Etched Doubles to $21B One Month After Its Last Round

Etched said on August 18 that it has raised $700 million at a $21 billion valuation, double the $10.3 billion the AI-chip startup was worth in July. The round was led by Jane Street, the quantitative trading firm, and was reported by TechCrunch.

The repricing has been relentless. Etched was valued at $5 billion in December 2025. In July it raised a $300 million Series C at $10.3 billion. Six weeks later the number has doubled again — a 4.2x move in eight months, with the last leg taking about a month. Nothing about the company's product roadmap changed in that window; what changed was who had touched the hardware.

That is the genuinely unusual detail in this round. Jane Street installed and tested an Etched cluster in its own datacenter before leading the investment, and told TechCrunch it "tested the chip and are pleased with the early results." Chip startups have been repriced this year on contracts and letters of intent rather than working silicon — Fractile jumped 6x on an Anthropic supply agreement for parts that will not ship until 2027. An investor who ran the cluster and then wrote the check is a materially stronger signal than one who read a spec sheet.

What Etched sells has also shifted. The company now describes its product as "frontier inference clusters" — complete systems rather than loose accelerators. Co-founder and COO Robert Wachen says two components are proprietary: a prefill chip that runs at low voltage to handle prompt processing, and what Etched calls "cluster-scale memory," which lets multiple chips share memory pools at high bandwidth and low latency. The memory-pooling piece is the more revealing of the two. Long-context inference is bounded less by raw arithmetic than by how much KV cache a system can hold and how fast it can reach it, which is the same constraint Cerebras and every other Nvidia challenger is designing around.

There is a strategic softening buried in the pitch. Etched made its name on Sohu, a chip hardwired for the transformer architecture, and the bet was that surrendering flexibility would buy an order-of-magnitude speed advantage. The company is now at pains to correct the impression that its silicon is built for any single model, saying it supports any frontier model. Broadening that claim widens the addressable market, but it also moves Etched toward the general-purpose ground Nvidia already occupies — and specificity was the whole argument for why a startup could beat Nvidia at all.

Kleiner Perkins, Sequoia Capital, Andreessen Horowitz, Peter Thiel, Tiger Global, Bain Capital Ventures, Blackstone, Neo, Stripes, Primary, Positive Sum, Diffusion and Argo also participated. Worth keeping in proportion: the datacenter test that anchors this valuation was run by a trading firm, and a quant shop's inference profile — small models, brutal latency budgets, predictable batch shapes — is not the workload a frontier lab brings. Etched has now been validated on hardware by a real customer, which is rare. It has not yet been validated on the workload that would justify $21 billion.