AI build-out hitting a compute and power wall
Reports say AI adoption is straining compute and energy capacity so severely that firms are rationing products and delaying launches. Journalists note rising costs and queueing pressures that shift the bottleneck from model design to physical power, capacity and operational allocation. ( )
Artificial intelligence companies are running into a basic limit: they can design new models faster than they can secure chips, data centers and electricity to run them. (ibtimes.com) Artificial intelligence systems are built and served from data centers, which are warehouses of servers, networking gear and cooling equipment. The International Energy Agency estimates data centers used about 415 terawatt-hours of electricity in 2024, or roughly 1.5% of global power consumption. (iea.org) Inside a modern data center, servers do most of the work and draw most of the power. The International Energy Agency says servers account for around 60% of electricity demand in modern facilities, while cooling ranges from about 7% in efficient hyperscale sites to more than 30% in less efficient enterprise sites. (iea.org) The squeeze is showing up in products people already use. International Business Times, citing a Wall Street Journal investigation, reported on April 13 that Anthropic tightened Claude usage during peak periods, CoreWeave pushed customers toward longer contracts, and OpenAI has had to make product-allocation decisions around limited compute. (ibtimes.com) Anthropic has also tried to shift demand away from busy hours instead of only adding more hardware. Its help center said that from March 13 to March 28, 2026, eligible Claude users got double usage during off-peak weekday hours, while peak-hour limits stayed the same. (support.claude.com) The bottleneck is no longer just semiconductor supply. Microsoft Chief Executive Satya Nadella said in a late-2025 interview that the “biggest issue” was power, not a shortage of compute on paper, because the company could not build out sites near electricity fast enough. (ft.com) Nvidia’s numbers show why the queue keeps growing. The company said on November 19, 2025 that quarterly data-center revenue hit a record $51.2 billion, up 66% from a year earlier, and Chief Executive Jensen Huang said “cloud GPUs are sold out.” (nvidianews.nvidia.com) The power side is tightening at the same time. Reuters reported on April 9 that big technology companies are putting money into next-generation nuclear projects to lock in future electricity supply for artificial-intelligence data centers. (reuters.com) Utilities and local grids are warning that new connections can take years in the busiest markets. A Reuters report distributed on April 14 said Dominion Energy had warned that some new data-center hookups in Northern Virginia could face multi-year delays because the grid cannot keep up. (reuters.com) The costs are spilling beyond the technology sector into electricity politics. CNBC reported on March 13 that U.S. residential power prices rose from 12.76 cents per kilowatt-hour in 2020 to 17.44 cents in February 2026, as communities and regulators argued over whether data-center growth is raising bills. (cnbc.com) Companies are still spending as if demand will keep climbing. Reuters reported in February that OpenAI was targeting about $600 billion in compute spending through 2030, a sign that the race has shifted from writing better code to securing enough physical capacity to turn that code on. (economictimes.indiatimes.com)
