The wire crossed my terminal at 06:14 Bogotá time. Google and Constellation, reportedly near a billion-dollar nuclear power deal. The source was not Bloomberg. It was not Reuters. It was Crypto Briefing — a crypto publication breaking an energy story.
That single fact is the trade. When the crypto press starts covering nuclear procurement, convergence is no longer a thesis. It is a settlement. And the number that should stop you is not "billion." It is the word "reportedly." One unnamed source. No 8-K, no docket number, no term sheet. The ledger was clean, but the vision was fragile.
Here is what the headline actually contains: one core fact, and a frame. The fact — a deal is reportedly near. The frame — a ten-figure number engineered to feel like proof. Everything else is inference dressed as reporting. I have audited enough contracts to know that the press release is marketing and the annex is truth. So let me give you the annex.
Context: the structure behind the number
The comparable is not hypothetical. In September 2024, Microsoft and Constellation announced a 20-year agreement to restart Three Mile Island's Unit 1 — now branded the Crane Clean Energy Center. 835 megawatts. Twenty years. Reported value around $16 billion.
Run the division. That implies roughly $100 to $115 per megawatt-hour. Now run the historical number. Existing US nuclear power purchase agreements have cleared in the $40 to $60 per MWh band for years.
That gap — call it a fifty-dollar premium — is the entire story. Not the billion. The premium. Google is not buying electricity. Google is buying scarcity priced in megawatt-hours, and it is paying double the old clearing price to lock it down.
Why now? Three load curves are colliding into one pool: AI training clusters, manufacturing reshoring, and the electrification of transport. US data center consumption ran about 176 TWh in 2023; forecasts for 2028 span 325 to 580 TWh — a range so wide it tells you nobody actually knows. Meanwhile PJM capacity prices have multiplied, and American power demand, flat for two decades, has turned upward.
Constellation is the largest nuclear operator in the country, roughly 22 gigawatts of carbon-free baseload running a capacity factor north of 90%. In a buyer's market, that asset is a commodity. In a seller's market, it is a toll booth. The market flipped. Constellation is now the toll booth.
And here is the part the energy desks miss, the part a crypto native sees instantly: the mining rig was the first AI data center. Since 2017, Bitcoin miners have been the marginal buyer of stranded power, the interruptible load grid operators could curtail at will. They wrote the playbook for monetizing wasted electrons. The AI boom is that playbook with a higher revenue per kilowatt.
But do not romanticize the miner. The rig is a call option on cheap power with a brutal strike. In 2020, running arbitrage across Ethereum lending markets and L2 testnets for three months, I learned that profit and volatility are the same variable measured from opposite sides. Miners live that lesson daily: their margin is power cost subtracted from block reward, and when the reward halves, only the cheapest electrons survive. That discipline is precisely what an AI data center now inherits. The compute economy did not discover energy scarcity. It inherited it from the miners.
Core: what "billion-dollar" hides
I spent six months in 2018 manually auditing token distribution contracts, and the lesson never left me: a number is a claim, a structure is a fact. "Billion-dollar" is a total. It is not a unit price. Ten billion over ten years at one gigawatt implies about $114 per MWh. Ten billion over twenty years at two gigawatts implies roughly $28. Same headline. Opposite economics.
This is the single most important gap in the reporting: total contract value and unit price have been deliberately blurred. A reader cannot tell whether Google locked in a premium or captured a scale discount. That ambiguity is not accidental. It is the difference between "we overpaid for scarcity" and "we bought smart." The press release wants you to feel the first number and assume the second.
Then there is the structure question, which the reporting does not touch at all. Two architectures exist, and they carry opposite risk profiles.
The first is co-location — the data center plugs directly into the plant, bypassing the grid queue. Fast, cheap, and legally exposed. In November 2024, FERC rejected the amended interconnection agreement for Talen and Amazon to co-locate at the Susquehanna nuclear station. The regulator effectively said: you cannot drink from the grid's cup and skip the grid's tab. That ruling is the direct precedent. If Google-Constellation is co-located, it walks straight into that uncertainty.
The second is a conventional grid-connected PPA. Cleaner legally, slower physically — US interconnection queues stretch for years. For an AI operator burning capital every quarter, waiting is the expensive part. For a hyperscaler, the time value of power exceeds the price of power. That is why co-location keeps resurfacing despite the regulatory slap.
Now layer on the fuel chain. Nuclear is not lithium or polysilicon; its supply risk is geopolitical. Russia holds an outsized share of global uranium enrichment, and HALEU — the fuel for advanced reactors — is even more exposed. Spot uranium spiked toward $100 per pound in early 2024, then settled into the $70 to $80 range. Read that carefully. The fuel got cheaper while the capacity got more expensive. That spread is the trade. The bottleneck is not the atom. It is the licensed, grid-connected, permitted plant.
The policy floor is real but fragile. IRA Section 45U hands existing nuclear a production tax credit of roughly $15 per MWh through 2032. The ADVANCE Act of 2024 accelerates licensing. Together they underwrite the confidence to sign a twenty-year contract. But subsidies live inside political cycles, and a long-dated PPA stacked on a variable credit is a bet with two moving parts. The contract hedges the operator's exposure to policy drift. That is the quiet reason the term is long.
In 2024, advising a Bogotá hedge fund on a $5 million crypto allocation, I argued for strict risk parameters against traditionalists who believed digital assets were the only volatility in the room. The real volatility was never in the token. It was in the infrastructure beneath it. The same is true here. The fragile variable in this deal is not Google's balance sheet. It is a regulator's docket and a grid's queue.
Contrarian: the blind spot
Retail reads this headline as Google saving the planet. The smart money reads it as a power-scarcity trade wearing a climate costume.
Watch the vocabulary. The industry says "carbon-free," not "renewable." That word choice is a regulatory maneuver. Google's 24/7 carbon-free energy goal matches supply and demand hourly, and to fold nuclear into that ledger, the "carbon-free" frame must be broad enough to admit it. "Renewable" would exclude the atom. The adjective was chosen to fit the asset, not the asset chosen to fit the adjective.
The second blind spot: the profit is migrating from the buyer to the seller. Before 2020, hyperscalers were the strong buyer, squeezing renewable developers on PPA price. Today the scarce asset is dispatchable carbon-free baseload, and the operator names the terms. Constellation, Vistra, Talen — the independent power producers — are the ones repricing upward. Google is the customer, not the winner.
And the third: the crypto miners who pivoted to AI hosting already front-ran this entire narrative. They held the interconnect agreements, the power contracts, the physical sites. They did not wait for the headline. We bet on the pattern, not the hype. The pattern was always this: whoever controls interruptible megawatts controls the compute economy.
Takeaway
Do not trade the press release. Trade the docket. The signals that matter are not Google's equity — they are the FERC filings, the PJM capacity auction clearing prices, and the spread between uranium spot and nuclear operator valuations. If capacity keeps repricing while fuel stays cheap, the toll booth keeps winning.
The summer was loud, but the profits were quiet. And the quiet money is sitting in the plants, not the press. In the void, we found the edge no one else saw.
