Oracle just signed a nuclear power subscription for its Lighthouse campus. The math says it covers less than 19% of the load. The rest is a mystery. In a bull market, nobody audits the energy bill.
On October 2, 2026, Oracle announced a power subscription agreement with We Energies for 125โ250MW from the Point Beach Nuclear Plant in Wisconsin. The headlines screamed "AI goes nuclear." The stock tickers twitched. The press releases glowed with carbon-free promise.
The code doesn't lie. But press releases do.
I pulled the numbers. Lighthouse Campus is designed for 1GW of AI compute with a total power footprint of 1.3GW. That implies a PUE of approximately 1.3 โ nothing revolutionary, just standard air-cooled or hybrid cooling. The nuclear subscription covers 125MW to 250MW. That's 9.6% to 19.2% of the campus's total power needs. The remaining 1.05โ1.18GW? Completely undisclosed.
This isn't a nuclear-powered AI campus. This is a nuclear-marketed AI campus with a massive energy gap dressed up as a sustainability win.
I've been auditing smart contracts since 2018, when I spent six months in a dorm room in Istanbul hunting reentrancy bugs in early Compound and MakerDAO code. The lesson from those days translates directly to infrastructure deals: the gap between the narrative and the implementation is where the risk lives. If a protocol claimed 100% collateralization but only locked 15% on-chain, I'd flag it as a red flag. The same logic applies here.
Oracle is locking in a small slice of existing nuclear capacity from a plant that's already operating. Point Beach is brownfield โ no new construction, no new technology, just a contract on electrons that were already flowing. Meanwhile, Microsoft locked up 835MW from Three Mile Island with a 20-year PPA and a $1.6 billion investment. Google ordered 500MW from Kairos Power's SMR program. Amazon dropped over $650 million on a Pennsylvania campus plus SMR equity.
Oracle's 125โ250MW is the smallest nuclear commitment among the hyperscalers. And it's the only one using a subscription model rather than a PPA or direct investment. That's not innovation born from strength. That's a smaller player using a novel structure to compete for scarce existing resources because it can't win on scale.
Let's talk about the cost curve, because this is where the story gets ugly.
Point Beach's levelized cost of energy was $45.94/MWh in 2016. By 2026, it's $75.51/MWh โ a 64% increase. By 2033, projections put it at $122.45/MWh. That's another 62% jump. Nuclear was supposed to be the cheap, stable baseload. The data says otherwise. The cost of nuclear electrons is doubling in less than two decades.
Now run the annual cost: 125โ250MW ร 8,760 hours ร $75.51/MWh = $82.7 million to $165 million per year at 2026 prices. By 2033, that balloons to $134 million to $268 million annually. Against a $15 billion campus investment, energy is a persistent operational cost that compounds every year. And if Oracle's subscription includes cost pass-through clauses โ which the article doesn't disclose โ then Oracle's cost per GPU-hour rises in lockstep with nuclear fuel and maintenance costs.
The subscription model itself is the interesting part. Oracle is committing to pay the full energy cost for its share of the project. That's not a PPA. That's a take-or-pay arrangement disguised as a subscription โ Oracle is the de facto offtaker for a slice of Point Beach's output, absorbing the risk of price increases in exchange for priority access to a scarce resource.
This is an options trade. Oracle is betting that electricity will be more expensive and harder to secure in the future. They're paying a premium now to lock in supply. It's a rational hedge if you believe AI compute demand will keep growing and grid capacity will remain constrained.
But here's the contrarian angle: the subscription model might be a sign of weakness, not strength. A company confident in its ability to secure cheap power through traditional PPAs or direct investment wouldn't need a novel structure. Oracle's position in the AI stack is weak. They're not the model layer โ that's OpenAI. They're not the dominant cloud โ that's Azure, AWS, and GCP. They're positioning as what the article calls a "compute landlord" โ providing the physical infrastructure but not the intelligence layer.
The landlord model has a moat: scarce land plus grid interconnection rights. But that moat can be eroded by policy. Wisconsin's Public Service Commission is currently considering whether to require tech companies to pay 100% of the cost of new power plants. If that ruling goes against the hyperscalers, Oracle's subscription economics could invert overnight. The $3 billion in "savings" that We Energies claims the deal will generate for customers by 2027โ2033? That's a utility's marketing number, not an audited figure. Divide $3 billion by seven years and you get roughly $43 million annually โ a rounding error against the $176 million rate increase We Energies is requesting, 20% of which is driven by this very deal.
The utility wins either way. Data center load expands the rate base. Regulated returns flow from the expanded rate base. Customers pay more today. The promise of future savings is just that โ a promise.
Now let's talk about what nobody is saying: the other 80% of the power. If nuclear covers less than 19% of Lighthouse Campus, where does the rest come from? Natural gas is the obvious answer. Gas plants can be built faster than nuclear, operate flexibly to match AI inference loads, and are cheaper per MWh. But gas isn't carbon-free. It doesn't fit the ESG narrative that Oracle is selling.
So the article doesn't mention it. The press release doesn't mention it. The narrative is "nuclear-powered AI," but the reality is likely "nuclear-marketed AI with a gas backbone." I've seen this pattern before โ in DeFi protocols that claimed "decentralized governance" while a multisig controlled all the keys. The marketing layer and the execution layer are different things. Always check both.
This is where my experience with restaking in 2023 becomes relevant. When I ran an EigenLayer node with $100,000 staked across multiple AVSs, I optimized latency and increased daily yield by 15% over the network average. The lesson was simple: the details of infrastructure configuration determine the actual returns, not the headline APY. The same applies here. The headline is "nuclear subscription." The detail is a 10โ19% coverage ratio with an unspecified gas fallback and a cost curve that's doubling every 7โ10 years.
The AI energy arms race is real. It's structural, not cyclical. Data center power demand hit 29.6GW by the end of 2025 โ roughly the entire peak demand of New York State. The IEA projects a 130% increase by 2030, pushing demand to approximately 68GW. Wisconsin's peak demand is up 40% in five years. These aren't projections born from hype cycles. They're load forecasts from grid operators who have to plan capacity years in advance.
But here's what the bull market doesn't want you to notice: the energy constraint might be the trigger that forces AI efficiency gains. When power becomes the binding constraint, the economic value of smaller models, sparse architectures, and MoE (Mixture of Experts) approaches increases. Google's Gemini and Anthropic's Claude already use MoE to reduce inference costs. If nuclear costs are rising 60%+ per decade, the incentive to use 10x less compute per inference becomes overwhelming.
The market is pricing AI energy plays as pure growth stories. Nuclear operators like Constellation, Vistra, and Talen are getting re-rated on the "AI power purchase" narrative. Grid equipment suppliers, transformer manufacturers, and HVDC vendors are seeing order books swell. All of that is real. But the market is not pricing the second-order effects: higher energy costs squeezing AI margins, forcing algorithmic efficiency, and compressing the demand curve for raw compute.
I'm not bearish on AI. I'm bearish on the assumption that energy is an infinite resource that can be tapped without cost consequences. Trust the math, fear the hype, ignore the noise. The math says nuclear covers 10โ19% of one campus. The math says costs are doubling. The math says the other 80% of the power is undisclosed.
The truth is in the gaps. The gaps are where the alpha is โ and where the losses hide.
So what should you watch? Three things. First, the Wisconsin PSC ruling on cost allocation. If the commission forces tech companies to pay 100% of new generation costs, the subscription model's economics shift dramatically. Second, Point Beach's cost trajectory. If LCOE hits $122.45/MWh by 2033, and Oracle's contract has cost pass-through, then Oracle's AI business margins compress from both sides โ higher energy costs and competitive pressure on compute pricing. Third, the undisclosed power supply. If the remaining 1.05โ1.18GW comes from natural gas, the "clean AI" narrative collapses, and the political risk of the entire deal rises.
Alpha isn't in the press release. It's extracted from the chaos of the footnotes. And right now, those footnotes are missing.
In a bull market, anyone can be a genius. Sustainable genius requires understanding the energy bill. Sleep on that โ it's still priceless.
