McKinsey: $5.2T AI spend by 2030

McKinsey says data centers built for AI workloads will need $5.2 trillion in capital spending by 2030. That is the bulk of a projected $6.7 trillion global bill for compute infrastructure. (mckinsey.com) The firm published that estimate in an April 28, 2025 report on the “cost of compute,” splitting expected spending between $5.2 trillion for AI-ready capacity and $1.5 trillion for traditional information-technology workloads. (mckinsey.com) McKinsey’s October 29, 2024 data-center analysis said global demand for capacity could grow 19% to 22% a year from 2023 through 2030, reaching 171 to 219 gigawatts from 60 gigawatts. In its midpoint scenario, about 70% of that demand comes from advanced AI workloads. (mckinsey.com) A data center is the warehouse behind cloud software and AI models: rows of servers, networking gear, power equipment and cooling systems running around the clock. AI changes the math because graphics processors and accelerators draw more power and produce more heat than older corporate computing jobs. (mckinsey.com 1) (mckinsey.com 2) McKinsey said mechanical and electrical systems alone could attract more than $250 billion of capital spending by 2030. The firm also warned that the United States could still face a supply deficit of more than 15 gigawatts by 2030 even if all currently known projects arrive on time. (mckinsey.com) That spending wave is already showing up in hyperscaler budgets. Dgtl Infra estimated annualized 2024 capital spending for Amazon Web Services, Microsoft, Google, Meta and Oracle at $166 billion, with a forward run rate of about $185 billion. (dgtlinfra.com) The bottlenecks are not only land and electricity. TSMC says its CoWoS advanced-packaging platform is built for high-performance computing chips that combine logic with high-bandwidth memory, and the company told investors in April 2025 it was working to double CoWoS capacity in 2025. (tsmc.com) (investor.tsmc.com) High-bandwidth memory, or HBM, is the dense memory stacked next to AI chips so models can move data faster. SK hynix said January 28, 2026 that record 2025 results were driven by AI memory and HBM products, and its January 2026 market outlook cited a Bank of America estimate for a $54.6 billion HBM market in 2026. (news.skhynix.com 1) (news.skhynix.com 2) ASML, which makes the extreme ultraviolet lithography tools used to print advanced chips, said in its 2025 annual report that AI drove a positive shift in wafer demand and stronger spending on advanced logic and dynamic random-access memory. The company’s January 28, 2026 investor-call transcript said demand for its advanced products, especially EUV systems, had been supported by AI in recent months. (asml.com 1) (asml.com 2) The result is a supply chain stretching from utilities and cooling contractors to memory makers and lithography vendors, all trying to catch up with a data-center buildout McKinsey pegs in the trillions. (mckinsey.com)