Compute running short, firms ration AI
Firms report they are rationing compute because AI workloads are consuming so much energy that available computing firepower is tightening. The coverage says companies are already deciding which customers, products and models get priority capacity as energy and capacity constraints bite (hindustantimes.com).
Artificial intelligence companies are no longer treating computing power like an unlimited utility; some are deciding which customers and models get scarce capacity first as power and data center limits tighten. (microsoft.com) (cloud.google.com) Training and running artificial intelligence happens inside data centers packed with graphics processing units, the specialized chips that do the math. The International Energy Agency said data centers used about 415 terawatt-hours of electricity in 2024, or roughly 1.5% of global electricity use. (iea.org) That electricity bill is climbing fast. The International Energy Agency said global data-center demand is set to more than double by 2030 to about 945 terawatt-hours, with artificial intelligence the biggest driver of the increase. (iea.org) The squeeze is already showing up in company disclosures. Microsoft told investors on April 30, 2025 that demand for its artificial-intelligence services was growing faster than the data-center capacity it was bringing online, and on July 30, 2025 it said demand still remained higher than supply. (microsoft.com 1) (microsoft.com 2) Alphabet gave a similar warning on October 29, 2025, saying it expected a “tight demand/supply environment” through the fourth quarter of 2025 and into 2026 even as it sped up server deployments and data-center construction. Google Cloud’s own infrastructure blog said in March 2026 that demand for tokens was outpacing the physical supply of chips and power in real time. (abc.xyz) (cloud.google.com) The industry response has been to lock up power as aggressively as it buys chips. OpenAI said in September 2025 that Stargate had nearly 7 gigawatts of planned capacity and more than $400 billion in planned investment over three years, then said in October 2025 that the buildout had grown to more than 8 gigawatts and over $450 billion. (openai.com 1) (openai.com 2) Other companies are doing the same. Meta said in October 2025 that a new El Paso site could scale to 1 gigawatt, and in November 2025 it said a Wisconsin project would become its 30th data center worldwide. (about.fb.com 1) (about.fb.com 2) The power system is becoming part of the competitive map for artificial intelligence. The International Energy Agency said data centers are on course to account for almost half of United States electricity-demand growth through 2030, and that countries able to deliver power “at speed and scale” will be best placed to benefit from artificial intelligence. (iea.org) (iea.org) That is changing how cloud computing is sold. Google Cloud now offers “assured capacity” reservations for artificial-intelligence workloads, and Amazon Web Services in December 2025 launched “AI Factories” built around customer-provided data-center space and power capacity. (cloud.google.com) (aws.amazon.com) The old promise of elastic cloud capacity is giving way to a market where electricity, land, transformers and permits can matter as much as software. In 2026, the companies with the most artificial-intelligence ambition are also becoming some of the biggest buyers of power. (cloud.google.com) (iea.org)
