AI’s next bottleneck: power and compute
The conversation about AI is shifting from model releases to physical limits on power and compute as data‑centre electricity needs and GPU shortages bite. Goldman Sachs projects global data‑centre power demand could surge ~220% by 2030, while reports show GPU rationing, outages and rising prices as demand outstrips supply ( ).
Artificial intelligence now faces a hardware problem: the industry can build bigger models faster than it can secure the electricity and chips to run them. (goldmansachs.com) Goldman Sachs said on March 4 that global data-center power demand is now expected to rise 220% by 2030 from 2023 levels, up from its earlier 175% forecast. The bank said that revision reflects higher artificial-intelligence server shipments and more power-hungry hardware. (goldmansachs.com) Goldman’s updated model points to total data-center electricity use of about 1,350 terawatt-hours by 2030, with roughly 60% of the increase coming from the United States. Benzinga reported that U.S. data-center demand alone could reach about 750 terawatt-hours by 2030. (benzinga.com) A data center is a warehouse full of servers, and artificial-intelligence servers draw far more power because they use graphics processing units, or specialized chips, to train models and generate answers. The International Energy Agency said on April 10, 2025, that global electricity demand from data centers is set to more than double to about 945 terawatt-hours by 2030, with artificial intelligence the biggest driver. (iea.org) The pressure is not only on the grid. The Decoder, citing a Wall Street Journal report, said on April 13 that demand for “agentic” artificial-intelligence tools has triggered a capacity crunch marked by outages, rationing and higher graphics-processing-unit prices. (the-decoder.com) Anthropic’s own status page shows the strain in real time. On April 10, 2026, the company reported elevated errors on requests to Claude models, and on April 11 it disclosed a separate email-login outage. (status.claude.com) OpenAI has also started cutting back on one of its most compute-heavy products. Its help center says the Sora web and app experiences will be discontinued on April 26, 2026, and the Sora application programming interface will end on September 24, 2026. (help.openai.com) The supply side is tight too. Nvidia said in its annual report that “significant mismatches” between supply and demand across its platforms have led to shortages before and could happen again, even as demand for data-center compute keeps rising. (sec.gov) The International Energy Agency said U.S. data centers are on course to account for almost half of the country’s electricity-demand growth by 2030. In that forecast, the United States would use more electricity for processing data in 2030 than for making aluminum, steel, cement and chemicals combined. (iea.org) That is why the artificial-intelligence race now runs through substations, gas turbines, transmission lines and chip packaging plants, not just research labs. The next limit on artificial intelligence is increasingly whether companies can get enough power and compute at the same time. (goldmansachs.com)
