US AI data center power gap

Global data center power capacity is projected to grow from 100.5 GW in 2024 to 211.1 GW by 2027, driven by AI servers that accounted for roughly one quarter of global data center power demand in 2025 and could exceed 40% by 2027, according to TrendForce. The Electric Power Research Institute estimates US data centers could consume 9% to 17% of all US electricity by 2030, up from approximately 4% to 5% today, with TrendForce projecting the US supply-demand gap could surpass 170 GW by 2030 and the global gap could reach 268 GW.

Grid connection delays are pushing AI data center projects toward portable natural gas turbine generators, but turbine blade production takes 60 to 90 weeks per batch and delivery queues extend to 2030. SpaceX confirmed plans to build a foundry to cast blades and vanes for natural gas power turbines, with Musk saying the move could bring turbines online up to 18 months faster. Howmet Aerospace and Berkshire Hathaway's Precision Castparts are the two dominant turbine blade producers, with Howmet holding roughly half of the global market share for natural gas turbine components. Howmet stock fell more than 7% following the announcement before partially recovering. The Financial Times reported that if established suppliers believe SpaceX can credibly cast those parts itself, they have a stronger reason to expand capacity preemptively, and the competitive threat may ease the bottleneck even if SpaceX never manufactures these parts at scale.

On the demand-management side, Emerald AI and NVIDIA demonstrated an automated flexible-load program with Silicon Valley Power in which NVIDIA's DSX Flex software receives grid signals, including load-shedding requests and pricing signals, and automatically adjusts workload priority, allowing critical AI jobs to continue while lower-priority jobs pause and resume after grid pressure eases. Redwood Materials is building off-grid power solutions for AI factories using 100% recycled electric vehicle batteries combined with an AI intelligence layer running on the NVIDIA Blackwell platform, with the company stating new power capacity can come online in months rather than years.