The startup Etched has raised $300 million in a Series C and reached a $10.3 billion valuation. That is double its December 2025 figure. The company designs chips specialized for AI inference, meant to run without GPUs. Sequoia Capital led the round, announced on July 23, 2026.
Key takeaways
- Series C: $300 million, at a $10.3 billion valuation
- Valuation doubled from $5 billion in December 2025 — in about seven months
- Sequoia Capital led; investors include SK Hynix, Andreessen Horowitz and Jane Street
- $1 billion in pre-orders; first silicon from TSMC in June 2026
- The chip supports transformers, Mixture of Experts and Mamba models — without GPUs
A bet against the GPU
Etched was founded in 2022. Gavin Uberti, Robert Wachen and Chris Zhu started it — all three Harvard dropouts. They built a chip specialized for one job: running finished AI models. It is not a general-purpose processor like Nvidia's GPU. It is an inference?inference: running a finished AI model, as opposed to training it ASIC?ASIC: a chip designed for a single task, not a general-purpose processor like a GPU. The company states flatly that GPUs are no longer needed to accelerate inference.
Two parts: prefill and decode
The system splits in two. The Prefill Chip processes the prompt and context using "low-voltage inference" — running at reduced voltage. That limits heat and lets the design pack in more transistors. The second part is Decode Memory: "cluster-scale memory" that lets many chips share memory at low latency. That speeds up token generation. Etched claims support for transformers, Mixture of Experts, Mamba, and state-space models.
The scale of the money
The Series C is $300 million. The valuation jumped to $10.3 billion from $5 billion in December 2025. That is a doubling in about seven months. Sequoia Capital led the round.
Andreessen Horowitz, SK Hynix, Jane Street and Diffusion Capital joined. Individual backers include Peter Thiel, Andrej Karpathy, Dylan Field and Amjad Masad. The presence of SK Hynix, a memory maker, is telling — memory is the bottleneck in inference.
Demand and manufacturing
The company cites $1 billion in booked orders. TSMC produced the first silicon in June 2026. Etched employs around 400 people. Alongside its San Jose office and a 2 MW data center, it is adding a new site in Milpitas: 80,000 square feet and 10 MW of power.
The skeptics are still here
When Etched started, specialized AI chips drew distrust. GPUs were the default, and they stayed that way. Even after silicon shipped successfully, the skeptics did not vanish. A co-founder himself tempers the hype.
I think we still have to be humbled by what it will take to actually get to scale.
Robert Wachen, co-founder and COO, Etched.
Why it matters
A $10.3 billion valuation for a company without mass shipments is a bet on one thesis: that AI's future is inference, not training, and that a specialized chip serves inference cheaper and faster than a general-purpose GPU. If the thesis holds, Nvidia loses part of its most profitable market and model operators gain a cheaper alternative. If it is wrong, Etched is left with expensive silicon tuned for architectures that may change.
The key point is that the company bets not on one model but on a whole class of workloads — transformers, MoE and state-space models. That spreads the risk tied to architecture shifts.
The presence of SK Hynix and TSMC around the startup shows the supply chain takes the bet seriously. The real test, though, is not the valuation but whether $1 billion in orders turns into working, deployed systems.
What's next
- Etched is opening its Milpitas site (80,000 sq ft, 10 MW) — added capacity for building systems.
- $1 billion in orders awaits fulfillment — deliveries of full systems will test whether the startup can scale.
- Silicon from TSMC (June 2026) is entering testing with early customers and investors.
