Power Is Now AI's Binding Constraint — and the Market Just Repriced It

The constraint that gates AI's growth has slipped from silicon to electrons. Chips are still the easiest part of the compute cycle to scale — you can buy more wafers, more capacity, more HBM. Electricity is the part you cannot order on a delivery timesheet. Data centers already take roughly 6% of U.S. power, and analysts at BloombergNEF project that share climbing toward a fifth by 2035, with grid access falling a reported 19 gigawatts short of what data centers will want. This is the concrete reason the AI trade's next big beneficiary may not be a chipmaker at all, but the companies that own the power those chips cannot run without.

The reason power binds so hard is that an AI cluster is not a negotiable customer. Model training and live inference need firm, around-the-clock, high-density electricity — the same load 24 hours a day, not the weather-dependent bursts a grid usually tolerates. That leaves nuclear as the only carbon-free baseload source available at meaningful scale. So the hyperscalers have stopped shopping and started locking up supply. Meta struck agreements across four partners for as much as 6.6 gigawatts by 2035; the Talen-Amazon deal pins 1,920 megawatts of Susquehanna output to Amazon's adjacent campus through 2042.

What matters for an investor is the economics hiding inside those contracts. A power purchase agreement does more than signal demand — it converts a scarce physical asset into a contracted, multi-decade revenue stream. Constellation's 20-year deal to sell Meta the output of its Clinton, Illinois plant — 1,121 megawatts, beginning June 2027 — largely replaces the state-funded credits that were set to expire in mid-2027, swapping ratepayer subsidies for a market contract two decades long. That is the moat equivalent in an industry that builds its moats out of baseload: once a fleet is under a 20-year PPA, a good portion of its future cash is already written down.

Yet check the price action, and the power trade has repriced hard even as that demand thesis stayed intact. Vistra is down about 14% year to date at $138, with a roughly $46 billion market cap; Constellation has fallen about 21% to $280. Both sit well off their 2026 highs and near their 52-week lows. A pure "more AI demand" story would not explain that. What changed is that the market stopped paying a scarcity premium and started pricing delivery risk.

The reason is timing. Look at the actual dates on these contracts: Clinton starts June 2027, Meta's Vistra uprates land "early in the 2030s," TerraPower's Natrium reactors as early as 2032, Oklo's as early as 2030. Signed contracts with 2027–2035 delivery are not today's earnings. "The physical world takes much longer to develop than what people might imagine it takes," Vistra CEO Jim Burke said on the company's August call — the operating signal behind the multiple reset. Regionally, data-center projects in Vistra's largest market, Texas, have been frozen, pushing more cost and delay onto developers. Structurally, Constellation was downgraded in May when analysts argued large parts of its fleet still behave like a regulated utility rather than dedicated AI infrastructure.

That downgrade cuts to the real question, and it is not "who has the most nuclear." It is who is better positioned for the stage we are entering. The first phase of the trade ran on nuclear scarcity — whoever owned the most baseload got the premium. The second phase is about deliverability: hyperscalers do not want a single asset, they want a firm, dispatchable, grid-connected system that can carry a multi-gigawatt cluster. That favors a diversified generator like Vistra or NRG, which can pair nuclear baseload with gas-fired flexibility and transmission, over a pure nuclear story whose capacity, however scarce, does not by itself move a cluster's physics. The leader everyone recognized — Constellation, with its Microsoft Three Mile Island restart — may be the one for whom that delivery burden is heaviest.

None of this kills the thesis. Power genuinely is the next binding constraint in the cycle, and the owners of firm baseload are signing the customer into 20 years of contracted revenue while the market has knocked their multiples down roughly a fifth. The judgment to hold onto is the inverse one: demand is not the issue anymore, so the question is whether execution and timing still justify the same allocation. That means separating what already reaches revenue — the operating fleet under PPA, which monetizes but later than today's price reflects — from what is only a claim: reactor restarts and small modular reactors that are not yet built, licensed, or delivered. Those are the parts of the power trade where a road map has not yet met a financial result, and they are exactly the ones to track by milestone, not by headline.