Data centres: more than GPUs
A recent video argues AI data‑centre constraints are multi‑dimensional — not just GPU shortages but capital, power, cooling and site readiness are gating capacity. The piece recommends stress‑testing deployments against utility commitments and longer lead times rather than assuming hardware orders equal immediate capacity. (youtube.com)
Artificial intelligence capacity is being held back by power lines, cooling pipes, land permits and financing schedules as much as by graphics processing unit supply. (iea.org / energy.gov) A data centre is a warehouse of servers, networking gear and backup equipment; the chips do the computing, but the building also needs electricity, transformers, chillers or liquid loops, and a utility connection before any model can run. The International Energy Agency said on April 10 that data-centre electricity demand is set to more than double to about 945 terawatt-hours by 2030. (iea.org / spglobal.com) The United States Department of Energy said in a July 30, 2024 report that hyperscale facilities are now requesting 300 megawatts to 1,000 megawatts of power, with lead times of one to three years, stretching local grids. In plain terms, ordering servers is like buying factory machines before the substation and water system are built. (energy.gov) That bottleneck is showing up in the real-estate market. Jones Lang LaSalle said North America data-centre vacancy stayed at 1% for a second straight year at the end of 2025, even with a 35-gigawatt construction pipeline. (jll.com) The hardware itself is also changing what buildings need. Nvidia says its GB200 NVL72 system packs 72 Blackwell graphics processing units and 36 Grace central processing units into a rack-scale, liquid-cooled design, and Hewlett Packard Enterprise says one such rack consumes 132 kilowatts. (nvidia.com / hpe.com) Cooling is no longer a side issue. Nvidia said in March 2025 that cooling has historically accounted for up to 40% of a data centre’s electricity use, and it pitched liquid cooling as a way to handle denser systems that air cooling struggles to support. (blogs.nvidia.com) Power access is becoming the deciding factor in where projects get built. Bloom Energy said in its January 2026 survey that power availability has moved from a planning issue to a hard limit on data-centre growth, with operators reporting longer interconnection timelines and more interest in on-site generation. (bloomenergy.com) Northern Virginia shows how that plays out on the ground. Dominion Energy has warned that proposed facilities there could require more electricity than it can deliver before the end of the decade, and earlier reporting said some projects were facing waits of up to seven years for hookups. (msn.com / energyconnects.com) Money is not the easy part either. Jones Lang LaSalle said North America could see more than 100 gigawatts of capacity break ground or deliver between 2025 and 2030, representing roughly $1 trillion of development, while utilities are planning heavy grid spending to keep up with load growth. (realestatedaily-news.com / cbsnews.com) The practical takeaway is simple: a graphics processing unit order is not the same thing as live capacity. In 2026, the winners are the operators that can show signed utility commitments, cooling designs, land control and realistic construction dates before the servers arrive. (energy.gov / bloomenergy.com)
