What “flexible” data-center power means
Kavya Balaraman’s October 8 Reuters explainer describes a growing effort to make data centers responsive to grid conditions. Under demand response, an operator temporarily reduces or shifts electricity use during peak demand or an emergency. For an AI facility, that might mean rescheduling some computing tasks rather than drawing maximum power at the same time as homes, businesses and other industry.
The appeal is straightforward: if a large customer can reliably ease off at the moments that matter most, a utility may need less new capacity to serve peak demand. Flexibility might also help some data centers connect to constrained grids sooner. But an occasional pilot is not the same thing as a dependable system-wide resource.
The scale of the opportunity—and the caveats
Reuters cites the Electric Power Research Institute’s estimate that U.S. data-center electricity use could rise from roughly 177–192 terawatt-hours in 2024 to 383–793 terawatt-hours by 2030. That broad range reflects uncertainty about how fast facilities are built and how much energy they consume. EPRI’s surveyed facilities reported potential peak-power reductions of 10%–30%, depending on facility type; those figures should not be treated as a guaranteed reduction for every data center.
A Duke University Nicholas Institute study estimated that greater data-center flexibility could avoid $40 billion–$150 billion in capital investment over the next decade. This is a scenario estimate, not money already saved. The result depends on how much load can actually move, when the grid needs relief, and how utilities plan around that response.
Pilots are becoming a bigger policy question
Most data-center demand-response efforts are still pilots or one-off agreements. Reuters points to an OpenAI agreement in Georgia that could reduce grid draw by up to one gigawatt during grid stress at a planned 3.2-gigawatt facility. Google, Nvidia and Emerald AI have also launched an AI Energy Management Alliance to advance flexible data centers. These examples show momentum, but not yet a proven nationwide operating model.
The missing pieces are both technical and commercial. Operators need to identify workloads that can pause, slow or shift without breaking customer commitments. Grid operators need to know that promised reductions will arrive at the right time. Utilities and regulators need tariffs, incentives and interconnection rules that reward reliable flexibility rather than merely announcing it.
What to watch next
The key question is whether flexible computing can be measured and trusted as a real grid resource across hundreds of facilities. If it can, it could reduce some peak infrastructure costs while making room for AI growth. If it cannot, the savings may remain largely theoretical. The Daily AI Chat episode and Short above unpack the opportunity and the practical barriers without treating projected savings as a done deal.
Source and listen
Source: Kavya Balaraman, Kavya Balaraman’s October 8 Reuters explainer, Reuters, October 8, 2026. Watch the YouTube Short or listen to the full podcast episode.
