For two years the entire AI trade has been a bet on one company’s chips and the handful of names that feed it. That bet has worked. But the constraint has quietly moved. You can order all the accelerators you want; if there is no power to run them and no substation to connect them to, the racks sit in a parking lot. I have spent the last few weeks reading grid-operator filings instead of earnings calls, and the picture they paint is the more investable story right now.
The number that reframes everything
US data centers used somewhere around 4% of the country’s electricity in 2023. Every serious forecast I have seen, from the utilities themselves to the Electric Power Research Institute, has that figure landing between 8% and 12% before the end of the decade. Call it a doubling in five years for a load category that took thirty years to reach its current size. The grid was not built for that slope, and it cannot be rebuilt on that timeline.
Where the actual bottleneck sits
It is not one thing. It is a chain, and every link is stretched.
Interconnection comes first. A new data center or a new power plant has to get in line to connect to the grid, and in the largest US market that queue now runs for years. PJM, the operator that covers thirteen states from Chicago to Washington, ran a capacity auction for the 2025 to 2026 delivery year that cleared near $270 per megawatt-day. The prior auction cleared around $28. That is not a typo and it is not a rounding difference. The price the grid pays to guarantee enough supply went up roughly nine times in a single year, because demand forecasts jumped and old coal and gas plants kept retiring on schedule.
Then there is the hardware to make the electricity. The big three gas-turbine makers are effectively sold out through 2028, in some slots into 2029. If a developer decides today it needs a gigawatt of new gas generation next to a campus in Texas or Ohio, the turbine itself is a three-to-four-year wait. Large transformers, the unglamorous boxes that step voltage up and down, have lead times that have stretched past two years. None of this is fixable with a checkbook alone.
The hyperscalers understood this before the market did. That is why Microsoft signed a twenty-year deal to restart a shuttered reactor at Three Mile Island, why Amazon bought a data center campus wired directly into a Pennsylvania nuclear plant, and why Meta and Google have both signed long-dated nuclear agreements including early bets on small modular reactors. When the richest companies on earth start contracting for 2028 power in 2025, they are telling you the constraint is real.
How the money is exposed
The clean way to think about this is by layer. Each has a different risk profile.
| Layer | Representative names | How it wins | Main risk |
|---|---|---|---|
| Merchant generation | Constellation, Vistra, Talen | Sells power into tight markets at rising prices; signs direct deals with hyperscalers | Prices are cyclical; a demand air pocket hits earnings fast |
| Turbines & grid gear | GE Vernova, Siemens Energy | Multi-year backlog already booked; pricing power on every new order | Execution and warranty costs; backlog is only as good as delivery |
| Regulated utilities with data-center load | NextEra, Southern, Dominion | Rate base grows with every new connection; earnings are contracted returns | Regulatory lag: they spend now and earn the return over years |
| Nuclear-heavy operators | Constellation, Vistra | Carbon-free baseload is exactly what a 24/7 data center wants | Restart and license timelines slip; one outage is material |
My own preference sits in the middle two rows. The merchant generators have already run a long way and their earnings swing hard with power prices, so you are buying momentum as much as a thesis. The turbine makers and the regulated utilities with genuine data-center load growth give you a slower, more contracted version of the same trade. GE Vernova in particular sells into every one of these projects regardless of which generation source wins, and its backlog is a real number sitting on a real balance sheet, not a forecast.
The regulated utilities work differently, and the difference matters. When a utility like NextEra or Dominion connects a new data center, it has to build the substations, the lines, and often new generation to serve it. That spending goes into the rate base, and the utility earns a regulated return on it, usually somewhere around 9% to 10%. The catch is timing. The capital goes out the door in 2026 and 2027; the approved return shows up in customer bills, and therefore in earnings, over the years that follow, after a rate case. So a utility with real data-center demand is a slower compounder than a merchant generator in a boom, but its earnings are a contracted stream rather than a bet on next winter’s power prices. In a market this uncertain about the pace of the AI buildout, I will take the contracted version.
What would break this
Two things, and I watch both.
The first is phantom demand. Grid operators have admitted that a large share of the load sitting in their interconnection queues is speculative or duplicative, the same project filed with three utilities at once to see who says yes fastest. Estimates of how much of that queued demand actually gets built range widely, and some run below half. If the real buildout comes in at the low end, the tightest forecasts loosen and the merchant-power trade deflates well before any plant gets turned off.
The second is the AI capex cycle itself. If the spending that is driving all of this slows, the way I laid out as a real possibility in my piece on whether this is 1999 again, the power demand curve bends down with it. Electricity here is a derivative of the compute buildout, not independent of it. A hyperscaler that trims its 2027 data center plans by 20% is also trimming its power procurement, and the contracts that look ironclad today have offramps.
What I am fairly confident about is the shape of the next few years. The chip shortage of 2023 and 2024 got solved the way shortages usually do, with more fabs and more supply. The power shortage cannot be solved that quickly, because you cannot permit, finance, and build a power plant or a transmission line in eighteen months. That mismatch is where the pricing power lives, and it is why I think the second phase of the AI trade looks less like a semiconductor stock and more like a utility bill. For the full map of who sits where in this spending wave, I went through it in detail in the AI infrastructure guide.
Financial disclaimer: The content on StockVane is for educational and informational purposes only and should not be construed as professional financial advice. Stock market investing involves risk of loss.