Artificial intelligence has been called the “new electricity” of the digital age, and while that description captures its transformative potential, it also highlights a growing problem: AI’s massive appetite for power. Data centers — the physical backbone of the AI revolution — are consuming more energy than ever before, pushing local grids to their limits.
On Monday, Google took a notable step forward in addressing this challenge by announcing new partnerships with Indiana Michigan Power and the Tennessee Valley Authority. These agreements will allow Google to throttle energy demand from machine learning workloads at its data centers in those regions during times of peak grid stress.
This initiative builds on Google’s earlier success with the Omaha Public Power District, where the company managed to reduce machine learning power demand during three grid events in 2023. The approach? Shift non-urgent computing tasks, like processing YouTube videos, to different times of the day when the electrical grid is under less strain.
Why AI is Straining the Grid
Generative AI and machine learning are computationally expensive. The chips that power these workloads consume enormous amounts of electricity, and the demand is scaling rapidly.
Morningstar Research Services predicts that U.S. data center power capacity will nearly triple to 80 gigawatts by 2030 — and some forecasts go as high as 100 gigawatts. To put that in perspective, that’s the equivalent of adding dozens of large power plants to the grid, solely to handle data center loads.
But building new power generation facilities is not a quick fix. Grid interconnection queues are already years long, and in some regions, there are physical bottlenecks, such as shortages of high-voltage transformers, that slow expansion even further.
Rob Enderle of the Enderle Group warns that without intervention, the situation could lead to extended brownouts or even catastrophic outages. Simply put, the grid as it stands cannot keep pace with AI’s growth without smarter demand-side solutions.
Demand-Side Solutions: A New Path Forward
Rather than solely focusing on adding more power plants, demand-side strategies aim to better manage when and how energy is used. For AI data centers, this means flexibly shifting workloads so they use less power when the grid is stressed.
Pete DiSanto of Enchanted Rock explains it plainly: “Without demand-side solutions, the grid simply won’t be able to keep up with AI data center growth, especially in regions already facing capacity challenges.”
Google’s agreements essentially allow its data centers to act like “virtual power plants,” scaling down certain workloads to free up capacity for the grid during high-demand periods. These strategies have the added benefit of making companies more attractive to utilities, potentially putting them at the front of the line for future power allocations.
Protecting Performance While Saving Power
One natural concern is whether reducing data center power usage during peak periods will slow down or disrupt AI services. The answer, according to experts, is that with smart planning, most users will never notice.
Wyatt Mayham from Northwest AI Consulting explains that not all AI workloads are created equal. Pausing a large-scale AI model training run or a batch analysis for an hour is far less impactful than interrupting real-time search or chatbot services. By targeting only non-latency-sensitive tasks for curtailment, companies like Google can maintain critical performance while still providing grid relief.
In cases where cutting power from the grid could threaten operations, data centers can draw on backup generation — such as onsite natural gas generators or battery systems — to keep performance uninterrupted.
The Real Bottleneck: Transmission, Not Generation
A key point often overlooked in the AI energy debate is that the problem isn’t only how much power we can generate — it’s whether we can get that power to the data centers.
Many projects are facing multi-year delays because of grid hardware shortages. Even if a utility can generate hundreds of megawatts, the local infrastructure might not be able to deliver it where it’s needed. Google’s demand-response approach sidesteps some of these delays by making better use of existing infrastructure, reducing the need for new, time-consuming grid buildouts.
An “All-of-the-Above” Energy Strategy
Industry leaders agree that demand-side solutions alone won’t solve the problem. The future of AI’s energy footprint will require a mix of approaches, including:
-
Advanced cooling technologies like direct-to-chip liquid cooling to reduce waste heat and power usage.
-
Long-term clean energy contracts with sources such as nuclear plants.
-
Onsite generation through small modular reactors or other dedicated power sources to bypass grid limitations entirely.
Ezra Hodge of EMA Partners adds that talent will play a crucial role in executing these solutions. Data centers will need leaders who understand both AI infrastructure and the complexities of the energy market — a rare but increasingly essential skill set.
Conclusion: The New Era of Energy-Aware AI
Google’s latest agreements are a glimpse into a future where AI innovation and energy management go hand-in-hand. Instead of AI simply being an energy problem, it can also become part of the solution — acting as a flexible, responsive partner to the grid.
As demand grows, strategies like these will move from being experimental to essential. The companies that master this balance between performance and sustainability will be the ones best positioned to lead the AI-powered world of tomorrow.
