Etched Raises $800M To Build Full AI Inference Racks And Chips
Reactions from ranked influencers
2 posts
WHOEVER MAKES THE MOST TOKENS, WINS @UbertiGavin and Rob Wachen, co-founders of @Etched, interviewed by @Patrick_oshag (@InvestLikeBest) Summary: @Etched bet that inference becomes the largest market on earth, and the company that produces the most tokens becomes the most valuable company in the world. Two Harvard dropouts raised about $800M, against more than $1B in customer demand, to build the entire inference rack: chip, board, power delivery, interconnect, and the factory to mass-produce it. The operating lessons for founders: vertically integrate, parallelize the schedule so nothing waits, and treat every impossible constraint as solvable. 1. Tokens as the prize. Inference will be the biggest market in the world, and whoever produces the most tokens becomes the most valuable company in the world. Etched runs every decision through one question: how do we get the most token capacity online? Today only about 1 in 1,000 people uses a paid AI plan, so the demand curve has barely started. 2. Production is the product. Etched builds the full inference solution: the chip, the board, the power delivery, the interconnect, and the factory that mass-produces the racks. As Uberti puts it, "really the production is the product," and "the best ability is availability." Etched is the only startup building its own rack and its own chip at once. 3. Everything is built on buffer. The semiconductor stack is designed general-purpose, which leaves slack once you know your exact use case. The default chip-design tools assume your chip might run in freezing temperatures, and as Uberti says, "I've never seen an AI data center with ice in it." Get 20% here and 2x there, and it compounds into a system far better for inference. 4. Low voltage inference. Etched's first technical bet runs the chip at under half the voltage of any other AI chip. Power rises with the square of voltage, so halving voltage cuts power to a quarter, and Bitcoin miners already run at under a quarter of GPU voltage, so the physics is settled. Lower voltage clears the thermal ceiling that makes GPUs throttle before you can add more compute. 5. Cluster scale memory. The second bet treats the whole cluster's memory as one shared pool instead of one chip's. A Blackwell chip takes about 4,000 nanoseconds to reach the next chip, and the speed-of-light floor is 2 to 3 nanoseconds. By custom-building everything above the second layer of Ethernet, Etched cut that latency by more than 5x. 6. Assume it is possible. The operating rule is to treat every hard problem as solvable, then work backward to the how. When a clock bug threatened to make every chip return wrong answers, some engineers quit and called it unsolvable; the team aligned two clock signals to within 50 trillionths of a second, 2 billion times a second, and fixed it in two weeks. As Uberti says, dozens of experiments fail and "I only got to get lucky once." 7. Prefetch the schedule. More than $1B of category revenue moves every day, so every day you don't ship costs you. Before their chip came back, Etched shipped empty racks to customer data centers, ran the full model on more than 700 FPGAs, and built cold plates against a mock thermal chip. One well-known AI chip company took 10 months to go from silicon to running inference; Etched did it in 40 days. 8. Flood the binding constraint. When a key vendor fell a year behind, Etched found a third option: send a dozen top engineers to Bangalore for 6 months to sit inside the vendor's team. Uberti lived there four and a half months, running 24-hour handoffs with the US team and deciding on the spot instead of waiting 12 hours for email. "Speed wins." 9. Legends plus naivety. Etched pairs the single best person in the world at a problem with raw, first-principles young talent, and the value is in the pairing. They hired Brian, who built NVIDIA's HGX and DGX systems and could point at a decision and say "that's a billion-dollar lesson I learned." Their phrase: "chips on shoulders put chips in data centers." 10. Damn the base rate. Every major Valley investor passed at once: two kids, no taped-out chip, inference might be a bubble. The biggest semiconductor Series A at the time was $40-50M, and Etched needed $100M, which they closed as soft commits scraped together in survival mode. As their investor Patrick O'Shaughnessy puts it, if you invest on base rates, buy an index fund. 11. It has to be existential. The best chip in the world gets built by the company that builds only that chip, which is why it is NVIDIA. An architect leaving a frontier lab said it plainly: Google survives if its TPUs fail, Meta survives if MTIA fails, OpenAI survives if its chip fails. For Etched, getting token capacity online is the entire company. 12. Agents per megawatt. Uberti's forecast: society measures productivity as GDP per capita now, and it will come to measure agents per megawatt. He calls 2026 the second-to-last year that most of the workforce is human, with more agents than people doing knowledge work by 2027. He calls it "the biggest proliferation of technology humanity's ever seen."