Etched, an AI chip startup founded in 2022 by three Harvard dropouts, has closed a $300 million Series C financing round at a $10.3 billion valuation, co-founder and COO Robert Wachen told TechCrunch. The round was led by Sequoia, with participation from Andreessen Horowitz, SK Hynix, Jane Street, Diffusion Capital and earlier investors. Individual backers include Peter Thiel, Andrej Karpathy, Dylan Field and Amjad Masad.
Rapid valuation growth and orders
Etched was valued at $5 billion in December when it raised a $500 million round, meaning its valuation has roughly doubled in about seven months. The company says this is the highest valuation for a Sequoia-led Series C. Etched also reports that its homegrown chips were successfully manufactured by TSMC, its first full systems are being tested by customers, and it has already booked $1 billion in orders.
What sets their technology apart
Etched began with the idea of building hardware optimized for modern transformer-based AI models, but the company stresses its systems can run any AI model. According to Wachen, Etched’s systems can handle Mixture of Experts models like DeepSeek and Qwen, as well as non-transformer architectures such as Mamba, which is built on a state-space model.
The company highlights two newly designed components that accelerate inference, the computation performed after a user submits a prompt. Wachen outlines inference as two stages: prefill (processing the prompt and context, which is compute intensive) and decode (generating output tokens, which requires large amounts of memory).
- Prefill: Etched developed a prefill chip that operates at much lower voltage than other AI chips, a design the company calls “low-voltage inference.” Lower voltage reduces heat and allows denser transistor packing, which Etched says yields dramatic speed improvements.
- Decode: For decoding, Etched built a new type of memory and an interconnect they call “cluster scale memory,” enabling many chips to share a common memory pool with very low latency. The combination is claimed to produce high throughput at lower cost.
The idea of etching model-specific elements into silicon to boost performance is becoming less speculative; the report notes Google is reportedly pursuing a similar approach with its Frozen v2 chip for Gemini.
Reception, demos and skepticism
Because Etched started before many in the tech world fully recognized AI’s specialized compute needs, the founders — CEO Gavin Uberti, Robert Wachen and CTO Chris Zhu — faced skeptics. Much skepticism stemmed from the fact few people had access to the systems: demos were largely private and shown to investors and early customers.
Wachen said notable figures who tried the hardware include Andrej Karpathy from Anthropic, Noam Brown from OpenAI and Geoffrey Hinton, and that investors in the financing round were enthusiastic after private demonstrations. Still, moving from manufactured silicon and early customer testing to mass-produced rack systems and broad deployment remains a significant challenge.
Founders’ early struggles and current scale
The trio famously dropped out of Harvard to launch Etched. Wachen recalled not knowing how to raise large sums of capital or how to hire in the beginning. After arriving in the Bay Area with no office or apartment, he slept on a friend’s unfurnished house floor. Early servers for chip design tools ran in a garage of an early employee, and reboots sometimes required the employee’s spouse to press the reset button.
Today Etched employs roughly 400 people in an office and runs a 2-megawatt data center. Wachen says they are running tokens in their lab and collaborating with some of the world’s largest AI companies.
Despite the difficulties, the founders persisted. “It’s come a long way. It’s a very, very different world. But I think, when you really think something’s possible, and you just work at it for a long time, you can do it,” Wachen said.
Why this matters
Etched’s rapid valuation increase and technical claims underline the importance of hardware innovation for scaling AI workloads. If the company’s low-voltage prefill chip and cluster-scale memory deliver the promised performance and cost benefits at production scale, they could be attractive to large model operators facing growing inference costs. The near-term milestones to watch are mass production of rack systems, broader customer deployments, and independent validation of performance and cost savings.
