Reducing demand 40-70 hours a year (1%) means the data center does not require generation upgrades that raise consumer prices. If the data center can cut demand or switch to a backup generator, the increased revenue would allow prices to drop.
Google supported studies at Columbia and Duke universities that confirmed the conclusion.
99% of the time, grids are not running at capacity. In the U.S., about 65% is typical. Grids are sized with enough uninterruptible peak power to avoid blackouts except during weather or other unexpected emergencies. If a new large load needed service all the time, new construction would be required to meet the higher peak.
Google’s announcement
Google published this post on August 4, 2025, under Michael Terrell, its head of advanced energy. It announces two new demand response agreements — with Indiana Michigan Power (I&M) for a new data center in Fort Wayne, Indiana, and with the Tennessee Valley Authority (TVA) — and calls them the first time Google is applying demand response specifically to machine learning workloads, rather than deferrable tasks like video transcoding.
The stated basis for confidence is a prior pilot with Omaha Public Power District (OPPD), where Google says it cut ML workload power draw during three grid events in 2024. Google runs a separate, older demand response program abroad — with Centrica Energy and transmission operator Elia in Belgium, and with Taiwan Power Company in Taiwan — but that program shifts non-urgent compute tasks, Terrell’s example being YouTube video processing, not ML workloads directly.
The post contains no megawatt figures, no curtailment duration or frequency caps, and no data on what the three OPPD events actually reduced. It’s an announcement of a capability and a partnership, not a technical filing.
Since publication, the I&M side of the deal has moved through Indiana regulatory review. The Indiana Utility Regulatory Commission approved the Google–I&M arrangement on April 29, 2026 — nearly nine months after this post. Local coverage describes a “capacity commitment framework” requiring large load customers to make long-term financial commitments proportional to their load, plus a rate discount for Google when it curtails. Neither company disclosed further terms. The Fort Wayne data center itself — previously known as Project Zodiac, first reported in April 2024, planned as up to 12 buildings across more than 700 acres — went operational in its first phase on December 11, 2025, before the demand response framework had regulatory sign-off. I&M separately filed in February 2026 to lower base electric rates, citing increased revenue from large customers including the Fort Wayne site as the justification; nothing in that filing ties the rate change directly to demand response performance. No comparable public filing or news update on the TVA side turned up in this search — that piece of the announcement remains unverified past the original post.
What the agreements do
This isn’t a study, so there’s nothing to project — but it’s worth being precise about what the post commits Google to versus what it merely gestures at. It commits to reducing ML workload power consumption at Fort Wayne and at TVA-served sites “during certain hours or times of the year” when asked. It gestures at, without specifying, which ML workloads are actually interruptible — training runs, inference serving, or something narrower — and at what fraction of total site load that represents. Terrell writes that reliability floors exist for services like Search, Maps, and healthcare cloud customers, which implies some ML capacity is walled off from curtailment, but the post gives no breakdown of protected versus flexible load at either site.
Author and credits
Michael Terrell, Head of Advanced Energy at Google. Published on Google’s corporate blog, August 4, 2025.
Link
Summary
The post runs three sections: an introduction framing flexible demand as a way to interconnect large loads faster and reduce transmission buildout, a middle section (“Delivering flexibility to strengthen grids”) laying out the I&M and TVA deals against the OPPD track record, and a closing section (“Evolving for future energy needs”) that hedges — demand flexibility is “still in the early stages,” available “only at certain locations,” and won’t replace new generation and transmission investment.
Structurally, it’s a policy-signaling document aimed at utilities, regulators, and grid planners as much as at the public: it positions Google as a flexibility resource that can be integrated into long-term resource planning, not merely a load that shows up and demands capacity. That complicates the standard framing of AI data centers as pure grid strain — but only rhetorically, since the post supplies no verifiable numbers on how much flexibility Google is actually offering or how reliably it delivers on curtailment requests.
Conclusions
The post lands on a simple claim: ML workloads can be made flexible, Google has done it at OPPD, and it’s now scaling the approach to two more utilities. It declines to quantify any of it — no megawatts, no percentage of load, no verification that curtailment didn’t degrade service for cloud customers during the OPPD events. What’s happened since narrows the gap between announcement and reality only partially: the I&M piece cleared Indiana regulatory review eight months late, arriving after the data center was already operational, while the TVA piece has no public follow-up at all. Read this as a signal of utility-hyperscaler coordination on interconnection speed and cost allocation, not as evidence of measured grid impact.
edited and authored by Dave with close collaboration by Claude