For years, the biggest energy question facing data center operators was whether enough electricity would be available to support growth. Today, that question has become more complex.

Total contracted offers in the UK’s electricity demand connection queue rose from 41GW to 125GW between November 2024 and June 2025 – nearly three times the country’s peak electricity demand of around 45GW. Ofgem has indicated the surge was driven largely by data center connection applications.

As AI workloads become larger and more dynamic, operators are discovering that access to power is only part of the challenge. The ability of electricity systems to respond to more variable demand is becoming just as important as the volume of generation available.

The resilience and continued growth of data infrastructure depend on reliable, responsive power, which means data security and energy security are increasingly intertwined – and both are rising up national agendas.

For those planning new facilities or expanding existing campuses, energy flexibility must be weighed as a critical consideration alongside grid connections, cost, resilience and long-term operational performance. Without it, operators risk higher costs, exposure to volatility, slower expansion and facilities that are less able to adapt as AI workloads continue to evolve.

The new shape of electricity demand

Historically, electricity systems have been designed around demand that could be forecast with a reasonable degree of confidence. Industrial facilities, commercial buildings, and even traditional data centers typically showed demand profiles that system operators could plan around.

AI workloads behave very differently. During large-scale model training, tens of thousands of GPUs can increase or reduce their power demand almost simultaneously as workloads begin, pause, or complete. Meta has documented fluctuations of this nature while training its Llama models, while wider research into hyperscale AI clusters suggests these swings can range from tens to hundreds of megawatts.

Average electricity demand may appear relatively stable across a 24-hour period, but the underlying load profile can be far more variable than that of previous generations of computing infrastructure.

Faced with increasingly unpredictable demand, how can operators, investors, governments, and ultimately customers have confidence in a data center’s ability to reliably power performance over time? The answer isn’t just more energy – it’s access to a dynamic system that delivers energy at the right time, with minimal waste and without compromising critical computing performance.

Building resilience through flexibility

There is no single technology capable of reliably meeting AI’s evolving electricity requirements. Instead, resilience comes from combining capabilities that operate on different timescales, from the sub-second swings inside a single training cluster to the hours-long shifts in demand across an entire grid, including:

• Near real-time: On-site battery storage can absorb and release power within seconds, while demand-side flexibility services help coordinate how facilities draw power in response to real-time grid conditions.
• Over minutes: Dispatchable generation – such as gas peaking plants and other fast-start sources – can ramp output up or down to help balance the wider electricity system as demand changes.
• Over hours: Flexibility becomes a scheduling question: shifting non-time-critical computing to periods, or locations, where power is more readily available, while drawing on a diversified mix of renewable and low-carbon generation, including pumped storage hydro.

No single technology can reliably solve this challenge on its own. But together, these approaches help electricity systems respond more effectively to AI’s rising dynamic demand.

Flexibility is becoming a competitive advantage

Across the industry, the common thread is diversification. Rather than depending on one technology or one source of electricity, operators are increasingly looking to combine multiple solutions that improve resilience while giving them greater confidence in their long-term ability to grow.

Some of the world’s largest technology companies are already diversifying their approaches to energy procurement:

• Microsoft has signed a 20-year power purchase agreement with Constellation Energy to bring the restarted Three-Mile Island Unit 1 – now renamed the Crane Clean Energy Center – back online, with power expected from around 2027.
• Google, meanwhile, has diversified beyond its earlier nuclear commitments; in July 2025, it signed a Hydro Framework Agreement with Brookfield Renewable Partners, giving it the ability to procure power from up to 3GW of hydroelectric assets as they are relicensed, overhauled or upgraded.
• Amazon has committed over $500 million to X-energy to develop small modular nuclear reactors, with an initial 320MW project under way in Washington State and an option to scale towards more than 5GW of new nuclear capacity across the US by 2039.

Reliable access to electricity increasingly depends on how much power is available and how effectively that power can respond to changing operational requirements.

Flexibility isn’t theoretical – it’s already working

Greater flexibility can strengthen electricity systems, while delays in deploying it are already creating avoidable pressures.

Battery energy storage has expanded rapidly across the UK over the past decade, with grid-scale battery capacity reaching around 7.5GW by 2025. This is equivalent to around 15 percent of the approximately 50GW of data center capacity currently seeking a grid connection in the UK.

Storing wind and solar power when generation is abundant and releasing it when needed isn’t just more efficient – it also helps reduce a cost the UK is already paying. NESO’s own data shows thermal constraint costs reached £1.7 billion ($2.26bn) in 2024/25, a 64 percent year-over-year increase, driven by rising curtailment as renewable generation continues to outpace grid capacity in some regions.

At the same time, National Grid has confirmed a UK-first trial demonstrating that AI data centers can dynamically adjust their power consumption in response to real-time grid signals without disrupting critical operations. Over five days in December 2025, a cluster of 96 Nvidia GPUs at a London data center responded to over 200 simulated real-time dispatch events from National Grid, cutting power consumption by up to 40 percent on request while maintaining 98.8 percent performance on its highest-priority workloads.

For data center operators, energy flexibility isn’t a future consideration – it’s already shaping who can scale, and how fast.