He kept Google's servers from overheating for a decade. Then he built an AI that knew the cooling plant better than he did.
Jim Gao runs Phaidra, a Seattle company that does something most operators find slightly terrifying: it lets an AI take the wheel of the physical world. Data centers, vaccine plants, the giant chillers that keep AI's servers from cooking themselves - Phaidra's reinforcement-learning agents run them, in real time, on their own. The pitch is not science fiction. Gao has already proved the core of it inside the most demanding infrastructure on earth.
Start with what Gao is building today, because it tells you where his head is. Phaidra makes control systems that learn. Not dashboards. Not alerts for a human to act on. Software that decides - which pumps to run, how fast, at what temperature - and keeps deciding, second after second, inside facilities where a wrong move costs millions or melts a server farm.
The timing is not subtle. The AI boom runs on data centers, and data centers run hot. Every chatbot reply and training run throws off heat that has to go somewhere, and the cooling bill is now one of the largest line items in the business. Phaidra sells the brain that shrinks it. In 2025 the company partnered with NVIDIA, CoreWeave and Applied Digital on agentic liquid-cooling management for the new generation of "AI factories" - the purpose-built data centers feeding the models.
The customer list reaches past servers. Phaidra's first big public name was Merck, where its AI helps run a vaccine manufacturing site that sprawls across 500 acres. The thesis is that any facility you can describe as a constrained optimization problem - here is the goal, here are the knobs, here are the rules you cannot break - is a candidate for an agent that runs it better than a control loop written by hand twenty years ago.
Gao reduces the whole field to a checklist. An objective function. A set of controllable actions. A set of operational constraints. Give a reinforcement-learning system those three things and a clean stream of data, and it can learn to drive almost anything. The hard part, he is quick to add, is rarely the AI. It is the data. Most industrial operators sit far down what he calls the "Maslow's hierarchy of data needs" - no storage, no cleaning, no streaming access - and the first job is often just getting the plant's own numbers into a usable shape.
The real promise of AI isn't automation. It's AI creativity - the ability to discover knowledge that didn't exist before.
Jim Gao, on Sequoia's "Training Data""This very AI agent that we created is telling me new things about the system I designed. That's a very, very powerful feeling."
- Jim GaoTwo bachelor's degrees from UC Berkeley: mechanical engineering and environmental science. One teaches you how to move heat. The other teaches you why moving it efficiently matters for the planet. Gao's whole career sits at the seam between them.
In 2013 he worked through Andrew Ng's Coursera machine-learning course and used Google's famous 20% time to test a hunch: that the math behind game-playing AI could run a building.
Gao did not arrive at AI through a PhD or a research lab. He arrived through the boiler room. He joined Google around 2011 as a data center engineer - designing the large cooling systems that pull heat off thousands of servers and running Power Usage Efficiency analysis to squeeze out waste. It was hands-on, physical, deeply unglamorous work, and he was good at it.
Then in 2016 a computer program named AlphaGo beat one of the best Go players alive, and Gao saw it differently than most people did. Where others saw a board game, he saw a system learning to make a sequence of decisions toward a goal under hard rules. That, he realized, was a description of his day job. He made the case internally and partnered with DeepMind to point reinforcement learning at Google's own cooling plants.
The result became one of the most cited examples in applied AI: up to a 40 percent reduction in cooling energy, inside a facility human engineers had already optimized for years. The savings were not from cutting corners. The AI honored every constraint and still found counterintuitive moves the experts had missed. Gao went on to lead DeepMind Energy, a team of more than forty experts building AI to control mission-critical data centers from the cloud.
One moment from those years stuck. Standing in a cavernous data center, his future co-founder Veda Panneershelvam pushed code that remotely switched on a chiller the size of a bus. The infrastructure that keeps the internet alive, controlled from a laptop in the cloud. That was the proof. The rest was a company.
Phaidra's investors are not buying a clever demo. They are buying the claim that reinforcement learning belongs in the control room of every heavy facility on the grid - and that the AI boom has made the timing urgent. The same servers training the models need cooling, and the cooling needs a brain.
Gao tends to point the credit elsewhere. He defers the deep technical questions to Veda, whose AlphaGo work seeded the whole idea, and frames himself as the domain guy who knew the plants. It is an unusually modest posture for a founder selling autonomy.
The real promise of AI isn't automation - it's AI creativity, the ability to discover knowledge that didn't exist before.
This very AI agent that we created is telling me new things about a system I designed.
Any problem you can map to constrained optimization - an objective, controllable actions, and constraints - is a candidate for reinforcement learning.
"Phaidra is Gao's argument that the next industrial revolution won't be built by faster machines, but by machines that learn what no one taught them."
From data centers to vaccine plants to the power grid