Rob Wachen was 16 when doctors told him that the bump on his back was stage 4 bone cancer, and that his odds of surviving were under 30%. Surgery meant he would probably live but might never walk again; radiation meant he would keep walking but might not survive. His parents told him the decision was his alone, and he chose surgery. The necrosis analysis came back below the threshold, which meant he needed radiation anyway. One of the only machines in the world that could deliver it was in Boston, so his parents moved there with him while he went through months of treatment in a wheelchair. He spent two years in chemotherapy, surgery, and learning to walk again. Before the diagnosis, he had been running a six-figure business, competing in Taekwondo, and playing semi-professional saxophone. He has said that surviving something like that requires hoping for something, and that he always knew if he got through it, he wanted to do something that mattered. Gavin Uberti's story is stranger in a different way. At 12, he was leading teams in Minecraft to build analog chips, sending precise architecture specs for hundreds of thousands of blocks to group chats. At 17, he got his first job writing high-performance code, and because a 17-year-old cannot sign a legally binding contract, the company skipped the NDA, sat him down, and said, "Gavin, don't share this information." That company sold to Apple for $200 million. The next one he worked at sold to Nvidia. In high school robotics, Gavin and his friend Sanford left their 20-person school team to compete as a two-person unit, on the theory that everyone else split their attention between documentation, outreach, and the robot, and they would only build the robot. At one point they held the world record for the highest score in the competition. The two met at Harvard, where Rob ran Prod, a nonprofit accelerator whose alumni companies, Cursor and Mercor among them, are now worth more than $100 billion combined. When GPT-4 launched with image uploads, Rob dug up a photo of his back from before his diagnosis and asked the model what the bump could be. It answered immediately that it could be a tumor and that he should get an MRI. His doctors had taken six months to reach the same conclusion. While he sat there absorbing that, a notification appeared on his screen telling him he was out of image credits for the day. The models were about to change everything, and the infrastructure to serve them did not exist. Every GPU and TPU running these models had been designed before ChatGPT. Gavin, who had spent a year optimizing AI models to run on GPUs, believed a system built specifically for inference could be radically better. In 2023, they dropped out to build it. They were 21. GPUs can't use all their advertised computing power because they overheat and slow themselves down. Etched developed what they call low voltage inference, running chips at under half the voltage of any other AI chip, borrowing an insight from Bitcoin miners that everyone in the industry insisted couldn't apply to AI. And where Nvidia's chips take 4,000 nanoseconds to send data to each other, Etched built a custom interconnect that cuts that by more than 5x, letting an entire cluster share its memory as a single pool. Mark Ross was the former CTO of Cypress Semiconductor, which sold for $9 billion. When they met him, they were a couple of guys in a dorm room telling a prestigious chip veteran they could be much faster than Nvidia, and he told them what everyone else had: "no, you can't, it will not work." But if they wanted to convince him, they should write a white paper, build a functional simulation, and show him. After a lot of very long nights, they came back with the simulation, and he looked at it and said, "huh, this works." Mark is Etched's full-time CTO. At the end of 2023, they had $15 million in the bank and needed at least $100 million more to finish the chip and build the rack. They wrote a 30-page memo and asked for it, double the largest semiconductor Series A ever raised, and every major investor in the Valley passed immediately. The market was uncertain, they were told, and two 22-year-olds could not build this type of company. Gavin started looking into how hard it would be to go back to Harvard. They called everyone they knew and survived on small checks, a million here, two million there, enough not to run out of money that month, then a five million check, then ten. At a board meeting, they looked at a spreadsheet of soft commits totaling $103 million. When a critical vendor in Bangalore fell a year behind, they shipped a dozen engineers to India for six months, and Gavin lived there for four and a half, running 24-hour development cycles with 12-hour handoffs between continents. When the chips came back and a clock-domain bug threatened to corrupt every result, people quit and called the problem unsolvable. The fix required aligning two clock signals to within 50 trillionths of a second, on every chip, two billion times a second. Gavin solved it in two weeks. They taped out a working chip on their first attempt, the first hardware company founded after ChatGPT to do it, and when silicon came back they had it running inference in a rack in 40 days. A well-known AI chip company took 10 months to do the same. Today Etched has more than 400 people, drawn from Nvidia, Google's TPU team, and Broadcom, and half of them live within five minutes of the office. The company has over $1 billion in customer orders and just raised $800 million. As Rob says, "We're in a new era of intelligence, where the cost of producing intelligence is so dramatically much cheaper than the value of the intelligence, that we are in a many-year, probably many-decade, supply shortage of these tokens." Whoever produces the most tokens, he believes, will be the most valuable company in the world.