George Hotz wants to own a zettaflop. That's 1e21 FLOPS, roughly a million times more powerful than what comma currently has, and about a thousand times beyond Frontier, the world's fastest supercomputer at Oak Ridge. In a blog post, the comma founder laid out his vision: a personal supercomputer running one million Claude instances at once. He describes wanting to command "50,000 people working for you, all aligned with you, all answering as one," quoting Vernor Vinge's "True Names" about experiencing sensory bandwidth thousands of times normal.
The numbers are wild but worked through with real math. Hotz sketches 100,000 chips delivering 10 PFLOPS each, powered by 250 acres of solar panels with pumped water storage. Current B200 chips get around 10 TFLOPS per watt, which would mean 100 MW of power draw. He bets efficiency improves 10x, bringing requirements to 10 MW. Total price tag: about $30 million across hardware, solar infrastructure, and land construction.
Tinygrad, Hotz's minimalist deep learning framework, would handle the software side. At roughly 4,000 lines of core code, it strips away abstraction layers and compiles computational graphs directly for specific hardware. That matters. At zettaflop scale, a 1% overhead means wasting 10 PFLOPS, enough to train a decent model on its own. Hotz has called "software being the bottleneck" a solvable problem, and Tinygrad's approach could make custom hardware setups viable where bloated frameworks would choke.
His post ends with five words: "I'll own this before I die." The math holds up well enough to take seriously. If compute keeps getting cheaper and solar keeps falling in price, one person commanding the computing power of a small country stops being science fiction. It becomes an engineering problem with a price tag.