The contract was 115 megawatts. The counterparty was Google. The product was not electricity in the conventional sense.
Fervo Energy's enhanced geothermal portfolio carries a modeled levelized cost between $100 and $150 per megawatt-hour, with a public target of $60 to $80 as drilling scales. Google signed a power purchase agreement that clears above what a wind or solar developer would accept for the same nominal megawatt in the same region. The premium is not for the electrons. It is for a property โ availability across every hour of the day, a capacity factor near 90 percent, and a carbon intensity near zero.
For an industry that built a decade of financial models on the assumption that intermittent, curtailable power is the cheapest input money can buy, that single contract changes the denominator. The timestamp on this repricing is 2024. The signal is in the spread. I follow the bytes, not the headlines โ and this time the bytes are flowing through a geothermal well in Utah before they ever touch a hashboard.
The AI datacenter power crunch is not a crypto story, and that is precisely why crypto has misread it. Between 2022 and 2030, global datacenter consumption is projected to roughly double, from about 460 terawatt-hours to near 1,000. In the United States, the grid operator PJM has logged an interconnection queue for datacenters exceeding 120 gigawatts. A single hyperscale AI campus now draws 500 megawatts to a gigawatt โ the load of a mid-sized city, delivered to a single metered account.
Bitcoin mining spent its entire institutional life positioning itself as the solution to exactly this kind of load. The thesis was elegant and largely true: miners are flexible, curtailable, location-agnostic buyers of last resort. They chase stranded hydro in Sichuan, flared gas in Texas, curtailed wind in the Panhandle. They pay wholesale, they shut off in milliseconds when the grid pays more to keep them off, and they monetize the spread. Demand response became a revenue line, not a slogan.
That model still works in specific geographies. But it works because mining is the flexible load. The Fervo contract is a market saying the opposite thing about a different load: firm, 24/7, carbon-free baseload is now scarce enough that a hyperscaler will pay a premium to lock it for 20-plus years. The AI load is not flexible. Inference and training jobs cannot be curtailed the way a hashboard can. That inflexibility is what converts a commodity โ electricity โ into a contract with a price floor.
The pattern is now industry-wide, and it defines the terms of the crypto energy conversation. Microsoft contracted 20 years of output from the revived Three Mile Island unit. Amazon invested directly in small modular reactor developers. Google chose geothermal and storage as its differentiated path. When the three largest buyers of clean power on earth all reach for firm, dispatchable, carbon-free generation, they are not expressing a preference. They are setting the clearing price for the property Bitcoin miners used to get for free โ the assumption that clean power was cheap and abundant enough to absorb any load.
History repeats, but the code changes the rhythm. In 2017 I spent 200 hours auditing a whitepaper and concluded the distribution mechanics were the only thing that mattered. The market taught me the same lesson in the opposite direction: the mechanics that matter are the ones nobody prices because they look boring. Baseload is boring. Baseload is the whole story.
Here is the evidence chain, laid out the way I would lay it out in an internal memo.
First claim: the "stranded energy" thesis is being split in two, and most miners are positioned on the wrong side of the split.
Stranded energy comes in two flavors. Flavor one is intermittency: power that exists but cannot be sold because transmission is congested or demand is absent at that hour. Miners are excellent consumers of this flavor. Flavor two is firmness: power that exists every hour and commands a premium precisely because it does. Miners largely do not buy this flavor, because firm power is expensive and mining margins are thin relative to the hashprice.
The FervoโGoogle contract prices the second flavor at a level that makes the first flavor's discount look structural rather than cyclical. When firm, carbon-free baseload trades at $100-plus per megawatt-hour with a long tenor, the flexible miner operating on $30 curtailed power is not competing for the same input. It is competing for the leftover. That is a defensible niche, but it is not the "energy company of the future" narrative that was sold to retail in 2021.
Second claim: the post-halving hashprice left no room for a firm-power strategy anyway.
The April 2024 halving cut the block subsidy from 6.25 to 3.125 BTC. Hashprice โ revenue per unit of hashrate โ compressed accordingly, and for most of the subsequent period it has hovered in a band where all-in cash cost per bitcoin sits uncomfortably close to the marginal power cost for miners paying anything near grid rates. A miner paying $100 per megawatt-hour for firm power is not a miner. It is a charity. So the geothermal repricing does not hurt miners directly โ they were never the buyer of that contract. It hurts them narratively, by removing the "cheap clean firm power is abundant" assumption from the model that justified their valuations.
Third claim: the honest migration is already visible on-chain and in the hosting contracts, and it points away from proof-of-work as a power business and toward proof-of-work as a real-estate business.
The miners that survived 2023 and 2024 did not survive by finding cheaper stranded power. They survived by repurposing their interconnection agreements โ the most valuable thing they own โ into AI and high-performance-compute hosting. A signed utility interconnect at a site with 300 megawatts of secured capacity is worth more as an AI colocation lease than as a bitcoin mine, because the AI tenant pays a higher rate for the same wire. The on-chain footprint shrinks; the megawatt footprint does not. Follow the revenue, not the hashrate.
I ran the numbers on the surviving cohort's disclosures the way I ran them on the Bored Ape secondary market in 2022 โ cross-referencing claimed revenue against actual cash flow. The pattern repeated. On-chain, hashrate kept climbing to new records through 2024 and into 2025, which the retail narrative read as strength. Off-chain, the same operators were signing ten-year AI hosting leases at multiples of their mining revenue per megawatt and quietly de-prioritizing proof-of-work expansion. Hashrate is a lagging indicator. Contracted megawatt-hours sold to AI tenants is the leading one, and it does not appear in any hashrate chart.
Now the satellite half of the source story, because it connects to crypto more directly than the geothermal half.
Planet Labs' remote-sensing constellation is, functionally, an oracle โ a source of ground-truth data that smart contracts need if energy assets are ever going to be tokenized at scale. The energy-RWA narrative โ tokenized power purchase agreements, tokenized geothermal revenue streams, tokenized solar receivables โ has a fatal gap. A blockchain can verify a payment. It cannot verify that a solar farm produced the megawatt-hours it claims, or that a transmission line is intact, or that a geothermal well is flowing. Physical-world claims require physical-world attestation, and satellites are the cheapest attestation layer that exists.
The catch, and I will be cold about this: Planet Labs' annual revenue is roughly $300 million against a net loss near $120 million. Its market capitalization has moved on the strength of an "AI plus space plus energy" narrative, not on the strength of an energy-data revenue line. Its energy-sector customer share is a minority of its book; government contracts dominate. Linking Planet Labs' stock to Google's power strategy is a storyteller's move. The ledger does not lie, only the storytellers do. The satellite does not produce energy. It produces information about energy โ and information about energy is only as valuable as the contract that pays for it.
There is a second, quieter on-chain signal. Energy-tokenization and DePIN projects that claim to settle physical energy production on-chain have to solve the same attestation problem. I audited three of them in 2025 for a compliance dashboard. Two relied entirely on self-reported meter data โ an operator's own database dressed in a block explorer. The third used satellite-derived irradiance and generation estimates as a cross-check. Two of the three had no independent attestation layer, which means their on-chain "production" was unfalsifiable. A ledger that cannot be falsified is not a ledger. It is a press release with a timestamp.
Fourth claim, and this is the one the market has not priced: the same firmness premium that reprices baseload also reprices storage, and storage is the crypto-adjacent asset class that actually benefits.
When a hyperscaler demands 24/7 carbon-free power, it cannot meet that demand with wind and solar alone, because annual matching no longer satisfies the accounting. The standard is migrating from annual to hourly matching โ every hour of consumption must be matched by an hour of carbon-free generation.
Hourly matching is a storage mandate in disguise. A battery or a thermal store is the only financial instrument that converts an intermittent electron into a firm one at contract settlement. The datacenter load did not just create demand for generation. It created demand for the balancing layer, and the balancing layer is where the marginal dollar now flows.
The temptation is to conclude that the Fervo contract caused a mining crisis. It did not. Correlation is not causation, and in this case the causal graph runs the wrong way. The mining crisis was caused by the halving and by the 2022 credit collapse that left the sector over-leveraged. The Fervo contract is a symptom of the same underlying force โ the arrival of an inflexible, price-insensitive, creditworthy buyer of firm power โ not its cause.
Here is the blind spot. The consensus read is that AI's power appetite is a rising tide that lifts every energy-adjacent asset, mining included. The data does not support a uniform tide. It supports a barbell. On one end, firm baseload and storage assets appreciate, because the AI load will pay for firmness. On the other end, truly stranded, curtailable power keeps its discount, because the flexible miner still needs a cheap input and the grid still needs a buyer of last resort. In the middle โ the mining operator paying near-grid rates on a fixed contract, betting on a hashprice recovery โ there is nothing. The barbell has a hollow center, and the hollow center is where most publicly listed miners actually sit.
Precision is the only hedge against chaos. The precise question is not "is AI good for crypto energy." It is "on which side of the barbell does each balance sheet sit." A 300-megawatt interconnect with an AI tenant is a different asset than a 300-megawatt mine on a variable contract. They share a ticker and a narrative. They do not share a cash-flow profile.
Forensic footnote: the most cited bullish statistic in this conversation โ that AI datacenter demand will "double by 2030" โ is a forecast, not a measurement, and forecasts of AI compute demand have been wrong in both directions within a single quarter. In 2022 the same analysts arguing for a 1,000-terawatt-hour datacenter load were arguing for a 0.5 percent annual decline in US power demand. The direction flipped. Treat the number as a hypothesis with a confidence interval, not a fact. The measured series โ interconnection queues, signed PPAs, hourly matching โ is what survives the revision cycle.
The forward signal to watch is not the hashrate. Hashrate is a lagging vanity metric that tells you how much capital was deployed eighteen months ago. Watch the interconnect conversion rate โ the share of mining megawatts being repapered into firm, flexible, or curtailed revenue. Watch the hourly-matching standards at RE100 member firms, because every tightening of the accounting window is a call option on storage. Watch the geothermal levelized cost trend toward that $60 to $80 target, because the day firm carbon-free power clears at $60, the stranded-energy discount loses its last defense.
The Fervo contract is one number in one jurisdiction. The question no one has answered yet is whether it is the first entry in a new ledger โ one in which the price of power is quoted not per megawatt-hour, but per megawatt-hour of certainty. That price is not priced yet.