The rise of artificial intelligence is creating an uncomfortable collision of timescales. AI infrastructure can be designed, financed and deployed at software speed. New transmission lines, substations and generation cannot. The result is a surge of connection requests arriving at a grid that is already managing congestion, ageing assets and the wider electrification of heat, transport and industry.
My earlier EnergyStory article on the surge in data-centre energy demand focused on the scale of the challenge. The next question is more constructive: must every data centre behave like an inflexible block of demand, or can the computing itself respond to the needs of the power system?
A March 2026 report from Emerald AI, EPRI, National Grid and Nebius – Power-Flexible AI Factories: A UK-First Demonstration of Grid-Responsive AI Infrastructure – provides the most tangible evidence I have seen so far. It does not make the energy demand disappear. It shows that a portion of that demand can be controlled with surprising speed and precision.
Why AI compute can be flexible
A data centre is not one indivisible machine. It runs a mixture of workloads. Some inference tasks are latency-sensitive: a user expects an immediate answer. Other work, including training, fine-tuning and batch inference, is more concerned with completing a large amount of computation over time. Within those jobs there are natural flex points, such as checkpoints, scheduling windows and choices about how many processors run at full power.
Modern GPU clusters also offer direct controls. Power caps can reduce instantaneous consumption. Schedulers can pause or deprioritise lower-value jobs. Tasks can move to a different time and, where latency permits, to another region. The critical insight is that flexibility need not mean pulling the plug on the facility. Software can protect priority workloads while less urgent work absorbs the reduction.
What the London trial actually tested
The demonstration took place at Nebius’s London AI Factory using a 130 kW cluster – roughly the electricity use of 400 UK homes, according to the report. It ran continuously for five days with commercially representative AI training workloads. National Grid Electricity Transmission and EPRI issued 22 real-time dispatch events, some with no advance notice and no knowledge of the workload schedule.
Emerald AI’s Conductor software sat between the grid signal and the computing infrastructure. It interpreted requests for a particular reduction, ramp rate and duration, then adjusted GPU power and workload scheduling. The report records 100% compliance with more than 200 requested power targets and ramp rates, including reductions of up to 40% and sub-minute responses.
Fast enough to help during a shock
One test mimicked the sudden demand spike associated with a major televised football match – the old ‘TV pickup’ effect. As the modelled household demand rose, the cluster reduced consumption in the opposite direction, effectively damping the change seen by the local network. Critical workloads continued while lower-priority work was throttled.
A more demanding test replicated the type of contingency seen during the 9 August 2019 lightning-strike event, when the loss of Hornsea and Little Barford contributed to a rapid fall in system frequency and disconnections affecting around 1.1 million customers. In the trial, the AI cluster reduced its consumption by 30% within 40 seconds. In a separate surprise event it reached a 40% reduction in about a minute while preserving the highest-priority workloads.
This matters because frequency events unfold faster than many conventional resources can respond. A large load that can reliably disappear within seconds is, from the grid’s perspective, similar to generation appearing. Engineers sometimes call this a ‘negawatt’: demand reduction that creates headroom at precisely the moment it is needed.
Flexible for hours, not only seconds
Resilience is not only about sudden faults. The system also experiences long periods of stress: a cold evening, low wind, an outage or a regional network constraint. The trial therefore requested 10% to 40% load reductions lasting from two to ten hours. The system followed those targets for the full duration. During the ten-hour case, the report says the highest-priority jobs retained 98.8% performance.
That duration is potentially valuable. Batteries are exceptionally good at rapid response, but many are designed for a few hours of discharge. Flexible computing is different: it is not releasing stored electricity; it is rescheduling work. If the deferred jobs can be completed later, the resource is limited by deadlines, service levels and cooling constraints rather than a fixed tank of energy.
Following carbon as well as price
The cluster also followed a five-minute carbon-intensity signal, reducing consumption when the grid mix was more carbon-intensive and increasing it when emissions were lower. This is a more demanding version of a time-of-use tariff. Instead of assuming that every night is clean and every evening is dirty, the workload responds to the actual changing generation mix.
At scale, this could make data-centre demand a better partner for renewable generation. Training could accelerate when wind is abundant and ease back during tight periods. The benefit would depend on location, network conditions and whether deferred work later creates a rebound peak, but the principle is powerful.
When does a flexible load become a grid asset?
A clever demonstration is not yet infrastructure. For network planners and system operators to rely on flexibility, it must be measurable, repeatable, enforceable and available when promised. The report’s distinction between non-firm and flexible connections is important. A non-firm customer may simply be curtailed without payment when the network is constrained. A flexible customer provides a defined service – response time, ramp rate, reduction, duration and number of events – in exchange for value.
That value might take several forms. A flexible facility could connect sooner because the network does not have to guarantee full demand during every stress condition. It could earn revenue in flexibility markets. Dynamic tariffs could reward operation when network capacity or renewable output is abundant. The report proposes all three: alternative non-firm connection agreements, a market for compute-as-flexibility and dynamic tariffs for very large loads.
The commercial bargain has to work for both sides
For the grid, the promise is faster connections and less unnecessary reinforcement. For the data-centre operator, the prize is earlier access to power and a lower or more predictable energy cost. But the contract must recognise that not all compute is flexible and that service interruptions have a real economic value. A system operator cannot assume a data centre will respond simply because the technology can.
The rules also need to address baselines, measurement, cyber security, control authority and accountability. Who verifies that the reduction is real? What happens if communications fail? Can an operator override the request to protect a critical customer workload? Could many facilities respond simultaneously and create a rebound when they restore demand? These are manageable questions, but they are infrastructure questions, not merely software features.
A promising pilot – with a large scaling challenge
The report is careful about scale. The London cluster was 130 kW; commercial AI campuses are discussed in hundreds of megawatts. Scaling changes the problem. Cooling systems, backup power, transformers, on-site generation, job deadlines and local network limits all have to be coordinated. The announced Aurora project in Virginia, at nearly 100 MW, is intended to test the concept at a more relevant scale.
We should therefore avoid a convenient conclusion that AI growth no longer requires new generation or networks. It does. Flexibility cannot produce energy, and a data centre that shifts work still has to complete it. What flexibility can do is reduce the size of the worst-case assumption, make better use of existing headroom and help align demand with the physical reality of the system.
So, could an AI data centre become a grid asset? The trial suggests the answer is yes – if flexibility is designed into the computing stack, verified at the meter and supported by credible connection and market arrangements. The more useful question may soon be the reverse: if a new AI facility can control a large share of its demand, why would we continue to treat it as if it cannot?
