The US power grid was designed around a load growth model of roughly 1% per year. AI data centers are asking for 20–30% increases in regional load on timelines measured in months, not decades. The mismatch between what utilities can deliver and what the AI buildout requires is not a short-term supply problem. It is a structural gap between infrastructure built for one set of assumptions and demand arriving under an entirely different one.
Deloitte projects US AI data center power demand could jump from 4 GW in 2024 to 123 GW by 2035. The IEA puts global data center power consumption potentially reaching 1,050 TWh by 2026. Nearly 100 GW of new data center capacity is expected to be added globally between 2026 and 2030, doubling total global capacity in four years. Utilities were not consulted on that timeline. They are now being presented with interconnection requests that exceed their capacity to fulfill.
US AI data center power demand: 4 GW (2024) → projected 123 GW by 2035 (Deloitte). IEA: global data center power consumption could reach 1,050 TWh by 2026. Grid interconnection for a new large data center: 4–5 years minimum in most markets; 10 years in congested regions. Nearly 100 GW of new data center capacity expected globally 2026–2030. Hyperscaler capex 2026: $600B+, ~$450B targeting AI infrastructure.
Grid interconnection for a new large data center takes 4–5 years at minimum in most US markets. In congested regions (Northern Virginia, the Pacific Northwest, parts of Texas) that timeline extends to 10 years. A data center developer who breaks ground today will not have the power contract confirmed for a facility that needs to be operational in 18 months. The response has been a proliferation of workarounds: behind-the-meter gas generation, battery storage systems sized to reduce peak grid draw, agreements with industrial operators to acquire existing load. None of these are solutions to the underlying bottleneck. They are deferments.
The cooling implication is direct. A facility running behind-the-meter generation at 30–40% efficiency is burning far more fuel per kWh than a utility-supplied facility running renewable power. The heat rejection load from that inefficiency compounds at the facility level. Operators building against a grid constraint are making thermal architecture decisions they would not make if power were available on schedule.
The structural response that data center operators are being pushed toward is co-investment in grid infrastructure. The largest hyperscalers, Microsoft, Google, Amazon and Meta, are now negotiating directly with utilities on transmission upgrades, substation construction timelines, and power purchase agreements that effectively pre-fund grid capacity. This is a different relationship than the industry had five years ago, when a data center developer filed an interconnection request and waited in the queue like everyone else.
That shift has implications for operators who are not hyperscalers. If Microsoft and Google are negotiating bespoke grid co-investment arrangements that move them to the front of the interconnection queue, mid-tier colocation operators and enterprise data centers are competing for whatever capacity remains. The grid is not a commons that scales smoothly with demand. It is infrastructure with physical limits, regulatory processes, and capital requirements that favor the largest buyers.
Cooling accounts for 30–40% of total data center electricity consumption in a typical air-cooled facility. At a 100 MW facility, that is 30–40 MW of load going purely to moving air and running chillers. A facility that converts to direct-to-chip liquid cooling with dry heat rejection at comparable workload can reduce that cooling load fraction to 5–10%. That 25 to 30 MW reduction is meaningful capacity inside the same power envelope.
For operators in grid-constrained markets, liquid cooling is no longer just a rack density solution. It is a power efficiency argument that buys more compute inside a fixed interconnection limit. An operator with a 100 MW interconnect who converts to liquid cooling frees up 20–25 MW of capacity for additional IT load without requiring a new interconnection agreement. The grid constraint, counterintuitively, is accelerating liquid cooling adoption in markets where air-cooled operators would have otherwise been slow to move.