Artificial intelligence (AI) data centres are imposing new stresses on power systems as rapid swings in demand resemble the loads of factories and towns. Training large models mobilises thousands of graphics processing units in unison, causing power draw to surge within milliseconds and at times to rise by one point five times design capacity. Mechanical failures have been reported as rotating equipment and small combustion engines suffer repeated shocks, and turbines at one US facility developed cracks with similar damage observed at smaller sites in the UK; crankshafts and other parts have broken under cyclical strain. Batteries, capacitors and flywheels are being deployed to smooth flows but some battery systems have required replacement within weeks, and operators warn that insufficient stabilising technology at some new builds is compounding risks. Reliability problems are delaying projects and cutting revenue at some AI campuses, with a planned two point six seven-gigawatt site pushed back by a year for engineering and power delivery. The cost of downtime is driven chiefly by lost compute revenue rather than the price of replacing pumps or breakers, and some facilities report uptime closer to 80 per cent. Regulators and grid operators have warned that many data centre load models do not capture dynamic behaviour and that greater coordination between load, generation, storage and controls is required. One agency evaluated more than 33 gigawatts of operational centres and issued a level three alert asking large centres to address risks and submit responses by an August third deadline. Chipmakers and equipment vendors are working with power experts and operators are using techniques to keep processors running steadily, though some have criticised side computations as wasteful. Test beds combining GPUs, on site generation and storage are being used to trial batteries, software and other systems intended to smooth oscillations between facilities and the grid. Stakeholders say there remains an opportunity to avoid long term instability if planning for load, generation, storage, controls and grid connection is integrated during design.