Ask most executives what's constraining the AI build-out and you'll hear chips, talent, or capital. Those answers made sense a year or two ago, but they're now out of date. Hyperscalers are on track to spend roughly $700 billion on AI infrastructure this year alone [1], which tells you capital was never going to be the limiting factor once the industry decided to build at this scale. What's actually holding projects back is something much less glamorous: a multi-year wait in line for grid power.
The median U.S. data center project now takes over 5 years from interconnection request to commercial operation [2], and Gartner projects that 40% of AI data centers will be power-constrained by 2027 [3]. Grid queues in the U.S. currently hold over 2,060 gigawatts of proposed capacity, nearly double the country's entire installed base [2], and most of that capacity will never actually get built. Somewhere in the last two years, the industry quietly shifted from a hardware race to a permitting and power race [7], and many companies haven't repriced their strategy to reflect it.
The businesses pulling ahead aren't necessarily the ones with the deepest pockets. They're the ones that have started treating power as a product they can engineer and sell, rather than a bill they simply pay. Talen Energy's nuclear power purchase agreement with Amazon is a good example: a deal worth roughly $18 billion in contracted revenue over 17 years [4], restructured specifically so AWS could get onto the grid faster while giving Talen a long-term, stable income stream. What started as a power company's side arrangement with a tech customer has become one of the more important vendor relationships in AI, built with the discipline of a financial product rather than a standard utility contract.
Fermi America's Project Matador, an 11-gigawatt private grid taking shape in the Texas Panhandle [5], tells a similar story. Project Matador blends gas, nuclear, solar, and battery storage into a system designed for one purpose. This lets AI tenants skip the interconnection queue entirely by generating and delivering their own power. This is evidence that Fermi isn't really a data center company, but it’s an energy company whose only customers happen to be compute buyers. The traditional boundaries between developer, utility, and power project sponsor are dissolving as companies like Femi find success outside these boundaries.
What should this mean for a company that isn't building a data center itself? For enterprise buyers, it adds a new question to vendor diligence priorities: does this provider control its own power path, or is it simply hoping the grid comes through on time? For investors, it suggests that power adjacent infrastructure, behind the meter generation, transmission equipment, and private grid operators, deserves to be evaluated as its own category. This category needs to have a risk and return profile that looks different from either traditional energy or traditional tech. For policymakers pursuing sovereign AI ambitions, it's worth remembering that national compute strategies are competing for the same transformers, turbines, and skilled labor as every hyperscaler on earth, and that sovereignty on paper carries little weight without a credible power plan underneath it.
None of this resolves quickly. Transformer lead times alone run three to five years [6], and new transmission infrastructure takes even longer to plan and build. For the rest of this decade, the pace of the AI buildout will be set by physical infrastructure timelines rather than by anyone's appetite to spend.
The companies, investors, and nations that recognize power as the product, not just an input they need to secure, are likely to be the ones still standing once the queue finally starts to clear.
Sources
This list reflects a representative sample of sources underlying the article's analysis and is not a comprehensive bibliography of every data point, broker note, filing, or report consulted in the underlying research.