Etched Raises $300M at a $10.3 Billion Valuation
Etched, the AI chip startup founded by three Harvard dropouts, said Thursday it has closed a $300 million Series C at a $10.3 billion valuation, a round the company calls the highest valuation ever for a Series C led by Sequoia. The raise doubles the four-year-old company's worth in roughly seven months and lands less than a month after it said booked orders for its AI inference systems had crossed $1 billion.
Sequoia led the round, with participation from Andreessen Horowitz, SK Hynix, Jane Street and Diffusion, alongside earlier backers. Etched was valued at $5 billion in December, when it raised a $500 million round, and its total funding now tops $1 billion. Individual investors in the company include Peter Thiel, Andrej Karpathy, Figma CEO Dylan Field and Replit CEO Amjad Masad.
The San Jose company plans to use the new capital to expand production and customer deployments. It recently opened an 80,000-square-foot facility near its headquarters for manufacturing and prototyping, employs about 400 people and, per Reuters, says demand for its systems "continues to outpace supply" as customers move from evaluation to deployment.
A bet that looked crazy in 2022
Etched was founded in 2022 by CEO Gavin Uberti, COO Robert Wachen and CTO Chris Zhu, who dropped out of Harvard to build silicon specialized for transformer models, the architecture behind systems like ChatGPT and Claude. At the time, betting an entire chip company on one model architecture struck much of the industry as reckless. Nvidia's general-purpose GPUs were the default answer to every AI workload, and custom silicon efforts had a long history of ending badly.
Four years later, the wager looks considerably less strange. Last month Etched announced that its first homegrown chips had been successfully manufactured by TSMC, that complete systems were in testing with clients, and that it had booked $1 billion in orders. Even Google is reportedly exploring the same core idea, working on silicon that bakes parts of its Gemini architecture directly into the chip.
"We had no idea how hard it was going to be," Wachen told TechCrunch, reflecting on the road from dorm room to a 400-person company running a 2 megawatt data center.
What the hardware actually does
Etched sells full rack-scale systems rather than bare chips, and the company has spent much of the past year pushing back on the perception that its hardware only runs specific large language models. According to Wachen, the systems can run any modern AI model, including Mixture of Experts designs like DeepSeek and Qwen, which split work across specialized sub-models, as well as non-transformer architectures like Mamba.
The company's pitch centers on inference, the computing that happens after a user submits a prompt. Inference has two phases: prefill, where the system digests the prompt and its context, and decode, where it generates the answer token by token. Prefill is compute-hungry; decode is memory-hungry.
Etched built new components for each. Its prefill chip runs at a much lower voltage than conventional AI accelerators, an approach the company calls low-voltage inference. Lower voltage means less heat, which allows more transistors to be packed into the same silicon. For decode, Etched developed an interconnect and memory design it calls cluster scale memory, which lets many chips share a common memory pool at very low latency. The combined promise is faster output at lower cost than GPU-based clusters.
Access to the systems has been tightly limited to investors and early customers, which fed skepticism but also became the company's sales strategy. Wachen says the round came together through private office demos, and that researchers including Karpathy, OpenAI's Noam Brown and Geoffrey Hinton have tried the hardware in person.
Why this matters for founders and operators
The inference market is where AI's economics get decided from here. Training a frontier model happens a handful of times a year at a handful of companies, but inference happens billions of times a day at every company that ships an AI feature. As usage scales, the cost of serving tokens, not training models, becomes the line item that determines whether AI products have real margins.
That is why capital keeps pouring into companies attacking the inference bottleneck from different angles, from Etched's specialized silicon to inference cloud platforms and custom accelerators from the hyperscalers themselves. For startups building on top of AI models, more competition at the silicon layer points in one direction: cheaper tokens over time, which expands what is economically viable to build.
The round is also a case study in conviction. Etched spent years as an industry punchline, three college dropouts building a chip for an architecture that might not last. The founders kept shipping through the doubt, landed TSMC production, converted demos into a billion dollars of orders, and are now backed at $10.3 billion by the most established venture firm in the business. Skepticism is not a verdict; it is a stage.
There is a caution flag too. Etched still has to mass-produce and deliver rack systems at scale, a challenge that has broken plenty of well-funded hardware companies. Booked orders are not shipped revenue, and Nvidia has crushed challengers before by simply iterating faster. Wachen himself acknowledges the company must stay "humbled by what it will take to actually get to scale."
What does Etched build?
Etched designs AI inference hardware sold as complete rack-scale systems. Its components include a prefill chip that runs at unusually low voltage and a shared-memory interconnect called cluster scale memory. The systems are designed to run AI models faster and more cheaply than general-purpose GPUs.
Who invested in Etched's Series C?
Sequoia led the $300 million round, with participation from Andreessen Horowitz, SK Hynix, Jane Street and Diffusion, plus earlier investors. The company says it is the highest valuation ever for a Sequoia-led Series C.
How is Etched's hardware different from Nvidia GPUs?
GPUs are general-purpose processors adapted for AI work. Etched builds silicon specialized for inference, splitting the workload into a compute-optimized prefill stage and a memory-optimized decode stage with dedicated hardware for each. The company says its systems can run any modern model architecture, not just transformers.
Why is AI inference such a big market?
Inference is the computing that runs every time someone uses an AI product, so its costs scale with usage rather than with model development. As AI features reach billions of users, serving costs become the dominant expense, making faster and cheaper inference one of the most valuable problems in the industry.
What will Etched do with the funding?
The company says it will expand production and customer deployments. It recently opened an 80,000-square-foot production and prototyping facility near its San Jose headquarters and plans to keep growing its roughly 400-person team as systems move from testing into delivery.