There is a comforting story we tell ourselves about artificial intelligence: that it lives in the cloud, that it runs on code, that it scales the way software always has, which is to say almost for free. As someone who has spent decades watching molecules and electrons move through real systems, I can tell you that story is about to collide with the part of the world that does not care about your roadmap. AI does not run on code. It runs on power. And power runs on copper, steel, water, and people, none of which you can download.
That is the thread running through the latest wave of institutional research crossing my desk this weekend. The Wall Street framing has quietly shifted. For two years the conversation was about chips and capital, about who could raise the most money fastest to build the most data centers. Now the sharper analysts are admitting that the binding constraint sits upstream of the chips, in the electricity system itself. Power is no longer a line item. It is the project.
The Constraint Moved Upstream
Start with the scale of what we are trying to plug in. The International Energy Agency reports that electricity demand from data centers jumped roughly 17 percent in 2025, far outpacing the 3 percent growth in global electricity demand overall, and it projects data center consumption more than doubling to around 945 terawatt-hours by 2030. In the United States, the EIA now expects data centers to account for close to half of all electricity demand growth through the end of the decade. We have not seen new load arrive like this since air conditioning went mainstream.
The problem is not that we lack the will to build power plants. The problem is timing. A data center can be poured, wired, and filled with servers in eighteen months. The grid it wants to connect to runs on a much slower clock, and that mismatch is where the whole story lives.
Lead Times Don’t Care About Your Press Release
Consider the humble transformer, the unglamorous steel-and-copper box that steps voltage up and down so electricity can actually move. Before the pandemic you could order one in three or four months. Today, lead times for large power transformers have stretched to roughly two and a half to three years, with generator step-up units running near 144 weeks. That is not a supply hiccup. It is a structural shortage, and it now dictates the schedule of every serious power project in the country.
The grid connection itself is no faster. Lawrence Berkeley National Laboratory’s latest tally found more than 2,000 gigawatts of generation and storage waiting in interconnection queues, a figure that rivals the entire installed capacity of the US power system, and the typical project now spends about four and a half years in line before it delivers a single watt. You can announce a gigawatt of AI compute on a Tuesday. The hardware and the connection to feed it are a half-decade away.
The Workforce You Cannot Conjure
Even if the steel showed up tomorrow, someone has to install it. The trade most in demand is the one we have neglected for a generation. Industry estimates point to a shortfall of hundreds of thousands of electricians, with roughly 20,000 retiring every year and a large share of those still working already past 55. Microsoft’s own president has called the electrician shortage the single biggest bottleneck slowing data center expansion. When the electrical workers’ union estimates that electrical work accounts for 45 to 70 percent of data center construction cost, you start to understand why a wiring crew has become a strategic asset. You cannot prompt your way to a journeyman electrician. That takes years of apprenticeship and real hands on real conduit.
Why the Money Is Merging
Put these constraints together and you get the most interesting shift of all. If you cannot wait five years for the grid, you bring the power plant to the campus. Developers are increasingly building behind-the-meter generation, natural gas turbines, fuel cells, and storage, sited right next to the servers. Goldman Sachs estimates that behind-the-meter solutions could supply a quarter to a third of incremental data center demand through 2030, and order books for fuel cell makers have roughly doubled in a year. The consequence is that the capital pools that used to be separate, energy on one side and technology on the other, are collapsing into one. The AI companies are becoming energy companies, whether they intended to or not.
This is the part I find genuinely clarifying. For years the energy transition was framed as a contest between the old physical economy and the new digital one. The AI buildout has ended that argument. The most software-driven enterprise in human history has discovered that it is, at bottom, an industrial problem: a question of turbines, transformers, water rights, and skilled hands. The firms that win the next decade will not be the ones with the cleverest models. They will be the ones who understood early that you can’t download a transformer. The physical world always sends its invoice, and it is coming due.
Further Reading
- IEA: Data centre electricity use surged in 2025
- EIA projects record US data center power use amid the AI boom (DCD)
- Lawrence Berkeley National Laboratory: Queued Up, interconnection queue data
- pv magazine USA: Transformer lead times extend toward four years
- CNBC: The AI data center boom and the skilled-trades shortage
- Goldman Sachs: Fuel cells and behind-the-meter power for data centers
- S&P Global: Beneath the surface, water stress in data centers