'Bad data means bad predictions': Experts advise Congress on AI load growth

Figuring out just how much energy artificial intelligence will require remains both complex and divisive. The energy demands of data centers are undoubtedly colossal, but experts disagree over just how colossal — some estimates place global data center load growth at more than 1,000 terawatt-hours globally by 2026, while others are expecting a more modest 800 TWh by 2028.

And it’s a divide that played out on Capitol Hill on Tuesday morning before a subcommittee of the House Energy and Commerce committee. Representative statements cited United States load growth statistics ranging from 5% annually to 20% annually, with estimates that data centers would soon consume anywhere from 9% to 12% of the country’s electricity.

• The top line: In a hearing that leaned heavily on concerns about a race to global AI dominance with China — and not-so-heavily on how to make data center load flexible during peak events — the subcommittee on energy, climate, and grid security heard testimony on how to make more accurate predictions of regional data center load growth.
• The current take: Philip Dion, senior vice president of customer solutions at the Edison Electric Institute, said consensus on growth comes down to transparent data. “If we don’t get the right load information, then we’re going to be wrong,” Dion told the subcommittee. “Bad data means bad predictions. We need to be working to find out what's the real load…and what efficiencies can we count on from the data centers and other large users.”

Getting those improved insights will require utilities and tech companies to work collaboratively, Dion said. Load transparency has to be a “two-way street,” he said. 

He added that EEI’s members are willing to cooperate, but doing so “may require some NDAs and other legal vessels.”

Tom Hassenboehler is currently the co-founder and managing partner at energy consultancy CO2EFFICIENT, but was once the chief council for the energy and commerce committee. On Tuesday, he told his former coworkers that getting transparent and accurate demand projections for data center load is a “top, critical issue” that everyone from utilities to tech companies should be working to address.

“We need to really better align economic and load forecasting with electric infrastructure development,” Hassenboehler said. 

And, he added, there needs to be more transparency in the forecasting process, with more stakeholders weighing in; there’s a “need to be at the table on an earlier basis.” Getting over the load growth uncertainty will require, in short, better collaboration between utilities and their AI-hungry customers.

Those conversations, however, aren’t yet happening at the necessary scale. Rep. Nanette Barragán (D-Calif.) pointed to Latitude Media’s previous coverage of the Georgia Public Service Commission’s back-and-forth with Microsoft over data center load predictions. Georgia Power, the tech giant said, was likely overestimating data center load in its updated integrated resource plan last year.

But that’s where the federal government can play a role in getting accurate estimations, said Columbia University professor Melissa Lott (who is also the host of the The Big Switch podcast, co-produced by Columbia and Latitude Studios). In facilitating those multi-stakeholder conversations, she recommended that policy makers consider elements like non-wires solutions, such as grid-enhancing technologies, and potential flexibility, such as during extreme weather events.

At the end of the day, these companies do have clean energy goals, Lott added, but they “don’t want to be utilities.”