Etched, founded by Gavin Uberti and Rob Wachen, sells a full inference rack including power delivery, boards and interconnects, not just a chip.
Rob Wachen argues chip design tools default to freezing temperatures while AI data centers never run below about eight degrees Celsius, exposing general-purpose assumptions in the industry.
Etched bets on low-voltage inference and cluster-scale memory, citing GPU model FLOPS utilization of roughly 20% to 50% and the quadratic link between voltage and power.
Etched bets purpose-built inference racks can outdo general-purpose GPUs
Etched is betting that AI inference will grow enormously and that chips built for the general-purpose past will serve it poorly. By designing a full rack around low voltage and cluster-scale memory, it aims to make tokens cheap and fast. The founders argue that speed, vertical integration and a willingness to defy the base rate are how a young team can compete with the largest chip companies.
Takeaways
Etched sells full inference racks, not chips. The product is a rack with the chip, power delivery, board and chip-to-chip interconnect, built at mass volume. Rob Wachen says the best vendor is no vendor. Etched hired Brian Loyler, who built Nvidia's HDX and DGX systems.
General-purpose assumptions limit today's AI chips. Both founders say chips designed before ChatGPT suit modern workloads poorly. Design tools default to freezing temperatures, though AI data centers never run below about eight degrees Celsius. GPU model FLOPS utilization runs roughly 20% to 50%.
Low voltage and cluster-scale memory are the bets. Voltage drives power quadratically, and Bitcoin miners run at under a quarter of GPU voltage despite experts calling it impossible. For decode, Gavin Uberti says cluster memory bandwidth matters, and a custom interconnect can cut Blackwell's roughly 4,000 nanosecond latency over fivefold.
Vertical integration and parallel work create speed. Etched built software, racks and production before chips returned, using 700 FPGAs to run a full-reticle design. That compressed bring-up to forty days. A delayed vendor once forced a dozen engineers to Bangalore for six months.
Fundraising was hard and most major investors passed. The founders wrote a 30-page memo that took about 100 hours. They found a debt provider, then raised a Series A of roughly $103m. Patrick O'Shaughnessy, a biased five-time investor, says investors must damn the base rate.
TSMC supply works for customers deploying at scale. Gavin praises TSMC's customer service and recalls a yield experiment it adopted across its line. Rob argues supply is not zero-sum for scale customers. Etched's first generation uses four-nanometer technology, while Nvidia's Rubin uses three-nanometer.
Cheap, fast inference needs far more infrastructure. Only a few million people pay for AI plans, so serving billions requires much more capacity. Rob wants tokens made with iPhone-like or car-like economies of scale, and says fast decode can compress year-long compute builds into a month.
Predictions for where AI goes
Rob Wachen predicts more agents than humans doing knowledge work by 2027.
Tap a name to go there · tap a line to see why they are linked
On this episode 3
Show 1
Rob expects inference to become a majority of global GDP.
Rob says dynamism will be a theme, and building an entire browser with coding agents will soon take under an hour.
Gavin Uberti expects math to get cheaper faster than memory. Mixture-of-experts models already use only a small fraction of parameters per token.
Rob expects the largest companies to produce most of the global token supply and own most of its supply chain.