Data‑centre limits: power, space, geopolitics
A recent video framed the real cost of data centres as a three‑part problem — infrastructure buildout, power availability, and geopolitical exposure — rather than a pure chip‑supply issue. That matters because even if GPUs are available, utility interconnection, cooling, and location risks can still bottleneck deployments; and researchers warning of active hack‑for‑hire campaigns underline that infrastructure growth also expands attack surfaces. Treating compute as a systems problem — not just a component play — is becoming essential for scaling AI. (youtube.com/watch?v=UGkyEYrtpHI (techcrunch.com))
The bottleneck in artificial intelligence is no longer just the chip. A modern data center can have servers ready to ship and still sit idle for months or years because the missing piece is electricity, not silicon. (lbl.gov) A data center is just a warehouse full of computers, and every computer turns electricity into heat. The more useful work those computers do, the more power lines, transformers, and cooling equipment the building needs around them. (iea.org) (lbl.gov) Lawrence Berkeley National Laboratory says United States data centers used 176 terawatt-hours of electricity in 2023, up from 58 terawatt-hours in 2014. The same report projects 325 to 580 terawatt-hours by 2028, which would put data centers at 6.7% to 12% of total United States electricity use. (lbl.gov) That is why “interconnection” has become a key word in this industry. Interconnection is the utility approval process that links a big new customer to the grid, and Bloom Energy’s 2026 survey says those timelines are widening as gigawatt-scale campuses move from plan to reality. (bloomenergy.com) A gigawatt is the scale of a large power plant, not a normal office park. Bloom Energy says about one in five data-center campuses are expected to exceed one gigawatt by 2030, which turns site selection into an energy hunt rather than a real-estate search. (bloomenergy.com) Cooling is the second limit, because the same rack of servers that earns revenue can also overheat like a car engine without a radiator. Bloom Energy says cooling capacity and water availability are rising alongside power delivery as critical constraints, and the Congressional Research Service says a 100-megawatt data center can directly use about as much water as 2,600 households depending on cooling design. (bloomenergy.com) (congress.gov) Location starts to matter in a different way once a company needs land, substations, fiber lines, and cooling water all at once. Bloom Energy says capital is concentrating in power-advantaged regions and expects Texas to become the leading United States data-center market within three years as legacy markets like California and Oregon lose relative share. (bloomenergy.com) The geography problem is also a political problem, because the most valuable compute clusters become fixed targets. TechCrunch reported on April 8, 2026 that researchers tied a hack-for-hire campaign called BITTER to phishing for Apple iCloud backups, Signal accounts, and Android device access across the Middle East and North Africa. (techcrunch.com) That story was about phones and cloud accounts, but the lesson travels upward to infrastructure. Every new data-center campus adds contractors, remote management tools, utility links, and customer networks, which means the attack surface grows with the building footprint. (techcrunch.com) (cisa.gov) So the real constraint on artificial-intelligence buildout is a stack of dependencies. You need the graphics processing units, then the building, then the grid connection, then the cooling system, then the local permits, and then the security posture to keep the whole thing from becoming a very expensive single point of failure. (lbl.gov) (bloomenergy.com) (cisa.gov)
