Phaidra runs the room you never think about.

Somewhere right now, a machine is deciding how cold a warehouse of GPUs should be.

It is not a person. It is not the old control logic a contractor hard-coded in 2014. It is a Phaidra agent, reading thousands of sensors, nudging chillers and pumps every few minutes, and quietly learning whether last week's guess was any good. The GPUs train models. Phaidra trains the building.

This is the unglamorous frontier of the AI boom. Everyone talks about the models. Phaidra took the other job - the one in the basement, near the coolant pipes, where a fraction of a degree turns into a power bill the size of a small town. The company sells AI that runs mission-critical facilities: data centers first, but also pharmaceutical and chemical plants where "good enough" control quietly wastes a fortune.

Founded in 2019 and headquartered in Seattle, Phaidra is roughly a hundred people, remote-first, with a backer list that reads like a who's-who of people betting on compute: NVIDIA, Collaborative Fund, Index Ventures, Sony's innovation arm, Mustafa Suleyman, Mark Cuban. By October 2025 it had raised more than $120 million. Not bad for a company whose product you will never see and whose best days happen when absolutely nothing dramatic occurs.

AI is hungry. The buildings that feed it are run by guesswork.

Here is the inconvenient arithmetic. Training and serving modern AI consumes staggering amounts of electricity, and a large slice of that never reaches a chip - it goes to cooling, to pumps, to fans, to the overhead of keeping the room alive. The industry has a name for that overhead, PUE, and for decades the standard tool for improving it was a skilled human operator squinting at dashboards.

Humans are good. They are also asleep eight hours a day, they retire, and they cannot hold ten thousand variables in their head at once. The control systems meant to help them were written years ago, for steady conditions that no longer exist. So plants run on conservative setpoints - a little too cold, a little too cautious - because nobody got fired for over-cooling. Multiply that caution across the global build-out of AI data centers and the waste is enormous.

That is the tension running through everything Phaidra does: the smartest software on Earth is being housed in buildings run by some of the dumbest. Closing that gap is the whole company.

Three people who had already done this once - at Google.

The bet was personal. Before Phaidra, Jim Gao designed cooling systems inside Google's data centers, then founded DeepMind's energy team. That team built an AI that cut cooling energy by around 40% in a facility that was already considered well-optimized. It worked. The obvious question followed: why should this superpower belong to one company that owns its own data centers?

So Gao left to sell it to everyone else. He brought along Veda Panneershelvam - a primary engineer on AlphaGo, the system that beat the world's best Go player - and Katie Hoffman, an industrial controls veteran from Trane and Raytheon who actually knew how chiller plants behave when nobody is watching. Frontier AI, meet the boiler room.

FIG. 2 - A researcher who beat a board game, an engineer who cooled Google, and an operator who knows what a real chiller sounds like at 3 a.m.

Meet Alfred, the operator who reads ten thousand dials and never blinks.

Phaidra's system - a virtual plant operator the team nicknamed Alfred - plugs into a facility's existing brains. No rip-and-replace, no forklift of new hardware. It speaks the native dialects of industrial buildings, BACnet and OPC-UA, connecting straight into the building automation system or SCADA infrastructure already on site.

Then it watches. It ingests thousands of sensor signals, builds a model of how the plant actually behaves, and uses reinforcement learning to find setpoints a human would never try. Before it touches anything, it runs in shadow mode - predicting, comparing, proving itself against reality. Only then does it take the wheel, adjusting equipment every five to ten minutes, with safety guardrails, a manual override, and one reassuring habit: if the network ever drops, it hands control politely back to the old setpoints and lets the GPUs stay cool the boring way.

The numbers do the arguing.

A pitch about efficiency lives or dies on whether the meter agrees. Phaidra's claim is blunt: 10 to 30% energy efficiency improvements in live deployments, with the software subscription paid back by utility savings inside months. The customers aren't hypothetical either - deployments span data center operators like STT GDC in Singapore and pharmaceutical manufacturing at Merck, plus a collaboration with NVIDIA on reference designs for the next generation of AI infrastructure.

FIG. 3 - Lower is better, which is a sentence energy managers rarely get to say with a smile.

Read the mission twice and the ambition sharpens. Phaidra doesn't just want to tune a chiller. It wants the whole AI factory - power, cooling, workload - to behave as one optimized organism, orchestrated by a portfolio of specialized agents rather than a patchwork of human shifts and frozen logic. Turn the data center from a cost center, the thinking goes, into something that quietly pays for itself.

Why it matters tomorrow is simple math meeting hard limits. The world is building AI factories faster than it is building power plants. Grids are strained, permits are slow, and the heat has to go somewhere. In that world, the company that squeezes 20% more useful compute out of the same megawatt isn't selling a nice-to-have. It is selling headroom - the only kind of growth a constrained grid still allows.

That machine, still deciding how cold the room should be.

Go back to the warehouse of GPUs. A year ago it ran a degree too cold, on setpoints a contractor picked when the building opened, burning power nobody chose to spend. Now a Phaidra agent holds it exactly where it should be - learning, adjusting, handing control back the instant anything looks off. The models upstairs still get all the credit.

Phaidra is fine with that. The best plant operator is the one you forget is there. The company spent six years and $120 million teaching a machine to run the room nobody thinks about - so the rest of us can keep not thinking about it, while the meter, quietly, runs slower.