Why the Chips, Power, and Data Centers Behind AI Can't Keep Up
In late March, heavy Claude users started posting screenshots of something very odd, their five hour usage limits were running out in twenty minutes. Anthropic blamed peak hour demand and blocked third party tools from using its flat rate plans. OpenAI quietly shut down its Sora video platform around the same time as its Codex tool surged past four million developers per week. What looked like routine product decisions were actually the first visible signs of an infrastructure problem that is only going to get harder to solve.
The math is uncomfortable. Running an AI model for users costs real money every single time and it scales directly with usage. If ten times more people use AI tools ten times more heavily, companies then need roughly one hundred times more computing power to keep up. The ratio is about to get worse. Agentic AI systems, that autonomously handle complex multi-step tasks without human input at each stage, can consume ten to one hundred times more compute per session than a standard conversation. As that model of use spreads across business workflows, the demand curve get very steep very fast.
The supply of chips is the first place that pressure is showing up in. Orders for Nvidia GPUs are growing to $1 trillion through 2027, double of that a year ago, with lead times stretching closer to a year. All three makers of high bandwidth memory, which specialized chips AI systems depend on, are sold out for 2026. TSMC fabricates about 90 percent of the world’s most advanced chips. Its CEO said in late 2025 that demand was running three times what the company could produce. The 2nm fabrication lines are booked through 2028, and its next planned U.S. facility is already fully committed before ground has been broken. Building new chip fabrication capacity takes two to four years. The chip shortage will persist.
Power is the next wall. A modern data center can be built in two to three years, but getting reliable electricity to run it takes much longer. Natural gas plants take five to seven years. Nuclear takes ten or more. Even solar, factoring in grid connection timelines, can take two to four years. In February 2025, Dominion Energy Virginia reported requests for 40.2 gigawatts of new data center power connections, nearly double of what had been requested six months earlier. Anthropic has projected that the U.S. AI sector needs at least 50 gigawatts of capacity by 2028. S&P Global estimates that the data center industry needs roughly $200 billion per year globally just to build planned capacity, before taking in the cost of equipment or grid infrastructure buildout.
Governments are treating this as a strategic issue, not just a technology one. The UK published a compute roadmap in 2025, committing up to £2 billion to expand its AI research computing 20x by 2030 and projections that frontier AI demand could grow ten thousand times by 2030. Industry groups have identified that semiconductor supply chains, relying on a small number of companies spread across global networks, remain a significant vulnerability that national AI strategies have barely begun to address. At the Pentagon, the Chief Digital and AI Officer said publicly that compute is the military’s top constraint. His framing was blunt “We’ve handed our warfighters a Ferrari, and my only sleepless nights come from making sure we never, ever run out of the high octane fuel that they need, which is compute.”
The leaders of OpenAI, Google, Meta, Amazon, and Microsoft have all said publicly they cannot get chips fast enough. NVIDIA’s CEO put it plainly by stating that doubling compute access for its top customers would increase their revenues fourfold. That demand is not slowing. AI is moving, or has moved, into coding, healthcare, legal work, finance, and military operations. Every one of those use cases adds to the load. The infrastructure to support that load is measured in years to build. The demand is not waiting.
References
Sanders, J., Egan, J., & Madigan, R. (2026, May 7). American AI companies can’t get enough chips. Center for a New American Security. https://www.cnas.org/publications/reports/american-ai-companies-cant-get-enough-chips
Harper, J. (2026, May 7). DOD planning to address compute ‘bottleneck’ that could hinder AI proliferation. Defense Scoop. https://defensescoop.com/2026/05/07/dod-planning-to-address-compute-bottleneck-ai-proliferation/
Aliaga, S. (2026, April 17). Is AI running out of compute? J.P. Morgan Asset Management. https://am.jpmorgan.com/us/en/asset-management/adv/insights/market-insights/market-updates/on-the-minds-of-investors/is-ai-running-out-of-compute/
Department for Science, Innovation and Technology. (2025, July 17). UK compute roadmap. UK Government. https://www.gov.uk/government/publications/uk-compute-roadmap/uk-compute-roadmap
Béchard, D. E. (2026, May 1). What is the AI compute crunch, and why are AI tools hitting usage limits? Scientific American. https://www.scientificamerican.com/article/what-is-the-ai-compute-crunch-and-why-are-ai-tools-hitting-usage-limits/
Morgan, K., & Partridge, B. (2025, December 2). Global AI power demand: Challenges and opportunities. S&P Global. https://www.spglobal.com/en/research-insights/special-reports/look-forward/data-center-frontiers/global-ai-power-demand-challenges-opportunities
Foster, L. (2025, January 13). Compute infrastructure and the AI opportunities action plan. techUK. https://www.techuk.org/resource/compute-infrastructure-and-the-ai-opportunities-action-plan.html