On a Friday afternoon in early October 2026, a regulatory filing in Madison, Wisconsin, did something three years of AI enthusiasm had failed to do: it put a number on the physical ceiling of the compute economy. Oracle — the company that spent 2024 recasting itself as the "compute landlord" of the Stargate era — agreed to subscribe to between 125 and 250 megawatts of nuclear power from We Energies' Point Beach plant. The headline wrote itself. The reality did not.
The Lighthouse Campus that Oracle is developing alongside OpenAI and Vantage draws 1.3 gigawatts at full footprint. The nuclear subscription covers, by the most generous arithmetic, 19% of it. The remaining 1.05 to 1.18 gigawatts — the overwhelming majority of the site's power — is nowhere in the document. Not a footnote. Not a placeholder. Absent.
That absence is the story. And for anyone who has watched crypto's own energy wars unfold since 2017, it should feel queasily familiar. The narrative always arrives before the infrastructure, and the infrastructure bill always arrives last.
From the ashes of 2017 to the fluidity of DeFi, crypto taught us that a market is a story first and a mechanism second. I learned that the hard way while finishing a cryptography PhD in Berlin, watching ICO whitepapers promise distributed utopias while their market caps were driven by Telegram sentiment. I analyzed more than 500 of them. The projects with the strongest community narratives outperformed the technically superior ones by roughly 300%. That finding pushed me out of pure engineering and into media, because it revealed something structural: capital does not flow toward code. It flows toward belief about code.
The AI energy race is the same phenomenon wearing a different jacket.
By late 2026, the pattern is unmistakable. Microsoft committed to a 20-year power purchase agreement for the 835-megawatt restart of Three Mile Island, backing it with $1.6 billion. Google placed an order with Kairos Power for 500 megawatts of small modular reactor capacity targeted at 2030. Amazon acquired a Pennsylvania campus and put $500 million into SMR equity. And Oracle — the smallest nuclear commitment of the four — chose the one route none of its rivals took: a subscription to existing, already-operating nuclear capacity, paired with a promise to absorb the full energy cost of its share of the project.
The competition has migrated. Getting electricity now matters as much as getting GPUs. For a decade, crypto miners understood this before anyone else. The industry that was mocked for burning power to produce hashes is watching the world's most sophisticated companies discover that power is the product.
But look closely at the Oracle number, because the gap is where the analysis lives.
Point Beach's subscription covers 125 to 250 megawatts against a 1.3-gigawatt footprint. Run the division: 9.6% to 19.2%. That is not energy self-sufficiency. That is an anchor — a green-lit headline that secures a scarce, brownfield resource and leaves the hard question, where does the other gigawatt come from, unanswered.
Based on my own experience tracing liquidity flows during DeFi Summer, when I followed $50 million across yield farms to map where capital actually sat versus where the narrative said it sat, I've learned to distrust the number that gets published and hunt the number that doesn't. Here, the unpublished number is enormous. The most plausible filler for the missing capacity is natural gas — cheap, dispatchable, and politically inconvenient for a project branded around clean nuclear power. The document simply declines to say.
Then there is the cost curve, which is where this story stops being about AI and becomes about everyone.
Point Beach's levelized cost of electricity ran $45.94 per megawatt-hour in 2016. By 2026 it had climbed to $75.51 — a 64% increase. By 2033, projections put it at $122.45 per megawatt-hour, another 62% on top. Nuclear power has spent fifty years marketed as the cheap, stable baseload of the future. On this data, that story is false. The cheapest generation source in the room is getting more expensive, not less.
I have argued for two years that the Layer 2 economics of Dencun were a trap — that blob space would saturate within two years and rollup gas fees would double once the subsidy faded. The mechanism here is identical in spirit. A temporary abundance was mistaken for a permanent law. Cheap blobs, cheap nuclear, cheap compute — each was a moment, not a floor. When the moment ends, the bill doubles and the narrative has to be rewritten in public.
There is a quieter inefficiency hiding in the engineering. The Lighthouse design implies a power usage effectiveness of about 1.3 — 1.3 gigawatts of total draw for 1 gigawatt of IT load. That is a conventional figure; aggressive liquid cooling can push a modern AI data center toward 1.1. The number suggests the campus is built on familiar thermal assumptions. More importantly, "1 gigawatt of AI compute" is a peak design capacity, not a steady-state draw. Real utilization may sit far below it, while the grid must be provisioned for the peak regardless. That mismatch — capacity reserved, capacity unused — is exactly the kind of structural inefficiency that utilities recover from ratepayers and hyperscalers never advertise.
The subscription model deserves its own skepticism. Oracle is not signing a power purchase agreement; it is subscribing, becoming the de facto offtaker for a defined slice of nuclear capacity while promising to carry the full cost — including, apparently, the upside risk of that $122.45 figure. That is an options trade dressed as a procurement decision. Oracle is betting that power in 2033 will be scarcer and more expensive than the market currently prices. It might be right. But a bet that requires the world to get worse for the bettor to win is not a moat. It is a wager.
The competitive picture sharpens the point. Against Microsoft's 835 megawatts, Google's 500, and Amazon's campus-plus-equity play, Oracle's 125 to 250 megawatts is a rounding error in the national energy race. Oracle is not leading the energy transition. It is renting a seat at the table, using existing plants because the future plants do not exist. American nuclear output has been essentially flat since 2020, and there is not a single small modular reactor under construction in the United States. Everyone is competing for a fixed, zero-sum pool of brownfield capacity. When one player locks a slice, everyone else's slice shrinks.
This is where crypto's institutional memory becomes useful. DePIN — decentralized physical infrastructure — was sold as the answer to exactly this bottleneck: distributed energy, verifiable supply, tokenized capacity. In practice, the hyperscalers are executing the vertical integration that DePIN promised, with balance sheets that no token can match. From the ashes of the ICO era to the fluidity of tokenized capacity, the decentralization narrative keeps being absorbed by the very institutions it was designed to route around. I have watched this before. It is the same gravitational pull that turns a "community-owned" protocol into a foundation with a treasury and a compliance officer — the USDC lesson, applied to electricity. A compliance-first stablecoin that can freeze any address within twenty-four hours is not decentralized finance; it is a bank wearing a blockchain. A nuclear subscription that anchors 10% of a campus is not energy independence; it is a hedge wearing a sustainability report.
The regulatory layer is the real precedent. Wisconsin's Public Service Commission is weighing whether to require technology companies to bear 100% of the cost of new generation. If it does, it becomes the first such ruling in the country — a template that decides, for a generation, who pays for AI's electricity. We Energies has proposed a $176 million rate increase for 2027, roughly 20% of it driven by this agreement, while projecting $300 million in fuel savings for customers through 2033. That savings figure is a utility's own forecast, unaudited, spread across seven years — about $43 million a year. It is not a gift to ratepayers. It is a public-relations line item.
And here is the contrarian read that most coverage will miss. The obvious story is "AI is eating the grid." The subtler story is that AI has just done crypto an enormous reputational favor. For years, Bitcoin's energy consumption was the industry's favorite cudgel — the argument that ended conversations. Now the world's most valuable companies are signing gigawatt-scale nuclear deals and being asked, in public hearings, whether ordinary families should subsidize their power. The energy villain has changed costumes. Crypto's "wasteful" narrative did not get refuted; it got company. That shift is worth more to the industry's political standing than any lobbying budget.
The second blind spot is the assumption that compute costs fall forever. If nuclear — the resource we were told was the cheapest — is doubling in cost toward 2033, then the long-run energy curve for AI is bending upward, not downward. Every model trained on the premise of declining marginal cost should be re-examined. Scarcity is returning to the input layer, and the market has not repriced it yet.
So watch three numbers, not the headline. Watch the missing 1.05 to 1.18 gigawatts, and whether it resolves to gas, grid, or storage. Watch the Wisconsin PSC ruling, because it sets the price of admission for every future campus. And watch the cost curve, because $122.45 per megawatt-hour is a forecast that, if it holds, rewrites the economics of both AI and the miners who now compete with it for the same electrons.
The next narrative is already forming: energy as the scarce asset, power contracts as collateral, and a quiet competition over who gets to be the landlord of the physical world. The question is not whether compute will need more electricity. It is whether the public will keep paying for the story that says it is cheap — or whether, like the ICO buyers of 2017, they will discover the bill only after the narrative has already moved on.
