The artificial intelligence gold rush is leading to a severe shortage of computing power, impacting product development and reliability. Demand for computational resources is outstripping companies' ability to access essential resources and build new infrastructure. For instance, hourly rental prices for Nvidia's most advanced Blackwell generation chips have surged by 48% in just two months, from $2.75 to $4.08, according to the Ornn Compute Price Index. This capacity crunch has prompted cloud infrastructure companies like Vultr to describe the situation as unprecedented, noting that lead times for data center construction are extended and power availability through 2026 is already fully allocated. Even major AI players like OpenAI are struggling, with CFO Sarah Friar admitting to making "very tough trades" and scrapping projects, such as the Sora video-generation app, to reallocate computing resources.
AI companies are already rationing their offerings and services, which is a significant warning sign for a boom that relies heavily on rapid adoption. Anthropic, known for its Claude chatbots, has experienced frequent outages and has begun to limit user token usage during peak hours, specifically from 5 a.m. to 11 a.m. Pacific Time on weekdays. This is despite the company's explosive growth, with its annual run rate ballooning from $9 billion at the end of 2025 to $30 billion just four months later. Customers have voiced complaints about hitting these limits too quickly, highlighting the strain on available compute resources. Token usage in OpenAI's API, primarily used by enterprise clients, soared from 6 billion minutes in October to 15 billion minutes by late March.
This burgeoning demand for AI is pushing electricity grids to their limits. PJM Interconnection LLC, the largest U.S. power grid serving 67 million people across 13 states, has introduced a new "capacity advisory" to warn customers of potential electricity shortages even outside of extreme weather conditions. The situation is particularly acute in Virginia, where data centers are projected to consume up to 57% of the state's electricity by 2030, significantly up from the current 4% to 5% share of U.S. electricity data centers consume. The Electric Power Research Institute projects that data centers could consume as much as 17% of U.S. electricity by 2030. Other states, including Arizona, Indiana, Iowa, Nebraska, Nevada, Oregon, and Wyoming, could see similar levels of data center power consumption by 2030. This enormous power demand, coupled with the risk of data centers suddenly disconnecting from the grid, poses new threats to grid stability.
Incidents in Virginia, where data centers using enough electricity for over a million homes simultaneously switched to backup power when high-voltage lines malfunctioned, illustrate the gravity of the situation. In February 2025, 40 data centers, and in July 2024, 70 data centers, simultaneously disconnected from the grid, causing rapid plunges in demand that forced grid operators to take emergency action. While these incidents didn't cause a crisis for PJM, they raised significant concern, particularly regarding the potential impact if 3,000 to 5,000 megawatts of demand were suddenly lost. The North American Electric Reliability Corporation (NERC) considers this an emerging and critical risk, working with tech companies to prevent such sudden drops in demand. Utility companies like Dominion Energy are collaborating with tech firms to find solutions that allow data centers to maintain grid connection during brief power disturbances, instead of automatically switching to backups.
