Etched has a simple pitch. Nvidia builds chips that do everything. Etched builds chips that do one thing, inference, and claims to do it far better. On Thursday the bet got a lot more money behind it. The startup raised $300 million in a Series C round that values it at $10.3 billion.
Sequoia led the round. Andreessen Horowitz, Jane Street, Diffusion and, tellingly, the memory maker SK Hynix all took part. It is, the company says, the highest valuation ever for a Sequoia-led Series C. Etched has now raised more than $1 billion in total.
The number that stands out is the speed. Etched was worth $5 billion in December. Seven months later it is worth twice that. The startup only emerged from stealth in June, when it revealed a working chip and $1 billion in signed orders.
A chip for one job
Etched started with a contrarian idea. Back in 2022, building a chip tuned for AI models based on the transformer architecture looked, in the words of TechCrunch’s Julie Bort, wild if not wacky. The three founders dropped out of Harvard to try anyway. They are chief executive Gavin Uberti, president Rob Wachen, and chief technology officer Chris Zhu.
The perception has been hard to shake. Critics assumed Etched’s systems could run only a handful of specific models. Wachen says that is wrong. The clusters are architecture-agnostic. They run large mixture-of-experts models such as DeepSeek and Qwen. They also run non-transformer designs like Mamba, a wholly different approach.
The 💜 of EU tech
The latest rumblings from the EU tech scene, a story from our wise ol’ founder Boris, and some questionable AI art. It’s free, every week, in your inbox. Sign up now!
The wider industry has come round to the core idea. Google is reportedly hardwiring Gemini’s architecture into its own Frozen v2 chip. Etched no longer looks so wacky.
Power and memory
Etched sells full rack-scale systems, not loose chips. Its edge, it claims, rests on two inventions. Each targets a different bottleneck in inference.
The first is power. AI chips overheat as they push more calculations, so they throttle their own speed. Etched runs its math blocks at under half the usual voltage. Less voltage means less heat. That lets the chip pack in more transistors and run harder. It calls this low-voltage inference.
The second is memory. Generating an answer demands huge, fast memory access. Etched built a shared memory pool that spans a whole cluster of chips, stitched together by a proprietary high-speed interconnect. That its new backers include SK Hynix, the world’s leading maker of AI memory, is no coincidence.
Time to ship
The hard part starts now. Etched has to turn prototypes into mass-produced racks. It has opened an 80,000 square-foot facility near its San Jose base, plus a factory in Taiwan. Its team of 400 includes engineers poached from Nvidia, Broadcom, Google’s TPU group and SK Hynix.
Demand, Reuters reported, is outrunning supply as customers move from testing to deployment. Uberti is blunt about what comes next. “Now is the time to be aggressive,” he said. “Our chips work, people want them, and it’s time to ship.”
Whether they ship at the scale Nvidia does is the open question. Etched has raised the money and the expectations. The race to make inference cheap now has one more well-funded contender, with a memory giant betting it can win.
