Generative AI and its agentic successors demand data centers that gulp electricity like small cities. Racks now push 100 kilowatts or more. Air cooling buckles under the heat. Operators scramble for transformers delayed years in supply chains. At Data Center World in Washington, DC, executives laid bare the crunch. Phill Lawson-Shanks, CEO of Aligned Data Centers, described a project upended by AI’s appetite: “We needed to restructure everything to be able to introduce enough cooling and find more power from the grid.” (TechRepublic)
Power shortages halt half of planned U.S. builds. Bloomberg reports close to 50% face delays or cancellation due to scarce electrical gear from China, even as Alphabet, Amazon, Meta, and Microsoft pledge over $650 billion in 2026 expansions. Transformers wait three to five years. Switchgear vanishes. Grid queues stretch seven years in spots like ERCOT. Utilities slam brakes on connections, per Wood Mackenzie.
But demand surges on. Omdia analysts Shen Wang and Alan Howard see AI factories bulldozing ahead. “People need AI so large AI factories will move forward somehow,” Wang said. Howard noted unprecedented construction: projections climb yearly. Aligned scouts “stranded power” sites, buys gear 2.5 years early, plans 88 buildings—all leased.
Cooling flips from afterthought to mandate. Rack densities leap from tens to hundreds of kilowatts. Liquid systems match air-cooled capacity in 2025, double it by 2026’s end. Cold plate shipments explode: eight million in 2025 to 356 million by 2030. Vertiv’s Scott Armul pushes simulation: “We can now simulate the whole environment, so we know that we need to adjust liquid cooling valves and alter CDU set points to optimize facility cooling and efficiency.” (TechRepublic)
Average racks hit 16 kilowatts today, up from 6.1 nine years back. AI craves 30-40 or higher. Only 20% of operators handle 50-70 kilowatts per rack. Workloads shift: AI from 15% to 40% by 2030. (Data Center World)
Global electricity for data centers reached 415 terawatt-hours in 2024—1.5% of total, growing 12% yearly versus 3% overall. Projections hit 1,050 TWh by 2026, rivaling Japan or Russia. Goldman Sachs sees 165% rise by 2030. IEA pins AI as prime driver, adding hundreds of TWh. In 2024, 40% came from natural gas, 24% renewables, 20% nuclear. (Brookings Institution)
Water strains mount too. U.S. centers slurped 17 billion gallons in 2023, mostly hyperscalers. Projections: 16-33 billion annually by 2028. One prompt: half a liter. Yet innovations beckon. Waste heat could warm towns in Appalachia or fuel greenhouses. Europe eyes 221 TWh yearly recovery—12% of district heating. London alone: enough for 500,000 homes. (World Economic Forum)
Operators pivot. CoreSite’s 2026 outlook flags on-site power, natural gas, liquid advances. Accenture calls it unprecedented momentum for AI, cloud, edge. JLL deems it history’s largest infrastructure supercycle, reshaping power, tech, real estate. (CoreSite)
Behind-the-meter generation skips grid lines. Natural gas turbines power sites in 12-24 months, not 3-5 years. Nuclear revives: Microsoft restarts Three Mile Island for 835 megawatts. Amazon buys Talen nuclear campus. Google inks small modular reactors with Kairos. Meta hunts sites. Hyperscalers bid fiercely for baseload.
Liquid cooling dominates. Cold plates for chips. Immersion for racks. Vendors like LiquidStack slash water, cool megawatts. China’s underwater servers tap ocean chill. Sandia tests submersion sans fans. China boosts efficiency 80%. NVIDIA’s Vera Rubin racks demand it—120 kilowatts, fully liquid-cooled. Microsoft deploys first NVL72.
Supply chains strain. CDU lead times triple. Vertiv’s $50 million Ohio plant waits till 2027. Pipe fittings bottleneck the chain. 83% of experts say cooling tech lags AI needs. Yet hyperscalers push: Microsoft, Google, Meta, Amazon signed power deals, secured land, ordered GPUs. Racks idle on cooling.
And heat islands emerge. Facilities spike local temps, hike bills, strain water, drone noise, pollute air. Cambridge urges chip tweaks, hybrid cooling: liquid at chip, air overall. Software efficiencies too.
China experiments boldly. Submerged servers cut energy, water. Theoretical only? No—operational tests.
Alberta queues 30 AI projects. Wonder Valley’s $70 billion park taps gas, geothermal for 1.4 gigawatts phase one. Synapse’s one-gigawatt bid falters on community pushback.
Gartner warns: power shortages curb 40% of centers by 2027. EnkiAI calls grid strain top barrier. Yet $582 billion AI infra spend in 2026, up 19%.
Operators integrate power, cooling, compute from start. Standardized blocks, defined interfaces between CDUs, thermal cycles. Scalable modules for gigawatt campuses.
So grids creak. But AI marches. Nuclear restarts. Gas surges. Liquids flow. Waste heat repurposes. Space beckons—T to the fourth in vacuum. Constraints bind. Solutions forge ahead. Industry insiders watch electrons and pipes, not just silicon.
WebProNews is an iEntry Publication