Etched raised $700 million at a $21 billion valuation on Tuesday, twenty-six days after Sequoia led its $300 million Series C at $10.3 billion. The lead check this time came from Jane Street, which is also Etched’s first paying customer: the quant firm took delivery of the startup’s inaugural inference rack, liked what it saw in its data center, and promptly wrote the round that doubled the company’s price.

The optics are almost too clean. Jane Street sits on Etched’s cap table alongside Hudson River Trading, Jump Trading, and Two Sigma. Four prop shops, each of them structurally long on latency-sensitive compute and structurally short on Nvidia allocation, have organized themselves into something between a customer consortium and a syndicate. The valuation has roughly quadrupled in eight months from a $5 billion mark in December.

The technical pitch is Low Voltage Inference, which promises higher compute density inside the same power envelope, and Cluster Scale Memory, a hybrid subsystem that pools memory across an entire cluster rather than a single chip. Etched says the resulting systems run Mixture of Experts designs including DeepSeek and Qwen, and non-transformer architectures such as Mamba. It claims more than $1 billion in customer contracts across frontier AI labs and clouds.

The execution story is where the WSJ reporting bites. Etched went from receiving test silicon back from TSMC to running inference workloads in 44 days, against a six-month industry norm. The person responsible is Brian Loiler, recruited from Nvidia in 2024 after 22 years building the HGX and DGX server systems. Loiler has since pulled roughly a dozen more engineers out of Nvidia; alumni now make up around 15% of Etched’s 400-person workforce.

Michael Burry, in the same WSJ piece, called Etched serious competition for Nvidia. Coming from the investor whose brand is built on shorting consensus, that quote is less a prediction than a tell about how the trade is currently positioned.

Sources