Applied Compute has Palantir's problem, on a clock

"Your data. Your model. Your edge." Applied Compute, a year-old San Francisco startup that trains custom AI models for other companies, puts the line on its homepage, and it means it: when an engagement ends, the customer walks away owning the model. That is an unusual thing for a company now in talks to raise at around $3bn to advertise, because it describes the moment its own product leaves the building.

The round, which the investor Elad Gil is in talks to lead, would more than double the valuation Applied Compute set in April, and it follows the kind of growth that makes allocators forgive a great deal. Revenue has nearly quadrupled since last November, to roughly $50m on an annualized basis. The founders — Yash Patil, the chief executive, with Rhythm Garg and Linden Li — came out of OpenAI, where they worked on the Codex programming assistant and on the reinforcement-learning methods behind the o1 reasoning model. Benchmark, Sequoia, Kleiner Perkins and Lux Capital have already put in $160m. On the growth and the pedigree, the enthusiasm is easy to understand.