M @ $10.3B
Harvard round manufactured Sequoia-led
dropouts + $1B orders Series C
Etched, AI2Work / TechCrunch / Value Add VC, July 2026
A chip built to do one thing
In 2022, three Harvard undergraduates dropped out to build a chip that could run exactly one thing: the transformer. It looked like a dare. Convolutional and recurrent networks still dominated production work, ChatGPT had not launched, and burning a single architecture permanently into silicon struck most chip engineers as a fast route to an expensive paperweight.
Four years later the dare has been repriced. The transformer is now the substrate of essentially every frontier model in production. Etched's answer, Sohu, is an application-specific integrated circuit (ASIC). It takes the transformer computation graph (attention, linear projections, softmax) and turns it into fixed-function circuits instead of instructions a general-purpose GPU has to fetch and schedule.
Inference, the process of generating tokens after a prompt, splits into two stages: prefill and decode. It designed for each separately. A low-voltage prefill chip handles the first pass; a cluster-scale memory system lets many chips pool memory for the token-by-token decode stage. The company sells full rack systems, not loose silicon.
The efficiency argument is straightforward. A GPU is programmable, which costs it overhead. Sohu gives that up. In exchange it can run transformer inference faster and with less power for as long as the architecture stays fixed. The founders' favourite precedent is bitcoin mining: once purpose-built SHA-256 ASICs arrived, GPU mining became economically obsolete within months.
The numbers behind the round
Sequoia led the $300 million Series C, joined by Andreessen Horowitz, SK Hynix, Jane Street and Diffusion Capital. In December Etched was valued at $5 billion after a $500 million round. Seven months later it is worth $10.3 billion.
Etched says it has booked more than $1 billion in orders and has manufactured its first chips, with full systems now in testing at customers. The company runs a two-megawatt data centre and has opened an 80,000-square-foot, ten-megawatt facility in Milpitas. The new capital is aimed at moving Sohu from working silicon to volume deployment against that order book.
That investor list reads as a supply-chain vote. SK Hynix is one of the world's largest memory makers, and a chip whose design depends on memory bandwidth wants its memory supplier inside the tent. The Wall Street Journal reports a parallel round in discussion at around a $20 billion valuation, led by Jane Street.
Why specialised silicon is suddenly the battleground
This is where the industry's weight has moved. Training was the expensive phase for years. Inference is now the recurring cost, and it scales directly with usage. Every agentic workflow, every coding assistant, every document pipeline multiplies inference calls per user action. The token-per-second ceiling of the hardware decides how many concurrent workloads a system can run within acceptable latency.
Etched is the most extreme version of a broader argument. The hyperscalers that buy the most compute are building their own inference chips — Google TPU, AWS Trainium, Meta MTIA, Microsoft Maia. Four days before the Series C, The Information reported Google is developing a server chip codenamed Frozen v2 that would etch part of Gemini's architecture into silicon, with engineers projecting six to ten times more tokens per unit of power than Google's newest TPUs. The logic runs through all of them: the general-purpose GPU, optimised for everything, is now the expensive bottleneck for the one thing everyone actually runs.
As we wrote in August, the cost-per-token economics of the ASIC-versus-GPU war are the real story underneath the funding headlines. The valuation is the market putting a number on that argument.
The risk welded into the die
Sohu's strength is inseparable from its constraint. The chip cannot run convolutional networks, LSTMs, state-space models or any non-transformer design. This is a physical property of the silicon, not a configuration choice.
The company says its systems do run mixture-of-experts models like DeepSeek and Qwen, and even non-transformer state-space designs such as Mamba. That widens the surface. But the core bet is unchanged: transformers remain the dominant production architecture for the useful life of the hardware, typically three to five years for data-centre equipment.
The evidence today favours that bet. Every frontier model in production is transformer-based, and no alternative has yet replaced it at the frontier. But favourable today is not locked in for five years. Hybrid architectures mixing attention with linear recurrence are an active research area. An ASIC is only faster than a GPU for as long as the architecture it encodes stays fixed. A GPU never has to guess which design wins.
The second risk is the benchmark gap. Etched reports $1 billion in booked orders and dramatic performance claims, all vendor-sourced. Nothing has been independently measured in a published third-party test. For an investor the question is whether the efficiency advantage survives production conditions, not the demo.
What changes if the bet pays off
If Sohu's performance holds in the field, the economics of running large language models at enterprise scale shift. Inference cost becomes a hardware problem instead of a GPU-pricing problem. That breaks the assumption, decades old, that a single silicon architecture serves every AI workload.
The order book says a swathe of buyers believe inference demand is large and stable enough to justify specialised chips. The valuation says capital agrees. The open item is whether the transformer stays the architecture everyone wants to run long enough for the wafer to pay for itself. That is not a technology question. It is a timeline question.
Racks start shipping to select clients this summer. General availability pricing has not been announced. Between now and published production benchmarks, the honest position is this: Etched has moved the industry's conversation from whether inference hardware will fragment to who will win the first contest.
What to watch next
Three things will test the thesis. The first independent benchmark of Sohu in a production environment, not a vendor demo. Named customer contracts beyond the $1 billion order book, which shows who actually deploys first. And a parallel Jane Street-led round reported in discussion near a $20 billion valuation, which would test how far the market is willing to extend the bet.
Published results that confirm the claimed inference throughput make custom silicon a durable third force in AI compute. Results that miss turn the $10.3 billion mark into a cautionary example of funding ahead of evidence.
The takeaway
The round is a bet that inference, not training, is where AI's compute spending settles, and that the architecture under all of it is stable enough to freeze into hardware. The money, the order book and the memory-maker on the cap table all point the same way.
The chips are manufactured. The systems are in testing. The benchmarks are not yet public. In the meantime, the fastest AI chip companies have moved from racing each other to racing the architecture itself.
