As attacks on energy systems accelerate adoption of distributed renewables, the resulting need for grid flexibility may turn AI data centers into a new kind of demand-response resource.
March 3, 2026
The impact of AI data centers on power grids and global warming is widely discussed, but an increasing amount of work is going into making data centers a source of energy demand flexibility.
A recent UK experiment reported by Bloomberg offers a concrete example. A group of data centers were able to reduce their energy demand by roughly 30% with about one minute’s notice from the grid, using software from a startup company called Emerald AI. Emerald’s investors include NVIDIA and former US Secretary of State John Kerry.
Emerald’s system, called Conductor, dynamically characterizes how specific AI workloads respond to power capping in real time and then starts or stops jobs accordingly. The company says the system can respond to changing grid conditions in roughly five-minute operational increments, aligning with the kinds of variability introduced by renewable generation.
Emerald’s white paper reports from the experiment:
“The AI cluster autonomously ramped down its power consumption with seconds-level precision.
Inverse Correlation: The AI power draw created a perfect inverse mirror of the demand spike. As residential load rocketed up, the AI load dropped, effectively neutralising the net load increase on the local substation.
SLA Preservation: Despite the rapid depowering, the Emerald AI Conductor ensured that critical high-priority jobs continued without latency spikes, while some lower-priority jobs were momentarily throttled.”
The result suggests that AI data centers may be capable of acting as a form of fast demand response. By shedding load during peak demand events, they can effectively create “negawatts,” reducing strain on the grid.
This kind of flexibility is becoming increasingly valuable as renewable energy becomes a larger share of electricity generation.
Solar and wind generation are variable and weather-dependent. As their share of the energy mix grows, power systems increasingly rely on demand flexibility, storage, and other balancing mechanisms to maintain grid stability.
Researchers have pointed out that data centers are particularly well positioned to help provide this flexibility. A study in Elsevier’s journal Energy Reports titled “Data centres as a source of flexibility for power systems,” by researchers affiliated with Cardiff University and the Research Institutes of Sweden, finds that characteristics such as workload classification — especially the distinction between delay-sensitive and delay-tolerant computing tasks — can enable meaningful demand-side flexibility while preserving service-level agreements.
One reason data centers may be unusually well suited to provide demand flexibility is the nature of many AI workloads themselves. Unlike most industrial electricity consumption, which must run continuously once started, many AI training and batch inference jobs can be paused, slowed, or rescheduled without breaking the service they provide. Cloud schedulers already move workloads between servers and data centers to optimize cost and performance. Extending that same scheduling logic to electricity availability — briefly throttling power consumption during grid stress or shifting work to times of abundant renewable generation — could turn large computing clusters into a new kind of controllable demand resource for power systems.
Bloomberg’s coverage of the Emerald experiment focused on one immediate incentive: new data centers may want to demonstrate demand flexibility so that utilities and regulators will allow them to connect to local grids more quickly.
AI infrastructure is already connected to conflict in more direct ways. Data centers themselves can become strategic targets, and AI systems are increasingly used for military operations and intelligence analysis. The connection explored here is different: how pressures on energy systems created by conflict may indirectly shape the role data centers play in civilian electricity grids.
But the implications may extend further upstream.
As geopolitical conflict increasingly targets energy infrastructure, there may be additional pressure toward distributed energy systems and the kinds of flexibility those systems require.
International humanitarian law places constraints on attacks against civilian infrastructure, including power systems. Civilian objects may not be directly targeted, and even attacks on military objectives can be unlawful if the expected civilian harm is disproportionate or if they deprive civilians of essential services.
In practice, however, electrical infrastructure has become a frequent target in modern conflicts.
One response has been increased interest in distributed renewable energy systems. Compared with centralized coal or gas plants, solar and wind installations can be geographically dispersed, meaning that disabling them often requires many more individual strikes. They can also sometimes be repaired or replaced more quickly than large centralized generation facilities, though they still depend on shared grid infrastructure like substations and transmission lines.
Bill McKibben wrote this weekend about how geopolitical conflict can accelerate renewable deployment in some contexts. Quoting reporting from Yale Environment 360 on Ukraine, he notes:
“Wind and solar arrays with independent transmission lines are scattered over the landscape, which makes them harder to hit and easier to repair. ‘A coal power station [is] a large single target that a single missile could take out,’ says Jeff Oatham of DTEK, Ukraine’s largest energy company and its largest private energy investor. ‘You would need around 40 missiles to do the equivalent amount of capacity damage at a wind farm.’”
The Yale360 reporting continues:
“Solar, too, makes an unattractive target. ‘Attacking decentralized solar power installations is not economically rational,’ says Ukrainian energy expert Olena Kondratiuk. ‘Missiles and drones are expensive, and significantly disrupting such systems would require a large number of strikes, while the overall impact on the energy system would remain limited.’”
If conflict and energy security concerns encourage greater adoption of distributed renewable energy, the implications extend beyond generation technologies.
Grids with higher shares of intermittent renewables require more flexibility — the ability to ramp demand or supply up and down quickly to balance changing conditions.
AI data centers, once seen primarily as inflexible and power-hungry loads, may increasingly become part of that balancing system.
If that trend continues, a more conflict-ridden world could indirectly increase demand for technologies that help large computing facilities modulate their power consumption in real time.
In an interconnected system, pressure applied in one place — geopolitical conflict — can propagate through the system and surface elsewhere, in this case as demand for software that helps data centers flex with the grid.
This post emerged from an effort to trace causal relationships across several developments in energy systems and AI infrastructure. The goal of this workflow is to surface relationships between developments that are usually discussed separately — in this case, geopolitics, renewable energy systems, and AI infrastructure.
The starting point was a Bloomberg article describing an experiment in which data centers reduced electricity consumption by roughly 30% with about one minute’s notice from the grid. I discovered that article through a network-scanning system I use called Hawkeye, which monitors developments across organizations and technologies in my broader business ecosystem.
After reading the article, I analyzed it using What’s Up With That? (WUWT), an AI toolkit I built to assist with research and critical reading. The system performs several analytical passes on a text, including red-team critique, scientific literature review, and identification of related topics worth investigating further.
One capability of WUWT is detecting causal claim chains across multiple pieces of content. In this case, it highlighted a possible connection between research on data-center demand flexibility and a Bill McKibben article I had read earlier discussing how geopolitical conflict can accelerate the deployment of distributed renewable energy.
That suggested the broader causal chain explored in this post:
conflict → distributed renewables → grid flexibility → AI data centers
The causal-chain detection approach is partly inspired by a recent U.S. Department of Energy RFP exploring AI systems capable of identifying causal claims across large bodies of text.
Both Hawkeye and WUWT are tools I use regularly in my research workflow. On war in Iran, I also recommend my friend Janet’s post “Let’s All Start Screaming Now.”
Chrome
Firefox