McKinsey: $5.2T AI data‑center capex

McKinsey says the global race to build data centers will require about $6.7 trillion by 2030, with $5.2 trillion tied to facilities built for artificial-intelligence workloads. (mckinsey.com) That estimate comes from McKinsey’s April 28, 2025 report on compute demand, which split projected spending into AI-capable sites and another $1.5 trillion for traditional information-technology workloads. (mckinsey.com) In March 2026, McKinsey sharpened the point: the bottleneck is no longer just chips. It said power equipment and cooling systems have become “pacing items” in the build-out. (mckinsey.com) A data center is the physical plant behind cloud and AI services: buildings full of servers, networking gear, transformers, switchgear, and cooling equipment that keep the machines running. AI systems need more of all of it because they pack in denser, hotter hardware than older enterprise applications. (mckinsey.com 1) (mckinsey.com 2) McKinsey’s August 2025 infrastructure analysis put non-IT data-center spending alone at more than $1.7 trillion through 2030. That bucket covers the construction, electrical, mechanical, and cooling systems outside the servers themselves. (mckinsey.com 1) (mckinsey.com 2) The power load is climbing with it. McKinsey said data-center electricity demand could reach 1,400 terawatt-hours globally by 2030, about 4% of total world power demand. (mckinsey.com) In the United States, McKinsey said data-center power capacity could rise from about 30 gigawatts in 2025 to more than 90 gigawatts by 2030. It also said hyperscalers are expected to capture about 70% of that forecast U.S. capacity. (mckinsey.com) That helps explain why investors have shifted from focusing only on Nvidia chips and cloud tenants to also tracking electrical and thermal suppliers. McKinsey said traditional equipment makers are seeing record demand but often cannot deliver fast enough or redesign products around rapidly changing chip road maps. (mckinsey.com) The market has also started treating specialized cloud providers as infrastructure proxies. Nebius said in September 2025 that it signed a multiyear Microsoft contract for dedicated capacity from its Vineland, New Jersey data center, and in March 2026 it said Meta signed a separate agreement worth up to about $27 billion. (nebius.com 1) (nebius.com 2) Those two disclosed deals are why analysts now talk about Nebius having roughly $46 billion of contracted AI-cloud business, though that figure is an inference from the announced Microsoft and Meta agreements rather than a single McKinsey number. (nebius.com) (nebius.com) McKinsey’s core message is narrower than the stock trade around it: AI demand is forcing a capital cycle that runs through land, grid access, transformers, cooling loops, and construction schedules as much as through semiconductors. (mckinsey.com) (mckinsey.com) If McKinsey’s forecast is even close, the next phase of the AI boom will be measured less by chatbot launches than by who can secure power, cooling, and construction capacity first. (mckinsey.com) (mckinsey.com)