Etched, the AI inference chip startup founded in 2022 by three Harvard dropouts, closed a $300 million Series C at a $10.3 billion valuation, led by Sequoia. Andreessen Horowitz, SK Hynix, Jane Street and Diffusion Capital also participated in the round. Jane Street is also a customer, buying Etched's full server rack systems—making its investor position a direct bet on its own infrastructure supplier.
The raise follows a $500 million round in December 2025, when the company was valued at $5 billion. Etched has doubled its valuation in about seven months and gone from zero public revenue to more than $1 billion in signed customer contracts.
Total capital raised now stands at roughly $800 million across the two most recent rounds alone. Earlier backers include Peter Thiel, former OpenAI research director Andrej Karpathy, Figma CEO Dylan Field and Replit CEO Amjad Masad.
Etched's core product is a chip built from the ground up to accelerate inference—the computational step that fires every time a user submits a prompt to a large language model. The company sells full rack-scale systems rather than discrete chips, a go-to-market decision that locks in tighter relationships with buyers and captures more of the hardware stack's economics.
Co-founder and COO Robert Wachen has said inference involves two stages: a prefill phase that processes the incoming prompt—computationally intensive and mathematically demanding—and a decode phase that generates the output tokens the user sees. Etched designed a separate chip for each stage. The prefill chip runs at lower voltage than competing AI chips, generating less heat and allowing more transistors to be packed into the same die. The decode chip introduces what the company calls cluster-scale memory, a shared memory pool that lets many chips connect together at high speed.
The company disputes the characterization that its chips run only transformer-based models. Wachen has said the systems support Mixture of Experts architectures—a design that routes tasks across specialized sub-models, used by DeepSeek and Qwen—as well as non-transformer designs like Mamba, which uses a state-space model architecture instead of the attention mechanism at the heart of GPT-style systems.
Etched came out of stealth last month, announcing that its first chips had been manufactured and that its initial full systems were in testing with clients. The $1 billion in signed contracts at that point gave the company a revenue backlog before it had shipped product at scale—an unusual position for a hardware startup at this stage.
The Series C carried a notable distinction: the company said it was the highest valuation ever for a Sequoia-led Series C.
Jane Street's dual role as customer and investor is a sharp data point in this deal. Trading firms run some of the most compute-intensive workloads outside hyperscalers—quantitative models, real-time risk, AI-assisted execution. A firm of Jane Street's scale buying rack systems and then participating in Etched's largest funding round is a concrete endorsement of the product's performance claims, not a financial speculation on the AI sector broadly.
The competitive context matters. Nvidia dominates AI chip revenue and trades at $225.01, essentially flat on the day. But Nvidia's strength is in training and broad deployment; Etched's specific claim is inference efficiency at a cost and thermal profile that Nvidia's general-purpose H100 and Blackwell architectures cannot match. Google is reportedly pursuing a similar concept with its Frozen v2 chip, designed to bake parts of its Gemini model directly into silicon—a sign that inference-optimized custom silicon, once considered fringe, now has traction at the largest AI spender on the planet.
At $10.3 billion, Etched is valued at roughly 10 times its current contract backlog. That multiple is aggressive for a hardware company with no shipping revenue at scale, but it reflects what investors are paying for: a position in the inference layer of AI infrastructure before that layer becomes as contested and expensive to enter as the training layer already is. Etched's next test is straightforward—deliver the systems, prove the performance numbers and convert that backlog into recognized revenue.
