The data shows a simple fact. The CEO of Constellation Energy says existing power plants are the bedrock for data centers. That statement is a marketing statement. It is also an engineering statement. When a hyperscaler needs 500 megawatts of continuous power within three years, a solar farm still waiting in an interconnection queue is worthless. A gas turbine that already has a grid connection is worth far more. The PJM capacity auction for 2025/2026 cleared at $268.92 per megawatt-day, up from $28.92 in the prior year. That is a 9x jump. Transformer delivery times have stretched from under twelve months in 2021 to two to four years today. The scarcity is not theoretical. It is measured in auction prices and lead times. Existing generation is the only asset class that can settle the AI load contract today.
Let me frame the market structure. AI data centers are not normal consumers. They demand 24/7 high-load operation, and their power demand is projected to grow two to three times between 2023 and 2030, rising from about four percent of US electricity consumption to eight to ten percent. New generation projects face five to seven year interconnection delays. Many projects sit in queues longer than the useful life of the GPU chips they are meant to power. The supply cliff is visible in the retirement records. Coal plant retirements are being postponed. The Three Mile Island unit is being restarted. Constellation operates the largest nuclear fleet in the United States, and nuclear power provides emissions-free baseload around the clock. The Inflation Reduction Act adds $15 to $30 per megawatt-hour in production tax credits for existing nuclear plants. Gas and nuclear are not the only solutions, but they are the only solutions already built. In energy, as in software, structure defines value; chaos destroys it.
Investors have noticed the shift. Vistra, Constellation, and Talen, the owners of baseload fleets, have seen their valuations re-rate upward from 2019 to 2024. Pure renewable developers have faced rising interest rates and compressed valuations. The market is no longer paying for promises; it is paying for operating assets. The order flow confirms this. Microsoft signed a 20-year agreement with Constellation to restart Three Mile Island. The effective nuclear price is estimated near $115 per megawatt-hour, far above the $30 to $50 operating cost. Google and Amazon have signed dedicated nuclear and geothermal agreements. These are not consumers buying on the spot market; they are locking in twenty-year capacity contracts. Public perception still treats renewable energy as the obvious winner. Smart money is signing long-term contracts for existing baseload plants. That divergence is the clearest order-flow signal in the sector.
I have been stress-testing systems for most of my career. In late 2017, I manually audited an ICO smart contract that promised decentralized storage; I spent three weeks tracing Solidity and found three integer overflow vulnerabilities in the fundraising function. The team responded with marketing. In 2022, while the market debated macro narratives about Terra and Luna, I wrote a 5,000 word technical autopsy of the death spiral. In 2023, I reverse-engineered EigenLayer's restaking contracts on a local testnet to test slasher conditions that their documentation did not cover. The lesson in every case is identical: theoretical performance is not live performance. Energy markets follow the same logic. A battery shows nameplate capacity in a data sheet; a power plant shows capacity when the fuel has arrived, the coolant is flowing, and the grid asks for power. For a data center, the question is not which technology is cheapest in a cost model. The question is which asset keeps running under sustained peak load. In that test, an existing power plant wins.
The core cost analysis confirms the point. Battery storage, mostly LFP, now costs between $0.30 and $0.60 per kilowatt-hour of delivered storage cost, with cycle lives of 6,000 to 8,000 cycles. But that metric applies to four-hour duration systems. A 500 megawatt data center that loses grid power for three days requires 36,000 megawatt-hours of stored energy. No battery plant in commercial operation is sized for that. Long-duration storage using flow batteries or compressed air remains in early commercialization, with poor round-trip efficiency and high integration complexity. Large US nuclear plants generate at $30 to $60 per megawatt-hour marginal cost. A natural gas combined-cycle plant has higher fuel cost but is still firmer than any intermittent resource. Meanwhile, solar and wind carry hidden integration costs. Solar LCOE is low, but when the cost of 24/7 firming is included, the system-level LCOE of a solar-plus-storage portfolio exceeds baseload generation. US solar projects sit in interconnection queues for more than four years on average. Onshore wind capacity factors run between 35 and 45 percent, well below the availability commitments a data center needs.
Storage still has a role. It is just not the role its advocates advertise. In ERCOT and PJM, storage participates in frequency regulation and peak shaving, earning arbitrage revenue while improving grid stability. It is not contracted as the primary capacity source for a hyperscale facility. Microsoft has piloted fuel cell backup in some new data centers, substituting hydrogen for diesel in emergency generation, but the scale is megawatt-based, not gigawatt-based. The center of the reliability stack remains continuous thermal generation.
Hydrogen is even further away. Green hydrogen production costs $3 to $6 per kilogram, and even with Inflation Reduction Act subsidies falling to $1 to $2 per kilogram, the delivered cost of electricity from that hydrogen is far above gas or nuclear. Fuel cells for backup exist at megawatt scale, while hyperscale loads are hundreds of megawatts. The industry is not close to replacing continuous baseload generation.
Now the contrarian angle. The CEO's claim is correct, but the reason he says it involves his own balance sheet. When the operator of the largest nuclear fleet says existing power plants are the bedrock, he is not describing a natural law; he is marketing the assets he owns. A solar-plus-storage developer would give the same speech with different nouns. The phrase "existing" conveniently pushes solar and wind into the slow, uncertain category, while framing nuclear and gas as instantly available. The framing directly shapes PPA negotiations, capacity market rules, and policy debates about reliability payments. This is not a neutral technical observation; it is a strategic narrative with a price tag.
The false binary is the bigger problem. The efficient system for a 500 megawatt data center is not baseload versus storage. It is a hybrid: existing nuclear or gas for continuous power, batteries for millisecond-to-hours response, and demand response for peak events. Batteries are not a substitute for baseload; they are its protection. A four-hour battery can shave peaks and regulate frequency, tasks that a nuclear plant performs poorly and at high mechanical cost. By framing storage as competition, the CEO obscures the complementary role that would reduce wear on his own turbines and raise overall reliability. In practice, the winning structure will be a portfolio, not a religious preference.
The profit reallocation is the hidden layer. Data center demand is shifting profit from electricity consumers and technology firms to generation owners with existing baseload assets. The gap between the effective PPA price and the operating cost at Three Mile Island is a transfer of margin from a hyperscaler to a power producer. Supply chain players capture some of the flow as well. Uranium suppliers benefit from the nuclear narrative. Copper producers benefit from grid buildout. Transformer manufacturers capture the delivery bottleneck. The largest single margin sits at the generation gate. If a utility can negotiate a fuel adjustment clause, it can push fuel price risk back to the data center operator while keeping the capacity premium for itself.
"Existing" is also a snapshot, not a guarantee. Many existing plants are aging coal and gas units or nuclear reactors that were scheduled for retirement. Keeping them alive requires re-licensing, environmental compliance, cooling water access, and fuel supply. The United States still imports about 25 to 30 percent of its enriched uranium from Russia; the ban takes effect in 2028, but a disruption before then is a live operational risk. Coal plants face mounting carbon regulation. Gas plants face methane disclosure rules. Old plants carry liabilities that new designs are intended to avoid. Meanwhile, the physical link between plant and data center is the real bottleneck. Distribution transformers, substations, and dedicated transmission lines are the binding constraint. Transformer lead times of two to four years mean an existing plant without a transformer is still years away. The CEO's claim is accurate at the generation layer and incomplete at the delivery layer.
The forward-looking signals are clear. Direct, long-term contracts between hyperscalers and existing generators will multiply, priced above historical baselines; the Three Mile Island deal is the template. Hybrid announcements will follow, where storage is bundled with gas or nuclear; that bundle, not any single technology, is the actual operating model for reliable AI power. And the supply chain will act as an early warning system: transformer producers, high-voltage cable makers, uranium enrichment, and fuel logistics are the canaries for whether this boom is physically real. If transformer backlogs shrink, the scarcity narrative is cracking. If copper prices spike further, the data center buildout is accelerating.
We do not predict the future; we hedge against it. The hedge is a set of structural facts: existing plants are scarce, new generation is slow, storage is fast but shallow, and delivery equipment is the real choke point. When those facts change, the narrative changes with them. Until then, treat every revolutionary energy solution with the same skepticism as a code audit. Check the capacity, check the fuel, check the connection, and assume the white paper is fiction until the load test passes. A plant that exists beats a plant that is permitted, but only if the electrons can reach the server. And if the transformer order is already in the queue, the server might get powered before the next hype cycle ends. We do not predict that either; we just position the portfolio so the outcome does not matter.