Etched Closes $300 Million Series C at $10.3 Billion to Challenge Nvidia on Inference

Etched Closes $300 Million Series C at $10.3 Billion to Challenge Nvidia on Inference

Four years ago, three Harvard undergraduates dropped out of school on a wager that most of Silicon Valley thought was reckless: that the entire AI industry had quietly standardized on a single design, the transformer, and that a chip built to run only that design could beat Nvidia at its own game. On July 23, that wager collected a $10.3 billion valuation.

Etched, the San Jose-based startup founded in 2022 by CEO Gavin Uberti, COO Robert Wachen, and CTO Chris Zhu, said it has closed a $300 million Series C led by Sequoia Capital, with participation from Andreessen Horowitz, SK hynix, Jane Street, and Diffusion Capital, alongside earlier backers. The round values the company at $10.3 billion, roughly double the $5 billion valuation it commanded in December when it raised $500 million. That is a doubling in about seven months, and Etched says it is the highest valuation ever for a Sequoia-led Series C. The company’s angel roster reads like an AI hall of fame, including Peter Thiel, Andrej Karpathy, Figma’s Dylan Field, and Replit’s Amjad Masad.

The bet: a chip that does one thing

Etched’s flagship product is Sohu, an application-specific integrated circuit, or ASIC, fabricated on TSMC’s N4P process and designed to run transformer models and little else. The logic is deliberately narrow: rather than build a general-purpose processor that does many jobs adequately, Etched hard-wired the transformer, the architecture behind ChatGPT, Claude, and nearly every frontier model, directly into silicon. The company claims Sohu delivers more than 10x the performance of leading GPUs such as Nvidia’s H100, with some reports suggesting a single Sohu server can stand in for roughly 160 H100 units.

Wachen is quick to push back on the perception that a transformer-only chip is a straitjacket. The systems, he says, can run any model, including Mixture-of-Experts designs such as DeepSeek and Qwen and even non-transformer architectures like Mamba. The company’s real edge, he argues, lies in re-engineering inference, the computation that happens after a user hits enter. “Inference is built in two stages,” Wachen told TechCrunch, “prefill and decode.” For the compute-heavy prefill phase, Etched built a chip that runs “at a much lower voltage than any other AI chip. We call this low-voltage inference.” Lower voltage means less heat, which means more transistors packed onto the die. For the memory-hungry decode phase, the company developed what it calls cluster-scale memory, an interconnect that lets many chips share a single low-latency memory pool.

Orders on the books

The funding lands a month after Etched said it had successfully manufactured its first silicon at TSMC, begun testing full systems with customers, and booked $1 billion in orders. The company now employs around 400 people and operates a 2-megawatt data center where, Wachen says, it is “running tokens in our lab today, working with some of the largest AI companies in the world.” Access has so far been limited to investors and early customers, many of whom became believers only after private demos. “Andrej Karpathy from Anthropic, Noam Brown from OpenAI, Geoffrey Hinton, as well as all the investors in the funding round, these are all people who actually tried the hardware and are very excited about it,” Wachen said.

Why it matters

Etched is a pure-play wager on the most important structural shift in AI economics: the migration of demand from training to inference. Training a model happens once; inference happens every time anyone uses it, forever. As AI moves from research labs into consumer apps, enterprise software, and agents that run continuously, the cost of serving tokens, not training runs, is becoming the binding constraint on the industry. Andreessen Horowitz, in backing the round, framed inference as “the most important workload of our era.”

That is also the pressure point on Nvidia, whose GPUs dominate both training and inference but were never purpose-built for the latter. A credible, cheaper, faster inference alternative strikes at the margin story underpinning Nvidia’s valuation. Notably, Etched is no longer alone in the thesis that etching a model into silicon is viable rather than “wacky”: Google is reportedly developing a Frozen v2 chip with Gemini’s architecture baked into the hardware.

The risk is the mirror image of the opportunity. An ASIC optimized for transformers is a concentrated bet that transformers stay dominant. If the field lurches toward a fundamentally new architecture, a chip built for the old one loses much of its value overnight. Etched is wagering that will not happen soon, and $1 billion in signed orders suggests a meaningful number of customers agree.

What to watch

The decisive test is delivery. Etched has promised its first rack-scale systems will ship in the summer of 2026, and the gap between a demo in a 2-megawatt lab and mass-produced racks in customer data centers is exactly where hardware startups tend to stumble. Watch whether Etched converts its $1 billion order book into deployed, revenue-generating systems on schedule, and whether independent benchmarks validate the 10x-plus claims outside the company’s own lab. Watch, too, for the next raise: reports already peg a future round near $20 billion. Wachen, for his part, is not declaring victory. “We had no idea how hard it was going to be,” he said. “I think we still have to be humbled by what it will take to actually get to scale.”