Key Takeaways
- The rapid growth of AI data centers is creating a new energy storage challenge. Unlike traditional industrial loads, AI computing workloads can fluctuate by hundreds of megawatts in fractions of a second, creating volatility that the grid was not designed to handle.
- Over the past decade, lithium-ion has become the default technology for stationary energy storage on the electrical grid. But lithium-ion batteries degrade quickly with repeated charge and discharge cycles. This could make them poorly suited for use with the large, frequent swings of data center power usage.
- To smooth out power swings, data centers run “dummy loads” that artificially inflate the amount of energy needed for AI computing.
- Flow batteries’ greater tolerance of these intense power swings makes them an attractive alternative to lithium-ion batteries for the data center market. Flow batteries could reduce the need for energy-intensive, workload-smoothing strategies and help stabilize data center power demand.
- The rise of AI data centers may upend the stationary battery market entirely. Much of today’s battery market was built around charging and discharging once per day to shift renewable energy from low-demand periods into evening peak demand. AI data centers increasingly prioritize power quality, resilience and asset longevity over lowest up-front cost.
Intro: The hidden cost of AI power volatility
Power systems are designed to manage constantly shifting electricity demand. But the rapid growth of AI computing is introducing a new class of load volatility, with data centers capable of increasing or decreasing power consumption by hundreds of megawatts in fractions of a second — an operating challenge unlike anything most grid operators have faced before.
The unpredictability of some AI workloads presents a unique challenge for grid stability. In June 2025, SemiAnalysis, an independent research firm focused on semiconductors, AI infrastructure and data center markets, described how multi-gigawatt-scale data centers are straining our century-old power grid to an unprecedented extent. The challenge is not only due to their massive scale, but also because of the unique load profile of AI training and some real-time analysis applications.
Unlike other large loads like factories or conventional data centers, AI clusters can swing from near-full utilization to partial idle states almost instantaneously. When thousands of graphics processing units (GPUs) enter communications, synchronization, or checkpointing phases simultaneously, power demand can rise or fall by tens or even hundreds of megawatts in seconds.
The electrical grid was not designed for this kind of volatility, and the consequences to the grid would be dramatic. SemiAnalytics cautioned, “At Gigawatt-scale, the worst-case scenario is a blackout for millions of Americans.” According to SemiAnalytics, the issue caught leading AI labs by surprise. Meta acknowledged the challenge of power load fluctuations in a 2024 paper, saying “This is an ongoing challenge for us as we scale training for the future.”
As a stopgap measure, software engineers are responding with the extraordinary tactic of generating dummy workloads to run when a data center’s actual power needs are low. This smooths out the power draw so that it is predictable — but at the expense of wasting an enormous amount of energy.
For a gigawatt-size facility, such as the largest data centers in the United States, this kind of brute-force energy-smoothing means tens of millions in additional and unnecessary annual energy costs.
Battery technologies could help alleviate power variability challenges — but not just any batteries.
Energy storage has a well-established track record of smoothing out grid power demands. For example, battery banks are now routinely co-located with new solar farm installations to even out their daily cycling and provide arbitrage opportunities for power traders.
But AI data centers may expose limitations in lithium-ion systems that are less important in conventional grid-storage applications. A different and underdiscussed class of energy storage technology known as “flow batteries” may prove to be a better match for the unique challenges of AI data center power loads.
Analysis: AI data centers create a different battery requirement.
The unique electrical demands of AI infrastructure could create the first major market opportunity for flow batteries following decades of lithium-ion dominance. Lithium-ion became dominant not because it was the best battery for every application, but because it was the first technology to achieve massive manufacturing scale. The electric vehicle market drove hundreds of billions of dollars of investment into lithium supply chains, creating a virtuous cycle of falling costs, expanding production and increasing investor confidence. By the mid-2020s, lithium-based chemistries accounted for over 90 percent of global battery storage deployments.
The rise of AI data centers may upend the market entirely. Much of today’s battery market was built around charging and discharging once per day to shift renewable energy from low-demand periods into evening peak demand.
AI data centers operate differently: Computing loads fluctuate continuously as processors ramp up and down. Industry observers describe AI facilities as highly volatile electrical loads requiring constant balancing.
Flow batteries appear better suited to rigorous data center demands than lithium-ion batteries.
Flow batteries have an almost-limitless ability to charge and discharge without degrading. Most battery types, including lithium-ion, degrade a bit with each charge and discharge.
Many grid-storage projects are designed for a daily cycling profile. AI data centers require batteries to respond continuously to rapid fluctuations in power demand. Flow battery developers argue that this operating profile favors technologies that can cycle repeatedly with little degradation.
Jon Parrella is the founder and CEO of TerraFlow Energy, a start-up manufacturer and integrator of vanadium flow batteries. Parrella argued that this distinction is particularly important for AI infrastructure. He noted that vanadium flow batteries can perform an “almost unlimited number of cycles,” and last 25 to 30 years while handling the volatility of data center loads.
Parrella explained that the electrical grid cannot be exposed to the extreme variability of AI data center power loads. “The bigger the data center is, the more pronounced the problem becomes,” he said. “When you start talking about gigawatt-scale data centers, you’ll trip the overages in the substations and you’ll cause a rolling blackout event.”
TerraFlow places flow batteries within the electrical architecture of a data center campus so that power continuously flows through the battery. With their approach, the battery acts as a shock absorber, smoothing power fluctuations before they reach the grid.
Lessons from bitcoin mining could inform data center design.
TerraFlow has quickly emerged as one of the most ambitious flow battery developers in the United States. Founded in 2024, the company broke ground in May for a manufacturing facility near Houston designed to produce approximately 300 megawatts of vanadium flow battery capacity annually.
Continue reading in The Washington Post
By Kathryn Clay, Lead Energy & Climate Analyst
