The U.S. Energy Information Administration's Annual Energy Outlook 2026 puts numbers on the data center electricity trajectory that most operators already sense but rarely see laid out at a 25-year horizon. Server electricity in data centers accounted for roughly 7 percent of all U.S. commercial sector electricity in 2025. By 2050, the EIA projects that share reaches 22 to 33 percent, depending on the demand scenario. The absolute numbers: 446 to 818 billion kilowatt-hours annually. The high-demand scenario standalone data center figure alone is 581 billion kWh.
These are not marginal growth projections. The commercial sector electricity intensity, which peaked historically in 2003 at 14.9 kWh per square foot, is projected to exceed that 2003 peak again for the first time in 2031 or 2032, driven primarily by data center density growth. The last time the sector hit that intensity level, most racks were drawing under 5 kW. The next time will be on the back of hardware drawing 100 to 500 kW.
The EIA report includes a finding that the cooling industry should read carefully. Space cooling energy intensity in data center floorspace runs approximately 2.9 times higher than in non-data center commercial buildings. Not 10 percent higher. Not double. Nearly triple. That ratio is a function of heat density, not facility design quality. The building above the raised floor is irrelevant. The density inside the racks is what drives the cooling load.
In the high-demand scenario, the EIA projects cooling electricity consumption in 2050 running 84 billion kWh above the baseline case. That 84 BkWh difference is entirely a function of what cooling architecture those facilities deploy. Facilities running air cooling at 40 kW per rack are not building toward 500 kW AI density. The cooling architecture question is not being deferred. It is being answered in every new facility that specifies liquid cooling in the initial build versus planning to retrofit later.
The baseline scenario assumes server efficiency improves 10 percent every three years starting in 2040. The high-demand scenario does not apply those improvements as aggressively. Neither scenario fully prices in the thermal management efficiency gains available from optimized liquid cooling. The UIUC cold plate research published this week targets a PUE of 1.011 using topology-optimized copper manufacturing. That is not in the EIA model. The 84 BkWh cooling gap between scenarios could narrow substantially if the liquid cooling adoption rate accelerates beyond what historical retrofit timelines suggest.
The EIA report is a forecast, not a plan. But it names the structural reality clearly: data centers are becoming the dominant electricity load in U.S. commercial buildings, and the cooling component of that load scales with the compute density decisions operators and hyperscalers make in the next five years. The facilities that are being permitted and designed today will still be operating in 2050. The cooling architecture locked into those builds is locked in for the duration.
At the same time, a separate North Carolina State analysis published this week projects residential electricity costs in Virginia rising up to 57 percent by 2030 as data center load saturates the grid. The EIA long-range projection and the near-term rate impact study are describing the same forcing function from two different time horizons.