‘Over-Revving the Engine’: AI’s Volatile Power Demand Is Damaging the Data Centers Behind It

Artificial intelligence is already powerful enough to crack turbines and burn out batteries in the data centres that feed it, as its surging and volatile electricity use shifts from straining power grids to physically damaging the infrastructure behind the boom.

Rapid, millisecond‑level swings in power demand from AI workloads are forcing cooling and backup systems to fail early, raising concerns over reliability, safety and cost just as investment in hyperscale sites accelerates.

The problem is not simply that AI uses a lot of power, but that it uses it in sharp, unpredictable bursts that traditional data‑centre infrastructure was never designed to handle.

Operators and engineers describe facilities as effectively ‘over‑revving’ their engines, with thousands of graphics processing units (GPUs) firing in sync and then dropping off almost instantly, creating repeated electrical shocks inside buildings and across the wider grid.

Infrastructure Under Stress From AI Workloads

Core systems in AI‑heavy data centres, including uninterruptible power supplies, backup generators and cooling units, are malfunctioning or reaching the end of their lives far sooner than expected.

In some cases, equipment is being retired ahead of schedule or suffering catastrophic failures, adding unforeseen capital costs for operators already under pressure to recover billions of dollars invested in hyperscale sites.

Specialists say the issue is most acute during large training runs, when tens or hundreds of thousands of GPUs start, stop and restart in tightly synchronised waves.

Pete Bowen, chief technology officer at UK colocation provider Carbon‑Z, notes that large AI clusters can drive power swings of hundreds of megawatts within seconds, a pattern that makes reliable grid operation significantly harder.

Sreemant Roy, a power‑quality expert at Schneider Electric, says that such ‘extremely dynamic or fluctuating’ loads can cause grid instability and, if unaddressed, may lead to blackouts or power outages. He warns that AI‑driven data centres could trigger sub‑synchronous oscillations that damage equipment connected to other parts of the network, a scenario that has worried utilities and system planners.

Downtime at AI‑focused facilities can cost from several thousand dollars to hundreds of thousands of dollars per minute, and that financial exposure is pushing developers to oversize power and cooling assets to cope with extreme peaks rather than average loads. The result is lower overall efficiency and longer delays before major projects break even.

Grid Risks and UK Data Centre Concerns

Beyond individual sites, grid planners are flagging AI‑driven load swings as a ‘high likelihood, high impact’ risk to wider system stability. A white paper from the North American Electric Reliability Corporation (NERC) said rapid power fluctuations from tens of thousands of chips during AI training could throw carefully calibrated grids out of balance and, in extreme cases, trigger cascading outages.

One event described by Data Center Knowledge saw dozens of data centres in Northern Virginia drop off the grid in a single incident, instantly removing roughly 1,500 megawatts of load and forcing operators to scramble to stabilise the system. In the UK, similar concerns are emerging as proposed data‑centre projects seek more electricity than the country’s entire current peak demand.

Ofgem’s recent call for input, cited in the CMS briefing, noted that around 140 proposed schemes, many linked to AI, could require roughly 50 gigawatts of power, far outstripping existing capacity and grid‑connection queue availability. Industry executives have also warned that energy, not compute, is now the main constraint on UK AI data‑centre expansion, with power shortages and grid delays affecting new builds.

Rethinking Where and How Data Centres Get Built

That pressure is feeding into debates over how new facilities should be planned, with some experts calling for stricter interconnection rules and others pushing operators towards on‑site generation and storage to handle volatile AI power demand.

For readers, the core risk is higher long‑term electricity costs and regional limits on where new data‑centre projects can go ahead, as regulators and utilities adjust to these new load patterns.

Others warn that the sector’s growing reliance on ever‑larger battery systems risks treating the symptom rather than the cause, and say deeper changes in how AI jobs are scheduled and priced will be needed to protect infrastructure and grids.

For now, the consensus among technical experts is that AI’s power problem is not just a software issue or a temporary kink in demand curves, but a fundamental engineering challenge.

How quickly operators, utilities and regulators respond will determine not only where the next wave of data centres gets built, but how reliably the systems behind everyday AI tools can stay online in the years ahead.