Electricity demand is booming on account of AI.

In a May 2024 report, Goldman Sachs predicted that data centers will use 8% of the U.S.’s total power supply by 2030, up from 3% in 2022, as cloud service providers expand to meet the demand for AI infrastructure. Assuming the current trend holds, U.S. utilities will need to invest around $50 billion in power generation capacity to support all the upgraded — and new — AI-running data centers.

There could be serious negative externalities. In Kansas, where Meta recently broke ground on a massive new server complex, power utility Evergy announced that it would delay the retirement of its coal plant by up to five years. Some experts say that power-hungry data centers — which are also big water guzzlers — could contribute to rising utility costs for everyday ratepayers, disproportionately impacting low-income people.

The data center power consumption problem would appear to be intractable. But Jim Gao, Katie Hoffman and Vedavyas Panneershelvam, the co-founders of Phaidra, believe that it’s possible to retrofit existing facilities to be more energy-efficient.

They’ve built a business out of it, in fact.

Phaidra, launched in 2019, creates AI-powered control systems for data centers as well as pharmaceutical and commercial building infrastructure. The company’s systems gather data from thousands of sensors around a facility and make real-time decisions about how to cool the equipment inside in a power-efficient way.

For many data centers, cooling is one of the most energy-intensive components. The average data center’s cooling system consumes about 40% of the center’s total power.

“The data center industry is in the midst of an arms race to build new capacity wherever land and power are available,” Gao told TechCrunch in an interview. “Phaidra’s service can deliver a more stable cooling system that runs on less energy.”

Gao previously led DeepMind Energy, the team within Google’s DeepMind AI research division responsible for commercializing tech to tackle climate change-related challenges. While at DeepMind, Goa — along with Panneershelvam, then a research engineer at DeepMind — developed an AI system to control and optimize Google’s data centers’ energy usage. It got quite a bit of coverage at the time.

DeepMind made the decision to quietly wind down DeepMind Energy after failing to ink deals with big industry players like British utility National Grid, per CNBC’s reporting. Gao left in August 2019 and Panneershelvam in May 2020 — a few months after the departure of DeepMind co-founder Mustafa Suleyman, who reportedly was a major driving force behind DeepMind’s climate change efforts.

After leaving DeepMind, Gao and Panneershelvam saw an opportunity to apply their lessons from the Google data center project to other data centers — and beyond. They recruited Hoffman, who was heading up innovation projects at Trane, a cooling manufacturing firm, to launch Phaidra.

Phaidra develops AI models for each customer trained on sensor data to optimize a facility’s (e.g. data center’s) cooling systems and overall energy management. These models self-improve, Gao claims, constantly learning from their own experience managing facility infrastructure.

“One of the unique approaches that Phaidra takes to AI is that we combine physics knowledge of how the facility operates with the learned models of the plant dynamics, based upon the sensor data,” Gao said. “The underlying models start with basic representations of standard components, but the semantics and hierarchy of data is configured uniquely from the actual system.”

Phaidra isn’t the only startup to attempt tackling the data center energy usage challenge with AI. Another vendor in the space was Boston-based Carbon Relay, at least until it decided to rebrand and pivot to DevOps and IT.

Elsewhere, Meta and Microsoft have also been experimenting with AI-driven data center optimization. But Gao sees Phaidra’s main competition as “the traditional way of doing things.”

“It’s typical for facilities to hire an outside engineering firm or consultancy to analyze the facility’s performance and manually update the back-end controls programming,” Gao said. “The problem with this approach is that traditional hard-coded controls logic forces the facility to operate the same way forever until somebody goes in to update the back-end programming — which happens every five-10 years in the industrial sector.”

One of Phaidra’s first customers wasn’t a data center operator, but instead big pharma company Merck, which deployed Phaidra’s tech to control a 500-acre vaccine manufacturing plant. Today, however, Phaidra’s clientele skews heavily toward the data center sector — a trend fueled by the AI frenzy, Gao says.

Relatedly, Phaidra was named a finalist in this year’s Amazon Sustainability Accelerator, which gives it an opportunity to pitch for the chance to pilot its tech in Amazon’s European operations with a potential investment of up to €2 million (~$2.15 million). Is Phaidra gunning for an Amazon tie-up? Gao wouldn’t say — but it’d certainly align with the startup’s long-term growth ambitions.

“We have our first international deployments up and running, and expect the higher energy cost regions of the world to fuel a lot of our growth in 2025,” Gao said. “Enterprises are seeking ways to do more with what they have … We’re well-positioned to execute our growth plan over the next two years.”

Phaidra’s making most of its money by charging a SaaS-like annual subscription to its AI. Gao explained: “The fee is a function of the complexity of the facility the AI is managing and the price of energy in the local region.”

Seattle-based Phaidra, which employs around 100 people, recently raised $12 million in a funding round led by Index Ventures. Bringing Phaidra’s total raised to $60.5 million, the new cash will be put toward R&D, implementation, customer success and expanded go-to-market efforts, Gao says.

He expects Phaidra to end the year with a team of 110.

“This was an opportunistic raise that enabled Phaidra to bring Index Ventures to our board and cap table,” Gao said. “Although Phaidra was not actively seeking additional capital, we are especially excited about Index Venture’s scaling expertise as Phaidra expands rapidly with our industrial customers, particularly in the data center industry.”