Oracle's Nuclear Subscription Covers 10–19% of Lighthouse's 1.3GW Load — The Missing 1.05GW Is the Signal

Oracle subscribed to 125–250 megawatts of nuclear baseload from the Point Beach plant. The Lighthouse Campus it feeds draws 1.3 gigawatts at peak. Do the arithmetic and the headline collapses into a footnote: the cleanest energy deal in AI infrastructure covers somewhere between 9.6% and 19.2% of the actual load. That is not a rounding error. That is a collateralization ratio. Glitch detected. Source traced. On October 2, 2026, per the filings We Energies submitted to the Wisconsin Public Service Commission, Oracle agreed to a nuclear "subscription" for its share of the Lighthouse Campus — a $15 billion AI data center developed alongside OpenAI and Vantage. It is the first arrangement of its kind in the state. It is being reported as a clean-energy triumph, an anchor for the Stargate buildout, and proof that hyperscalers are finally paying their own way. I spent two days reading the underlying numbers the way I read a Solidity contract before a mainnet deploy. What I found was a structure that is 85% undisclosed, priced against a cost curve that is rising, and validated by a savings figure that no one outside We Energies has audited. The megawatts are real. The story around them is not. To understand why a nuclear subscription matters to anyone holding compute tokens, you have to read the AI arms race as a series of collateral pledges. Microsoft went first and went biggest. In 2024 it committed to restarting Three Mile Island's Unit 1 — 835 megawatts, a 20-year power purchase agreement, roughly $1.6 billion behind it. The restart targets 2027. Google took a different exit: a 500MW order with Kairos Power, betting on small modular reactors that do not yet exist in commercial form, with first power optimistically in 2030. Amazon bought its way in — a Pennsylvania campus plus a $500 million equity stake in X-energy's SMR program. Then came Stargate, the buildout consortium, which turned a power problem into a portfolio problem. Oracle sits at the bottom of that table. It is the smallest nuclear commitment of the four hyperscalers — 125 to 250MW, one-tenth of Microsoft's. But Oracle did something none of the others did: it bought from the shelf, not from the future. Point Beach is already running. There is no restart risk, no regulatory first-of-a-kind, no 2030 horizon. Oracle rented existing baseload. The mechanism is called a subscription. Oracle does not sign a power purchase agreement. It does not buy equity in a reactor operator. It commits to absorb the full energy cost of its share and locks the supply — an arrangement We Energies says is a first for Wisconsin. The utility, for its part, proposed a $176 million rate increase for 2027, roughly 20% of which it attributes to this agreement, while projecting the deal will save customers about $300 million in fuel costs between 2027 and 2033. That is the framing the market received. Now let me audit it. Start with the ratio no press release mentioned. Lighthouse is designed for 1GW of AI compute and a 1.3GW total power footprint. That spread implies a power usage effectiveness of about 1.3 — the multiplier between what the chips burn and what the facility draws from the grid, including cooling and distribution losses. Against 1.3GW, Oracle's 125–250MW subscription covers 9.6% to 19.2%. Which means 1.05 to 1.18 gigawatts of Lighthouse's load is unaccounted for in the disclosures. Nobody has said where it comes from. Not in the filings, not in the announcements, not in the coverage. If I were reading this as a bridge contract, I would stop here. A bridge that is 15% collateralized is not a bridge. It is a promise. The nuclear subscription is not powering the campus; it is anchoring a narrative and buying a green credential. The remaining 85% — the part that actually determines whether the lights stay on — is silent. This is the first place the crypto framing pays off, because I have seen this exact structure before. In 2021 I spent two weeks reverse-engineering the Bored Ape Yacht Club contract, chasing the ERC-721 implementation down into the off-chain metadata retrieval path. The traits were served from centralized servers. The "digital scarcity" was real on the token and fictional in the metadata — the team could have rewritten a trait without any on-chain record of the change. NFT metadata mismatch found. Lighthouse has the same shape: the on-chain claim (nuclear) is verifiable and small; the off-chain dependency (the other 1.05GW) is unverified and dominant. You are being shown the audited line and asked not to ask about the rest of the balance sheet. Now price the subscription against its own trend line. Point Beach's levelized cost of energy ran $45.94 per megawatt-hour in 2016. By 2026 it is $75.51 — a 64% increase in a decade. The projection for 2033 is $122.45 per megawatt-hour, another 62% on top. Liquidity draining. Logic broken. Nuclear power's entire pitch is that it is expensive to build and cheap to run — a fixed cost amortized over decades, with fuel a minor line item. That pitch assumes the running cost is flat. The Point Beach curve says the running cost is compounding faster than most equity indexes. Run Oracle's number. 125–250MW times 8,760 hours times $75.51 is roughly $83 million to $165 million per year at 2026 prices. At the 2033 price of $122.45, the same contract runs $134 million to $268 million per year. Energy is not the dominant cost of a $15 billion campus, but it is the one line that is contractually rising while everyone else assumes compute gets cheaper. And here is the crypto-native read: if the cheapest baseload on earth is doubling, the "compute gets cheaper forever" assumption that underwrites every DePIN token is not a thesis. It is an unhedged short. Nuclear is rigid. It runs 24/7 at a near-constant output because ramping a reactor is expensive and slow. AI training is a decent fit — it is a long, flat, predictable load. AI inference is not. Inference is spiky, bursty, and correlates with human attention cycles. Attaching a volatile load to a rigid baseload source requires either storage, curtailment, or a second, flexible source — usually gas. None of that is in the disclosure. The subscription locks the supply. It does not explain how a 125–250MW slab of flat nuclear output maps onto a variable inference profile. That mapping problem is the operational core of the deal, and it is missing. Compare it to how proof-of-work miners solved this years ago. Miners are the ideal grid citizens precisely because their load is interruptible — they can shed in milliseconds when prices spike. AI inference cannot do that. You cannot tell a user their completion request will resume when the reactor settles. So the AI campus needs the flexibility that PoW mining provided for free, and it has to buy that flexibility with storage or gas capacity. The coverage did not mention gas once. In a 1.3GW facility covered 10–19% by nuclear, the word "gas" is doing a lot of load-bearing work by its absence. The PUE of 1.3 is itself a confession. If Lighthouse were built on aggressive direct-to-chip liquid cooling, a modern AI facility can approach a PUE of 1.1. A value of 1.3 is the signature of conventional air or hybrid cooling. That 0.2 difference at 1GW of compute is roughly 200MW of pure thermal overhead — a Point Beach worth of capacity spent on fans and chillers instead of silicon. And "1GW of AI compute" is a nameplate. Data centers quote design capacity, which is a peak ceiling, not a steady-state load. Real utilization in AI facilities is frequently below 50% for long stretches. So the grid must be sized for a peak the facility may rarely hit, while the actual draw is lower and lumpier — a mismatch that inflates the infrastructure the utility builds and the rate base it recovers. Exchange volume anomaly flagged: the number the public is told is the number the grid must provision for, not the number the grid will see. Strip the vocabulary and Oracle's subscription is a priority gas auction in physical form. In a congested block, whoever pays the highest priority fee gets the scarce resource — block space. In a congested grid, whoever signs the biggest take-or-pay commitment gets the scarce resource — firm baseload. Oracle is bidding for priority access to a zero-sum resource. The US produced essentially the same nuclear output in 2025 as in 2020, and there is not a single commercial SMR under construction. Every megawatt Oracle locks is a megawatt someone else cannot. Which brings me to the naming coincidence I cannot ignore. The company is called Oracle. It has just built a single-source, single-counterparty dependency for a critical input — the physical equivalent of routing an entire protocol through one price feed. We Energies is the node operator. The subscription is the feed. If We Energies misprices, delays, or renegotiates, Oracle has no second source and no fallback. For years I have argued that oracle feed latency is DeFi's Achilles' heel and that solving decentralization with a permissioned node set is itself a joke. Oracle the company has now constructed the same joke in concrete: a centralized trust anchor dressed as a resilient system. The irony is not lost on me, and it should not be lost on anyone pricing Oracle's AI margin. There is a second reading, and it is the charitable one. Oracle is choosing to become a regulatory partner rather than wait to be regulated. By signing a subscription the utility can point to as customer-funded clean capacity, Oracle buys goodwill in a state that is actively deciding whether to make tech pay for its own generation. That is the same move PayPal made with PYUSD — pre-empt the regulator by becoming the counterparty the regulator needs. It is a hedge, not a religion. But a hedge still has a price, and that price is the 85% of the load that remains undisclosed. Now audit the $300 million. We Energies projects the deal saves customers roughly $300 million in fuel costs across 2027–2033. That is a single-party projection. It has not been independently verified. Spread across seven years, it is about $43 million a year — against a $176 million rate increase, 20% of which the utility attributes to this very agreement. The savings are forecast; the increase is filed. One of these numbers has a signature on it and the other has an asterisk. If the $300 million does not materialize, there is no disclosed mechanism for who absorbs the shortfall. There is no take-or-pay clause visible in the public record, no price collar, no index. The subscription's term, its price adjustment mechanism, and its minimum volume are all undisclosed. I am being asked to price a derivative with no published strike. This is where the Dencun analogy lands hard. When EIP-4844 shipped, blob space was effectively free and rollups priced their entire cost model around that abundance. My position then, and now, was that blob demand would saturate within two years and rollup gas fees would double once it did. The lesson is that subsidized abundance is a temporary state, and the cost model built during the subsidy always breaks. AI compute is living through its own blob era. Energy is being presented as abundant and clean and locked. The Point Beach curve says the abundance is priced to expire. In 2024 I built a Python model to track real-time institutional inflows into BlackRock's IBIT and found a correlation between traditional market volatility and crypto ETF outflows that the mainstream press ignored because it did not fit the narrative. I bring that up because the same narrative filter is operating here. The story the market wants is "hyperscalers are going green." The story the data tells is "hyperscalers are securing scarce baseload and passing the flexibility bill to the grid." Model it. We Energies expands its rate base by serving a hyperscale load, earning a regulated return on the infrastructure it builds for that load. The utility is not a neutral party in this deal. It is a beneficiary with a regulatory incentive to attract exactly this kind of customer. That does not make the deal wrong. It makes the deal's sponsor interested, and it makes the $300 million projection a marketing input as much as a financial one. Here is the part that should worry anyone long compute tokens. The entire decentralized-compute sector — render networks, GPU marketplaces, DePIN nodes — prices its future on a single assumption: that centralized compute is expensive and getting more so, and that distributed supply can undercut it. That assumption has a hidden dependency. It requires that the marginal cost of compute trends down over time. Energy is the floor under that marginal cost. If the floor is rising — and a 62% increase in the cheapest baseload by 2033 says it is — then the arbitrage that decentralized compute is supposed to exploit narrows on both ends. The centralized incumbents are locking energy now, at today's prices, with contracts that insulate them from the curve. The decentralized networks are buying energy on the spot market, exposed to every uptick. Oracle's subscription is a hedge. Most DePIN networks are not hedged at all. They are short energy without knowing it. Then there is the regulatory question that determines everything downstream. The Wisconsin Public Service Commission is weighing whether to require technology companies to bear 100% of the cost of new generation built to serve them. If it rules yes, the subscription's economics change shape: the utility's incentive to expand its rate base shrinks, and the full cost lands on Oracle's side of the ledger. If it rules no, retail ratepayers inherit a share of infrastructure built for a single hyperscaler — a cross-subsidy that will be litigated for a decade. There is no national precedent for this ruling. Whoever writes it sets the default for how AI's energy bill is split in the United States. That is a bigger prize than 250 megawatts. And beneath the generation question sits a bottleneck nobody in the coverage touched: interconnection. The US grid's queue to connect new load and new generation runs into years. The physical grid is aging and capacity-constrained — the coverage acknowledged as much and then moved on. A 2028 completion target for Lighthouse depends not only on 1.3GW of supply existing but on that supply being physically connectable on schedule. The generation is the visible problem. The queue is the invisible one, and it is frequently the binding constraint. Every risk here compounds on the same variable. The coverage gap is 1.05GW. The cost curve is rising 62% by 2033. The regulatory outcome is undecided. The interconnection queue is unquantified. Any one of these can be absorbed. All four at once, on a $15 billion asset with a 2028 deadline, is a stress test, not a strategy. Everyone reading this story is reading the wrong risk. The consensus worry is that AI will overwhelm the grid — that data centers will brown out cities. The contrarian risk is the opposite: that the grid will overwhelm the AI buildout, and the nuclear subscription is the tell. Here is the blind spot. The deal is celebrated because it is clean. But the cleanest available baseload is the one whose cost is rising fastest, and the coverage is only 10–19% of the load. Read together, those two facts describe a system that is not solving its energy problem — it is decorating it. The 85% gap will be filled, and the cheapest way to fill it is gas, which contradicts the entire green narrative the subscription exists to sell. The second blind spot is the one crypto should own. We spent a decade defending proof-of-work against the charge that it wastes energy, and the industry's answer was to move to proof-of-stake and call the problem solved. The AI industry is about to discover that you cannot stake your way out of thermodynamics. Compute has a physical settlement layer, and that layer is electricity. Bitcoin miners understood this in 2017. The hyperscalers are relearning it in 2026. The DePIN tokens are still priced as if the lesson never happened. Watch three numbers. The Wisconsin PSC ruling on who pays for new generation. The disclosed source of Lighthouse's missing 1.05GW. And the Point Beach price curve in 2033 — because that is the number that decides whether "cheap compute" survives as a thesis. Oracle has bet that energy gets scarcer and more expensive. The market has not yet decided whether to believe it. When it does, the repricing will not start with Oracle. It will start with every compute token that assumed the floor under its marginal cost would hold. There is a version of this where Oracle is right and everyone else is late. If energy is the new scarcity and Oracle locked a slab of it at 2026 prices with a utility that needs the customer, then the subscription is a cheap call option on a world where megawatts trade like blockspace. But that version requires the other 1.05GW to appear, the PSC to rule gently, and the cost curve to bend. Three conditions, one signature on the document. The math is public. The rest is a promise.