AI data centres turn to on-site power

AI data-centre developers are moving power generation onto their own sites because utilities in several markets cannot connect new projects fast enough. DigiTimes reported on May 24 that operators are increasingly pairing large AI facilities with gas turbines and battery systems as computing loads rise toward the gigawatt scale. (taiwannews.com.tw) Taiwan has become one of the clearest examples of the strain. The Ministry of Economic Affairs said in April that electricity demand from AI data centres and computing facilities in Taiwan could reach 1 gigawatt by 2030, and said a single AI server site could consume as much power as a nuclear reactor unit or a large coal-fired plant produces. (moeaea.gov.tw) The result is a shift in how projects are built. Instead of waiting years for grid upgrades, developers are increasingly trying to secure “behind-the-meter” supply — power generated at or near the site and used directly by the data centre. (digitimes.com) DigiTimes said that change is being driven by higher GPU power draw, grid bottlenecks, and shortages of equipment including gas turbines and transformers. ### Why are AI data centres asking for so much power so quickly? AI servers are denser and more power-hungry than earlier cloud infrastructure. (taiwannews.com.tw) DigiTimes analyst Sabrina Yu said AI data centres face four energy pressures at once: rising GPU thermal design power, a shift toward high-voltage direct current architectures, grid constraints, and tougher carbon requirements. The International Energy Agency has also said there is “no AI without energy,” particularly electricity for data centres, and has been studying how AI deployment is changing power demand. (digitimes.com) ### Why is Taiwan drawing special attention? Taiwan’s grid operator has already warned that AI facilities cannot be accommodated simply by extending local lines. Taipower said the electricity needs of AI data centres must be weighed against local supply conditions and has steered operators toward central and southern Taiwan, where capacity is less constrained. (digitimes.com) December reporting from Taiwan said Taipower was preparing new rules for AI data centres covering siting near power sources, self-supplied power and tariff reviews, with stricter scrutiny for new or expanded projects above 5 megawatts. (iea.org) Taipower Chairman Tseng Wen-sheng said AI data centres raise grid risk because of their very high load density. ### What does “on-site power” actually mean in practice? On-site power usually means turbines, batteries, and other dedicated equipment installed at or near the campus so the operator can guarantee round-the-clock electricity. (taiwannews.com.tw) DigiTimes said operators are increasingly adopting that model because grid access is becoming the limiting factor for project timing. In the United States, xAI has become a prominent example of the trade-offs. Digital Today reported that xAI data centres are running on natural gas, and Mississippi Today reported on May 11 that xAI had 46 gas turbines tied to facilities serving its Memphis operations. (technews.tw) ### Why is nuclear back in the conversation? A report cited by ANI and other outlets said rising electricity demand from AI data centres is helping revive interest in nuclear power because it offers firm, low-carbon generation alongside energy-security and decarbonisation goals. (digitimes.com) Taiwan’s Energy Administration said on May 14 that the government was carefully assessing nuclear power to strengthen supply stability and system resilience in response to global energy developments and AI-driven electricity demand. (digitaltoday.co.kr) ### What happens next? Taiwan’s government has said new management rules for AI data centres are under discussion, including location, self-generation and tariff arrangements, with a fuller framework expected next year, according to Taipower reporting cited in December. (energy.economictimes.indiatimes.com) Developers in Taiwan and elsewhere are likely to keep pursuing projects that combine computing capacity with dedicated power supply rather than waiting for conventional grid build-outs. (technews.tw) (moeaea.gov.tw)