Compute and energy strain rising
Surging enterprise AI demand is triggering compute and energy constraints that are forcing some providers to ration access and raise prices, according to industry reporting. The trend suggests buyers should scrutinise not just model performance but the underlying service economics — especially power and data‑center limits (enterpriseai.economictimes.indiatimes.com).
Artificial intelligence demand is running into two hard limits at once: the number of chips in data centers and the electricity needed to run them. Microsoft said on April 30, 2025, that demand for its artificial intelligence services was growing faster than the company could bring new capacity online. (microsoft.com) That squeeze is showing up in how providers sell access. OpenAI’s current application programming interface pricing page offers lower-cost batch processing at 50% off standard rates, “Flex processing” with slower responses and occasional unavailability, and separate enterprise options for reserved capacity. (openai.com) Amazon Web Services has pushed a similar split between premium and discounted usage. Its Bedrock pricing page advertises batch inference at 50% lower prices than on-demand inference, while Anthropic said in November 2024 that Amazon Web Services had become its “primary cloud and training partner” as Amazon expanded its investment to $8 billion. (aws.amazon.com) (anthropic.com) The bottleneck is not only chips. Google said in its 2025 environmental report that its total data-center electricity consumption rose 27% in 2024, up from 17% growth in the prior year, as artificial intelligence use expanded across its products and infrastructure. (smartenergydecisions.com) Utilities are now treating data-center demand as a grid-planning problem, not just a real-estate one. Dominion Energy Virginia told regulators in a February 16, 2026 filing that it had about 70 gigawatts of large-load requests in its interconnection queue, versus a 24.7-gigawatt all-time system peak in January 2025. (datacenterdynamics.com) Dominion said about 25 gigawatts of those requests had projected connection dates through the end of 2031, while another 45 gigawatts remained under study. The utility proposed a formal queue for projects of roughly 100 megawatts or more, with individual requests capped at 300 megawatts. (datacenterdynamics.com) That changes what buyers have to evaluate when they shop for artificial intelligence services. A model can look cheap on a benchmark and still become expensive if the provider charges more for priority access, limits peak-time throughput, or needs customers to reserve capacity in advance. (openai.com) (aws.amazon.com) The industry response has been to redesign the stack from the bottom up. Microsoft has been expanding custom and partner hardware for supercomputing, Google has been emphasizing more efficient Tensor Processing Units, and Anthropic has been working with Amazon Web Services on Trainium chips and low-level software to extract more performance from the same power budget. (microsoft.com) (blog.google) (anthropic.com) The result is that artificial intelligence pricing is starting to look more like airline seating than commodity cloud computing: standard, batch, flex, and reserved tiers tied to scarce infrastructure. As long as data-center buildouts and power hookups lag demand, access itself will remain part of the product. (openai.com) (datacenterdynamics.com)
