Etched hits a $10.3B valuation betting the AI behind ChatGPT deserves its own silicon

Sequoia, SK Hynix and Jane Street just priced a $300 million Series C around a single bet: that the compute step that turns a prompt into an answer is becoming its own infrastructure category, with chips built for one job: running transformers, the AI design behind ChatGPT and most modern AI systems.

Etched, an AI chip startup founded in 2022 by three Harvard dropouts, is now valued at $10.3 billion on the round, according to TechCrunch. That is roughly double the $5 billion the company commanded in December on a $500 million raise, and the company says it is the highest valuation ever for a Sequoia-led Series C. The price tag is the round. The product thesis is the news.

Etched sells full systems designed to run any AI model, including mixture-of-experts architectures like DeepSeek and Qwen, as well as non-transformer designs such as Mamba and state-space models. The specialization is in the silicon: two new components built from scratch for the two stages of inference, the compute step that produces a model's answer after a user submits a prompt. Prefill, the compute-heavy phase that ingests the prompt and context, sits on one block. Decode, the memory-heavy phase that generates the visible answer one token at a time, sits on the other. That split, etched directly into the chip rather than split across software and a general-purpose GPU, is the technical bet.

Sequoia led the round. Andreessen Horowitz, SK Hynix, Jane Street, and Diffusion Capital joined. Individual backers include Peter Thiel, Karpathy, Field, and Masad. SK Hynix is the memory chipmaker that supplies the high-bandwidth memory used in Nvidia's GPUs; their participation is a vote of confidence from the company that would otherwise supply Etched's competitors. Jane Street is a quantitative trading firm that runs some of its own inference workloads in-house; their check signals a buyer willing to underwrite the silicon instead of just waiting for Nvidia.

Etched is also sitting on roughly $1 billion in orders, the company told TechCrunch, and last month announced that it had successfully manufactured its homegrown chips. First full systems are being tested by clients. The $1 billion figure and the "dramatically faster" prefill claim both come from a single on-the-record company source, co-founder Wachen.

Etched's chips are designed for transformer-shaped AI. If the field moves toward non-transformer architectures (state-space models, Mamba, or whatever comes next), the company's bet ages in real time. Google is reportedly pursuing the same idea of etching parts of a specific AI model into silicon, with a chip referred to as Frozen v2 for Gemini. The parallel bet from the largest AI lab in the world is both validation of the inference-silicon thesis and a signal that the bigger players see the same workload specialization.

The round is one Series C price for a specific hardware layer: chips optimized for the workload that actually answers a user's prompt, with the prefill/decode split baked in. Whether that layer becomes a durable category or a transitional bet depends on whether transformer-shaped silicon ages well as the models change. The next $10 billion of inference will tell.