The dirty secret: data centre limits
- A recent video highlights the non-obvious constraints on AI data-centre expansion like power, cooling, and networking. - The piece argues GPU supply is only one part of a broader stack problem that includes long-term reservation economics. - That viewpoint reframes vendor selection toward reserved capacity, hybrid execution, and infrastructure resilience as core procurement questions (youtube.com).
The bottleneck in artificial intelligence build-outs is no longer just chips; it is whether operators can secure enough power, cooling, and network capacity to run them. (mckinsey.com) A data center is the warehouse behind cloud computing: rows of servers drawing electricity, shedding heat, and moving data across switches and fiber. New artificial intelligence systems pack far more computing into each rack, pushing facilities from ordinary air-cooling toward liquid systems and denser electrical gear. (techrepublic.com) Nvidia’s GB200 NVL72 shows the shift in one product. The rack links 72 Blackwell graphics processors and 36 Grace central processors in a liquid-cooled system, and Hewlett Packard Enterprise says one rack draws 132 kilowatts, with 115 kilowatts liquid cooled and 17 kilowatts air cooled. (nvidia.com, buy.hpe.com) Networking is a second limit, because these systems only work at full speed if thousands of chips can talk to each other quickly and reliably. Google says a TPU v5p Pod contains 8,960 chips, with 1,200 gigabytes per second of inter-chip bandwidth per chip, and it added link-resiliency features to route around optical faults. (docs.cloud.google.com) Power is the hardest constraint because it arrives on utility timelines, not software timelines. Deloitte said in June 2025 that United States power demand from artificial intelligence data centers could rise to 123 gigawatts by 2035 from 4 gigawatts in 2024, and that some grid interconnection requests already face waits of seven years. (deloitte.com) That squeeze is already changing the market for space. JLL said North American colocation vacancy fell to 2.3% by mid-2025 and 73% of capacity under construction was already preleased, while CBRE said global vacancy averaged 6.6% in the first quarter of 2025 and new projects in some hubs were slipping to 2027 or later because of power shortages. (jll.com, cbre.com) Northern Virginia, the biggest U.S. data-center market, shows how reservation economics now work. Dominion Energy told regulators in February 2026 that data-center power requests had reached nearly 70,000 megawatts, almost triple its January 2025 system peak of 24,678 megawatts. (virginiabusiness.com) That is why buying “access” can matter as much as buying hardware. McKinsey said hyperscalers are expected to capture about 70% of forecast U.S. capacity through owned and leased sites, and Anthropic advertised a six-figure role on April 23, 2026 focused on negotiating data-center deals for its Europe expansion. (mckinsey.com, cnbc.com) The procurement question has widened with that shift. A buyer now has to ask how much power is reserved, what cooling loop is installed, how network failures are handled, and which workloads can run in a public cloud, a leased colocation hall, or an in-house cluster. (docs.cloud.google.com, buy.hpe.com, mckinsey.com) The short version is that the “chip shortage” story now misses the larger constraint. The real race is for powered, cooled, connected capacity that can stay online when thousands of expensive processors arrive. (deloitte.com, jll.com)
