AI inference chip startup Etched has raised $700 million at a $21 billion valuation in a round led by Jane Street, just 26 days after being valued at $10.3 billion. Jane Street was already a backer, and the round coincides with Etched shipping its first customer rack to Jane Street itself. The valuation has roughly quadrupled in eight months despite the company having just delivered its first paying customer order. Etched's chips split workloads into a low-voltage prefill chip and decode via shared cluster-scale memory, and now support Mixture of Experts models like DeepSeek and Qwen plus non-transformer architectures like Mamba. Over $1 billion in contracts was signed as of June, and its first chip worked on TSMC's N4P process on the first attempt. The funding will go toward factories, supply chains, and fleet software as it scales toward gigawatt-level deployments, amid competition from European rivals like Olix.
Questions this post answers
What is Etched's inference chip architecture and how does it handle different AI model types?
Etched splits its chip design into a prefill chip that runs at low voltage to pack in more transistors without excessive heat, and a decode stage that relies on cluster-scale memory allowing many chips to share one fast memory pool. Originally built around etching a single model into silicon, the systems now support Mixture of Experts models like DeepSeek and Qwen, plus non-transformer architectures such as Mamba. daily.dev surfaces coverage like this for engineers evaluating specialized inference hardware against GPUs.
How much funding has Etched raised and at what valuation?
Etched raised $700 million at a $21 billion valuation in a round led by Jane Street, up from a $10.3 billion valuation just 26 days earlier following a $300 million Series C led by Sequoia, and $5 billion in December. The valuation has roughly quadrupled in about eight months, alongside the company's first customer delivery, a rack shipped to Jane Street. Track fast-moving AI infrastructure funding rounds like this one on daily.dev.