US AI Data-Center Expansion Faces Power, Grid and Supply-Chain Constraints
DataNews Editorial Team
Power availability, grid interconnection capacity and component supply are emerging as central constraints on US AI data-center expansion. Federal agencies are streamlining selected permitting and interconnection processes, while major cloud providers are pursuing nuclear agreements, gas generation and grid-investment commitments—but these measures are not expected to provide immediate relief.
- US AI data-center delivery is increasingly dependent on power availability and grid interconnection, not only accelerator supply.
- Goldman Sachs projected US data-center electricity demand to reach 66 GW in 2027, up from 31 GW in 2025.
- Federal streamlining applies to AI training and inference data centers with loads above 100 MW.
- The approximately 9.8 GW of disclosed nuclear commitments is not commissioned or guaranteed capacity for data centers.
- PJM's 2026/2027 capacity auction reached the FERC-approved cap of $329.17/MW-day.
US AI infrastructure build-outs are increasingly constrained by electricity supply, transmission capacity and the timing of grid interconnection, alongside shortages of high-performance chips, advanced packaging and HBM memory. The source describes demand growth outpacing additions to generation and network capacity, pushing power procurement and interconnection into the critical path for data-center delivery.
Global data-center electricity demand rose 17% in 2025, according to figures cited by the source, while electricity consumption at dedicated AI data centers increased 50%. Goldman Sachs projected US data-center power demand to rise from 31 GW in 2025 to 41 GW in 2026 and 66 GW in 2027. These figures describe electricity demand rather than commissioned data-center capacity.
Grid reliability and capacity-market pricing underscore the pressure. The US Department of Energy warned that, under existing retirement and new-build plans, most power-system regions could face unacceptable reliability risks within five years. In PJM, the capacity auction for the 2026/2027 delivery year reached the FERC-approved price cap of $329.17/MW-day. Separately, the source reports that the PJM capacity-auction price for the 2025/2026 delivery year was nearly nine times the prior delivery year's level.
Federal measures include streamlined permitting for AI training and inference data centers with loads above 100 MW, as well as processes intended to support joint interconnection applications for load and generation. The response also includes opening federal sites and supporting energy projects, according to the source.
Microsoft, Google, Amazon and other companies are using nuclear agreements, gas generation and commitments to fund required grid infrastructure as part of their response. As of May 2026, four major cloud providers had disclosed nuclear-project commitments totaling about 9.8 GW. That figure represents disclosed commitments, not operating or guaranteed data-center supply. The source gives no project-level breakdown or a complete list of the four providers included in the total.
The timing gap remains material. The source's general assessment is that nuclear restarts and new nuclear-build projects are not expected to enter service before 2027, while small modular reactors have not yet entered commercial operation. In March 2026, seven major AI technology companies signed a White House-brokered commitment to help fund required grid infrastructure so that costs would not be shifted to household consumers.
For operators, investors and developers, the implication is a growing competition for interconnected power rather than only for sites and compute equipment. Projects with substantiated power supply and interconnection arrangements could become more valuable, while some demand could shift toward locations with available generation or toward models that colocate generation with large AI loads. These are market implications, not evidence that such capacity is already available.
For US data-center development, access to reliable power and timely grid interconnection is becoming as consequential as access to AI accelerators. This could extend delivery timelines, increase the value of projects with credible power pathways, and support demand for dedicated generation and transmission infrastructure; announced nuclear capacity and regulatory measures should not be treated as power already available to new campuses.
- — / October 16, 2026 · ChineseSource →
This article was prepared by the DataNews editorial team based on the sources listed above.