Data center operators have a new answer when communities push back on water use. Switch to a closed-loop system, fill it once, and the facility uses about as much water as a fast-food restaurant.
That answer is true. It's also incomplete.
"I remind you that water is 3,500 times more efficient in rejecting heat than air is," said Scott Charter, head of go-to-market for data centers at Physical Superintelligence (PSI), when we spoke during Yotta 2026. Closed-loop designs rely on dry coils, which means more fans running harder to move the same heat. "So that means I'm going to be wasting a lot of electricity in order to be able to say we don't need to use your water."
You save water, and you pay for it in power. That's the kind of coupled trade-off PSI built its platform to find.
PSI launched September 1 with a $58 million seed round led by Breakthrough Energy Ventures. It calls itself an AI-native physics lab, and its platform is called Emmy. The first commercial target is the data center: new builds on the ground and in orbit, plus retrofits of facilities already running.
Charter spent 35 years in the industry, the last eight at Oracle, before joining PSI in July. What stood out in our conversation was how Emmy reasons.
"A transformer is great at predicting the next, whether that's word or pattern or whatever, but that is based just on patterns," he said. "That is not always made on doing backend math to get to the correct answer."
Emmy uses a multi-physics engine with surrogate models. Charter described it as starting from the result you want and solving back to where you are. "We want to come to Z with an output, and then back into A."
A human engineer with a digital twin might test three ideas. Charter described a "deep twin" that starts much wider. In his example, Emmy looks at 500,000 possible designs and narrows them to about 750, possibly in 14 minutes. Engineers then go deep on that shortlist and bring three options to the client. Compare that to a team of 20 people spending four to six months on four designs.
The water example shows what that wider search can surface: a third path. Take dirty, compromised local water that no one wants, treat it, and use it to run evaporative cooling towers, still the most efficient way to cool a facility. Then return clean water to the community.
"I go from villain to hero," Charter said.
The obvious question is trust. When an LLM is wrong, you get a bad paragraph. When a thermal model is wrong, you get a hot aisle and throttled GPUs.
Charter's answer is to show the math. Candidate designs get modeled in a simulation environment, such as NVIDIA Omniverse, before anyone pours concrete. "You have to be able to show your log data and be able to show your math on every step of that turn," he said. "So it is the opposite of a black box."
He was candid about where the platform stands. "There is probably more manual handholding of Emmy today than there will be in six months," he said. Automation will grow, but not without oversight. "This will always be with a human in the loop. Let's call it a physicist in the loop."
The engineering pressure behind all of this is rack density. Enterprise racks averaged 8 to 10 kW in 2020. The latest systems run around 130 kW, and NVIDIA's roadmap points toward a megawatt. "Anything when you eclipse beyond anywhere from 50 to 70 kW, you cannot use air only," he said. "We have to go to liquid."
Liquid doesn't eliminate air, though. Most facilities run a hybrid. At these densities, power, cooling, networking, and compute stop behaving as separate systems. Change one, and the others move.
PSI's commercial model lowers the buyer's risk. There's no upfront fee. PSI agrees on a customer's current design baseline, finds improvements, and takes a share of the gains. "If we can't prove ourselves, you kick us to the curb," Charter said. On retrofits, he said the team can find meaningful changes in days.
Conversations start in the C-suite, where leaders decide how to spread billions across a portfolio of sites. "Whoever is the loudest in that team probably gets a little bit more of a spend, but that's not always the most logical place to go spend that dollar," Charter said.
There's a gap worth watching. PSI hasn't published independently verified performance numbers yet. Asked about that, Charter said the company will keep publishing research separate from its customer work. PSI is a public benefit corporation, and he said the goal is to share what it learns. "We want to be the next Bell Labs."
His biggest concern isn't the technology. It's people. PSI has about 32 employees, mostly PhDs, including theoretical physicists and astrophysicists. "It's easy to hire a 50-person dream team," he said. "It's much more difficult to keep that same level when it's a 500-person dream team."
The water-versus-power trade-off is a useful lens. Data center decisions that look simple rarely stay simple once you account for everything they touch. Physics-based AI that explores thousands of options and shows its work could help teams see those trade-offs before they build. Now PSI needs to put numbers behind that promise.
