Data-centre limits matter more than chips
The bottleneck for scaling enterprise AI may be facilities and power, not just semiconductor supply: a recent analysis highlights grid access, cooling, interconnect timelines and permitting as decisive constraints. That means deliverable compute depends on site-level resilience and geopolitics as much as on GPU availability. (youtube.com)
The shortage in artificial intelligence is no longer just chips. In 2026, developers can often buy graphics processors faster than they can secure enough electricity, cooling pipes, and utility approvals to turn those processors on. (iea.org) (datacenterfrontier.com) A modern artificial intelligence data center is a warehouse-sized power plant in reverse: electricity comes in, computation comes out, and the leftover energy leaves as heat. Lawrence Berkeley National Laboratory estimated U.S. data centers used 176 terawatt-hours in 2023 and could reach 325 to 580 terawatt-hours by 2028. (lbl.gov) The chip story changed when one rack started behaving like a tiny supercomputer. NVIDIA says its GB200 NVL72 packs 72 Blackwell graphics processors into one liquid-cooled rack linked by 130 terabytes per second of internal connections. (nvidia.com) That density shifts the bottleneck from silicon to the building around it. Supermicro’s GB200 NVL72 system is sold with direct liquid cooling, cold plates, and a cooling distribution unit because ordinary air cooling struggles when so much compute is concentrated in one cabinet. (supermicro.com) The slow part is usually the grid queue. Lawrence Berkeley National Laboratory says more than 10,300 U.S. power projects were seeking interconnection at the end of 2024, and the median wait from request to commercial operation has stretched to more than four years for projects built in 2018 through 2024. (lbl.gov) That delay is now showing up in the data-center pipeline itself. Wood Mackenzie said planned U.S. data-center capacity additions were cut in half from the third quarter of 2025 to the fourth quarter of 2025 as load-queue problems persisted. (woodmac.com) Developers are reacting by chasing places with faster power, not just bigger labor pools or better fiber lines. Bloom Energy’s 2026 survey-based report says Texas could become the largest U.S. data-center market by 2028 while Northern Virginia slows under grid constraints. (datacenterfrontier.com) Reliability is part of the bottleneck too. A July 10, 2024 voltage event in Northern Virginia caused 60 data centers to disconnect at once, dropping about 1,500 megawatts of load and forcing emergency grid action to avoid wider instability. (belfercenter.org) (datacenterdynamics.com) That is why “available power” now means more than a utility map with spare capacity. It means backup generation, water access, cooling equipment, substation upgrades, and rules that decide who pays when a 75-megawatt or larger customer shows up asking for service. (legiscan.com) (belfercenter.org) The result is that deliverable artificial intelligence compute is becoming a site-selection problem before it becomes a semiconductor problem. The winners are the regions that can permit, power, cool, and stabilize these facilities on real construction timelines, not just the companies that can reserve the next batch of graphics processors. (iea.org) (woodmac.com)
