Partnering with Etched: Building the Inference Machine
Sequoia By Sonya Huang and Abhishek Malani ● Covered by 3 sources
Etched is shipping its first AI inference chip system and Sequoia is leading its $300M Series C. The bet: inference is where AI gets huge, and hardware built for it could win big.
Based on reporting by Sequoia, By Sonya Huang and Abhishek Malani — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Sequoia is putting serious money behind Etched’s idea that inference, not training, will be the giant AI market. The firm says the startup’s first product is a production-ready custom silicon system, set to ship in 2026, and that the company has already shown it can run frontier models at the cluster level.
That was a contrarian call when Gavin, Chris and Rob were still in Harvard dorm rooms in 2022. Now it looks much less lonely. Etched has spent the last few years on the hard, unglamorous parts of AI hardware: low-voltage inference, cluster-scale memory, and systems designed around racks and clusters instead of a single chip. The goal is simple to say and brutal to execute: more throughput, less latency, better interactivity.
The company’s pitch is not tied to one model family. Sequoia says Etched’s system works across large sparse MoEs, dense transformers and even alternative architectures like Mamba. That matters because the model mix keeps shifting. Context lengths grow. Attention gets reinvented. Hardware teams have to keep moving while the target keeps changing shape.
Etched has also been doing the operational stuff that separates real hardware companies from slide decks. The team stood up a live lab in San Jose for its first racks, opened an office in Taiwan to work closer to suppliers and speed up testing, and keeps repeating a blunt internal slogan: production is the product.
Earlier this year, Etched taped out its first-generation chip at TSMC, which Sequoia says made it the first post-ChatGPT-era company to complete a full-reticle A0 tape-out on TSMC’s leading-edge nodes. In 40 days, the team brought up its first chip cluster and ran inference on a wide range of frontier AI models. Sequoia says the earliest customers are now getting access. The round itself is a $300 million Series C at a $10 billion pre-money valuation, with Jane Street, Andreessen Horowitz, Diffusion and SK Hynix joining in.
My take — AI-written commentary, not fact-checked reporting
This is the sort of bet AI investors love: giant market, hard engineering, and just enough physics to make everyone feel brave. The real tell is the phrase “production is the product” — that’s what hardware people say when they’ve been burned and learned to stop romanticizing the demo. Also, if inference really is the money machine, then the gold rush is moving from model fireworks to very expensive plumbing.
Read more about this at: Sequoia