In July 2026, Etched closed a $300 million Series C at a $10.3 billion valuation — up from $5 billion in December 2025. For a hardware startup founded in 2022 by three Harvard dropouts, doubling valuation in seven months is the kind of number that deserves a look at what is underneath it.
The Numbers
- ›$300 million Series C at a $10.3 billion valuation, July 2026
- ›Up from a $5 billion valuation in December 2025
- ›Led by Sequoia Capital
- ›Joined by Andreessen Horowitz, SK Hynix, Jane Street and Diffusion Capital
- ›Angel investors including Peter Thiel, Andrej Karpathy and Dylan Field
- ›Roughly $1 billion in booked pre-orders
The SK Hynix name on that list is the interesting one. Memory bandwidth, not raw compute, is the binding constraint on inference at scale — a memory manufacturer taking a position in an inference chip company tells you where they think the bottleneck is.
What They Actually Built
Etched's system splits inference into its two natural halves and builds different hardware for each. There is a prefill chip — the phase where the model ingests your prompt — which runs at a lower voltage than competing parts. Lower voltage means less heat, and less heat means you can pack more transistors into the same thermal envelope. Then there is the decode side, handled by what the company calls cluster-scale memory: a memory system built for the token-by-token generation phase, where the workload is memory-bound rather than compute-bound.
That split is the whole thesis. Prefill and decode have genuinely different performance characteristics, and a GPU running both is a compromise optimised for neither.
Not Transformer-Only After All
Etched was widely described early on as building transformer-specific silicon, and the obvious objection followed: bake one architecture into hardware and you are one research breakthrough away from an expensive paperweight. The company now says that framing was wrong. COO Robert Wachen has stated the systems support any AI model — including mixture-of-experts designs like DeepSeek and Qwen, and non-transformer architectures such as Mamba.
The original criticism was the right question to ask. That the answer changed is worth noting honestly, rather than pretending the concern was never reasonable.
Why Specialised Inference Silicon, and Why Now
Training gets the headlines; inference gets the bill. Once a model is deployed, the cost of serving it runs continuously for as long as anyone uses the product, and at current volumes that operating cost dominates the economics of every AI company. A chip that meaningfully lowers cost per token is attacking the largest recurring line item in the industry — which is why a company with limited public benchmarks can book a billion dollars of orders.
The Skepticism Is Still Fair
Etched has not published performance figures against Nvidia parts, and its customer list is not public beyond a reference to some of the largest AI companies in the world. The founders faced plenty of doubters even after successfully manufacturing silicon, partly because so few people have hands-on access to the hardware. A $10.3 billion valuation is a claim about the future, not evidence about the present — and until independent benchmarks exist, that distinction is the whole story.
Written by Manas Garge
Founder & Data Engineer