Recursive self-improvement is the new AGI — and George Hotz is raising $20M to serve tokens from a $11.5M Oregon building
Mar 9, 2026
Key Points
- George Hotz is raising $20M at a $200M valuation to buy an $11.5M Oregon building and preorder 3,000 AMD RDNA 5 cards for token serving on OpenRouter.
- His cost structure, cheap consumer GPUs at roughly $2,500 each and $0.03/kWh Oregon power, is designed to make him structurally impossible for well-capitalized rivals to undercut.
- Hotz publicly dismisses AGI as a categorical threshold while raising capital on the explicit assumption that token demand is about to skyrocket.
Summary
George Hotz is raising $20 million at a $200 million valuation for Tiny Corp through what he describes as a friends-and-family round rather than a traditional VC process. Hotz wants to buy an $11.5 million building in Oregon with five megawatts of power, preorder 3,000 AMD RDNA 5 96GB cards at a target price of $2,500 each, and assemble them into 520 Tiny Boxes with six cards apiece. The revenue target is $600,000 per month selling tokens on OpenRouter, which Hotz says represents just 1% of OpenRouter's market depth. Oregon power pricing runs roughly $0.03 per kilowatt hour before peak pricing, putting the electric bill around $50,000 a month. An additional $100,000 per month comes from leasing colocation space to Comma. With TinyGrad optimizations, Hotz projects revenue could reach $5.4 million per month at 3x throughput, with room for three times more machines on the existing power contract. His baseline case has the building and cards paid off in three years.
The competitive argument rests on cost structure. Consumer GPUs and cheap colocation make it structurally difficult for better-capitalized operators to undercut him on price.
On AGI, Hotz rejects the idea of a binary threshold moment. He argues that researchers changing a random seed from 42 to 137 and declaring AGI is no more rigorous than praising a junior engineer for the same move. Hotz argues that software development is getting cheaper and that overpaid engineers are struggling to compete, and he frames recursive self-improvement not as a coming singularity but as something humanity has been doing for centuries.
Hotz is dismissing the notion that anything categorically new is happening while simultaneously raising money to build infrastructure on the assumption that token demand will skyrocket. It is still unclear how those two positions resolve, and at least one observer reads the fact that Hotz is raising at all as a signal of genuine acceleration.