U.S. grid risks threaten AI buildout
- EPRI and the IEA now say AI’s next bottleneck is power infrastructure — not chips — as U.S. data-center demand collides with grid connection delays. - EPRI’s 2026 outlook puts data centers at 9% to 17% of U.S. electricity by 2030; the IEA says their electricity use jumped 17% in 2025. - That shifts the AI trade toward utilities, transmission, transformers, and onsite generation as the real gating factors for new capacity.
The AI buildout is running into a much less glamorous limit than GPUs. It’s the grid. Not in some abstract way, either — in the very physical sense of transformers, switchgear, transmission capacity, interconnection studies, cooling systems, and enough generation to keep giant campuses running. What changed over the past few months is that this stopped sounding like a niche utility problem and started showing up in the core forecasts for AI growth. EPRI’s 2026 data-center outlook and the IEA’s April update both frame power availability as a real near-term constraint on how fast new AI capacity can come online. ### Why is the grid suddenly the bottleneck? Because AI data centers are huge, concentrated loads. EPRI now projects U.S. data centers could consume 9% to 17% of national electricity by 2030, up from roughly 4% to 5% today. That doesn’t mean the whole country runs out of power. It means specific regions get hit with giant new demand clusters that local utilities and transmission systems were not built to serve on short notice. (powering-intelligence.epri.com) ### What exactly is in short supply? The boring hardware matters most. The IEA says supply chains for gas turbines and transformers tightened over the past year, while developers also face slow grid connections and other approvals. DOE has been funding transformer R&D and manufacturing support because demand pressure is coming from several directions at once — aging infrastructure, electrification, storm hardening, renewables, and now AI-heavy load growth. (powering-intelligence.epri.com) ### Why can’t developers just wait for utility hookups? Because time-to-power is now the whole game. Bloom Energy’s 2026 survey says power availability has become a defining boundary on data-center growth, not just another planning variable. It also says capital is shifting toward “power-advantaged” regions and that onsite generation is moving from backup plan to core strategy. Basically, if one market can offer megawatts in two years and another needs five, the first market wins even if the land is worse. (iea.org) ### Why are transmission lines such a headache? Because the U.S. grid connects slowly even before you add AI. FERC is now actively rewriting rules for how very large new loads — including data centers — connect to the transmission system, with action due by the end of June 2026. That tells you the problem is no longer hypothetical. The regulator is treating large-load interconnection as urgent enough to need a new framework. (bloomenergy.com) ### Is this just a local utility issue? Not really. NERC’s 2025 long-term reliability assessment, which covers 2026 through 2035, warns that demand growth from data centers and other large loads is rising faster than the infrastructure needed to support it. Utility Dive’s summary of that report says NERC lifted its 10-year peak-demand forecast by 24% on new data-center loads. The catch is that reliability problems show up region by region, not as one national blackout story. (ferc.gov) ### Does efficiency solve this? It helps, but not enough. The IEA says power use per AI task is falling fast, which is the good news. But total AI usage is rising even faster, and more workloads are energy-intensive. That’s the classic Jevons-style problem — each task gets cheaper, so people do a lot more of them. Net demand still climbs. (prod.nerc.com) ### Who benefits if this is the real constraint? Utilities, power developers, grid equipment makers, and anyone sitting on scarce deliverable capacity. If AI deployment depends on actual electrons rather than promised capacity, then companies with generation, transmission access, or fast interconnection pathways become strategic. That’s why the market keeps circling names tied to power-rich regions and dispatchable generation. This is less a pure semiconductor story now and more a race to secure megawatts. (iea.org) ### So what’s the bottom line? AI is not just a software and chips story anymore. It’s an infrastructure story. The next winners may be the companies that can deliver power, transformers, hookups, and cooling on schedule — because a data center without electrons is just an expensive warehouse. (iea.org) (bloomenergy.com)
