The Idea
In January 2025, a small Chinese lab released DeepSeek R1 and Silicon Valley lost its composure. Marc Andreessen called it a Sputnik moment. Inside OpenAI, three young researchers read it differently: if they could build the infrastructure to optimise open models, it could mean big business. Yash Patil, Rhythm Garg and Linden Li were all in their early twenties and all out of Stanford. Patil had got in by emailing Sam Altman for a job as a sophomore, been told he would have to drop out, and had Altman offer to talk his parents round personally. He spent two years there on post-training infrastructure and Codex, and left with one conviction: any company running its critical workflows on someone else's model is building on shifting sand. Five months after R1 landed, the three of them quit. This was the beginning of Applied Compute.
The Execution
- May 2025: Patil, Garg and Li leave OpenAI and incorporate in San Francisco on the bet that companies don't want generic AI, they want intelligence trained on their own data. Garg had been a core contributor on o1, Li worked on RL training infrastructure.
- June 2025: A $20M seed led by Benchmark's Victor Lazarte values the company at $100M pre-launch, with Sequoia, Conviction, Hanabi and Definition alongside. Investors were circling at 5x that price before the round was even announced.
- October 2025: The company comes out of stealth with $80M raised from Benchmark, Sequoia and Lux at roughly a $500M valuation, pitching "Specific Intelligence" for the enterprise.
- November 2025: A sceptic on X jokes that Sequoia shouldn't be asked about the revenue. Patil replies with a number, $12.8M annualised six months after incorporation, and the replies pile in asking how much of it is consulting rather than product.
- April 2026: An $80M round led by Kleiner Perkins prices the company at $1.3B post-money, with Elad Gil, Lux, Greenoaks, Neo and Hanabi participating, taking total funding to $160M.
- June 2026: Satya Nadella sits down with Patil and lands on the line the company now leads with, that there should be as many models as there are firms.
- August 2026: The team ships the Agent Cloud, one platform to train, serve and continuously improve models the customer owns outright, alongside frontier-scale RL research on Kimi K3.
- August 2026: Revenue hits $50M annualised, nearly four times the November figure, across customers including DoorDash, Cognition, Harvey and Mercor.
- Today: With about 25 staff, Applied Compute confirms it is in talks for $350M at $3.25B, 15 months after incorporation, with Patil claiming their models can run 10x cheaper than frontier ones.
The lesson?
They never tried to build a frontier model. They built the layer that makes everyone else's models useful, and they spotted it from inside the one company that couldn't sell it. When the replies called it consulting in a trench coat, revenue was $12.8M. Nine months later it was $50M. The best ideas are usually the ones your employer isn't allowed to have.