Power infrastructure has become the binding constraint on AI scaling. Industry surveys show 70%+ of respondents cite power and grid capacity as the primary buildout obstacle. Global AI data center demand projections reach 68 gigawatts by 2027 and 327 gigawatts by 2030—nearly three times current total data center capacity of 114-122 gigawatts, excluding traditional growth. Individual training runs may require one gigawatt by 2028 and eight gigawatts by 2030. As of mid-2025, over $162 billion across 36+ projects had been blocked or delayed by power constraints. The bottleneck is transmission and distribution infrastructure, not generation capacity. Grid connection requests take four to seven years in Virginia, while new transmission lines require five to ten years. Texas has tens of gigawatts requested but only one gigawatt approved. Regulatory reform efforts are underway but timelines remain compressed at best to three to four years.

Natural gas turbines have become the immediate solution, installing in weeks to months and bypassing utility approvals entirely. xAI deployed 500+ megawatts of truck-mounted turbines for Colossus in July 2024 while awaiting grid connection through November 2024. OpenAI and Oracle ordered a 2.3 gigawatt on-site gas plant in Texas in October 2025. xAI’s Memphis facility is ordering up to 60 turbines. Companies deploy aeroderivative turbines—retired aircraft engines in trailers—for rapid deployment. But limitations mount: turbine wait times approach 18-24 months, direct carbon emissions trigger scrutiny, multi-gigawatt facilities require gas delivery infrastructure, and operating costs exceed grid electricity. Co-location with existing power plants or retiring coal facilities reduces connection timelines to three to four years by leveraging existing transmission, but suitable sites are limited and competition intense.

Longer-term firm, carbon-free solutions remain on extended timelines. Small modular reactors offer 50–300 megawatts per unit with multi-decade operating lifetimes, but development cycles typically span ten to fifteen years, including licensing and construction. Enhanced geothermal systems present a potentially faster path to firm baseload in certain geologies, though commercial-scale validation is needed. Both technologies may meaningfully contribute post-2030 but are unlikely to relieve near-term constraints at scale. More speculative concepts—including space-based solar power and orbital compute facilities—attract capital and technical exploration but appear outside plausible 2030 deployment windows for material relief.

Ground-based renewables combined with storage face scale and duration limits at multi-gigawatt load profiles. Data centers cannot tolerate multi-hour interruptions, and current lithium-ion systems provide hours—not days—of coverage. Long-duration storage technologies are progressing but are not yet commercially deployed at multi-gigawatt, multi-day scale.

Efficiency gains have become critical: “tokens per watt per dollar” maximizes output from available power. Blackwell chips promise 25x energy reduction per inference task versus previous generations at the chip level. Combined with operational improvements—advanced liquid cooling (immersion, direct-to-chip) reducing overhead and enabling higher density, increased GPU utilization minimizing idle time, and workload scheduling optimizing for power availability—the industry might achieve 5-10x effective capacity gains over two to three years. Demand flexibility offers additional headroom: training versus inference have different power profiles, workload migration can follow power availability geographically, and time-shifting to off-peak hours reduces grid stress. But efficiency faces diminishing returns and cannot deliver the 50-100x compute demand increase that aggressive scaling trajectories suggest.

All companies pursuing long-term solutions simultaneously deploy gas turbines because they cannot wait a decade. The power constraint remains the defining infrastructure challenge through 2030, with solutions spanning immediate fossil fuel deployment, medium-term efficiency extraction, and long-term carbon-free baseload.