Key Takeaways

  • Google will target machine learning workloads for demand response under agreements with Indiana Michigan Power and Tennessee Valley Authority.
  • Long-term curtailment commitments could help data centers connect sooner without forcing utilities to build immediately for every peak-demand scenario.
  • Flexible computing has limits, but it may reduce grid costs and create a new competitive advantage in the AI infrastructure race.

Google is extending its data center flexibility strategy to the machine learning workloads behind large language models, changing how utilities and technology companies plan for rapidly growing electricity demand.

Google’s head of advanced energy stated that agreements with Indiana Michigan Power (I&M) and Tennessee Valley Authority (TVA) represent the first time the company is delivering data center demand response by targeting machine learning workloads. Both utilities operate in regions experiencing substantial interest from data center developers.

The basic idea is familiar: A large electricity customer reduces consumption when the grid is under unusual stress. The difference is how Google, I&M and TVA intend to use that flexibility. Rather than treating demand response only as an emergency measure for existing customers, the agreements incorporate curtailment into long-term planning for new load.

That distinction matters. AI data centers can consume as much electricity as small cities, while generation projects, substations and transmission lines can take years to complete. A data center willing to temporarily reduce selected workloads may gain access to available grid capacity sooner than a facility requiring uninterrupted power at its full nameplate demand.

“We can’t do it everywhere. Some of our loads can’t be curtailed,” the head of advanced energy said. Where flexibility is possible, however, “there’s value to being able to secure capacity without having to wait for new infrastructure.”

Not every AI task is equally flexible. Customer-facing inference, security systems and time-sensitive services may require continuous operation. Some model training, batch processing and other machine learning jobs can potentially be paused, slowed or shifted to another location. Coordinating that activity requires software controls, workload visibility and clear contractual rules covering when curtailment can occur.

Utilities do not typically plan large new loads this way. A Duke University doctoral fellow and former special adviser at the Department of Energy noted that it is not standard practice to let major customers connect on the condition that they reduce consumption during specified periods.

Research from the Duke University Nicholas Institute estimated that U.S. grids could accommodate nearly 100 gigawatts of additional data center capacity if participating facilities accepted targeted annual average load curtailment. In practical terms, that could involve cutting less than half of a facility’s consumption for roughly two hours during peak events occurring across about 100 hours each year.

Why can such a small amount of curtailment have an outsized effect? Power systems are built for their most demanding hours, not merely average consumption. If data centers avoid adding load during those relatively rare peaks, utilities may be able to use existing infrastructure more efficiently while longer-term upgrades proceed.

The stakes are rising quickly. A Lawrence Berkeley National Laboratory assessment estimated that U.S. data centers consumed about 176 TWh of electricity in 2023, equal to roughly 4.4% of national electricity use. Consumption is projected to reach 325 to 580 TWh by 2028, equivalent to 6.7% to 12% of U.S. electricity use.

A Department of Energy report also found almost no examples of grid-aware flexible data center operation in the United States, apart from Google. The I&M and TVA agreements therefore provide an early test of whether flexible AI demand can move from isolated experimentation into utility planning.

TVA’s involvement is especially relevant across the Southeast, including Alabama, where data center activity is developing around Birmingham, Huntsville and Montgomery. Those markets offer fiber, industrial infrastructure and potential grid access, but new projects will still compete for generation and transmission capacity.

There are unresolved questions. Utilities will need ways to verify reductions, define notice periods and protect residential and commercial customers from infrastructure costs created by speculative projects. Developers also tend to withhold facility size and power requirements, making load forecasts difficult to evaluate.

Still, flexibility could change the economics of data center siting. Google may secure electricity sooner, while utilities gain a contractual tool for managing peaks. If the model performs as intended, access to power may depend not only on how much electricity an AI operator wants, but also on how intelligently it can step aside when the grid needs room.