In 2024 Bitcoin's code cut in half the new coins it pays its miners, as it does every four years, and what a megawatt of power spent on mining earns fell by more than half. Artificial intelligence companies now pay more to rent a miner's powered site than mining takes in on the same power: about twice what mining keeps after paying for electricity, and nearly three times what it keeps once its machines are replaced too. The miners who held the connections to that power are becoming landlords. What the switch is worth, who has made it, how the stock market noticed, and a dated test of whether the rent will last.
The most valuable thing a Bitcoin miner owns turned out to be its electricity. Bitcoin pays new coins to whichever machine first guesses a number that fits a rule. The more guesses a second, the more often a miner wins, and every guess costs electricity. So for a decade the industry bought power, land and connections to the electric grid, and filled them with machines that guess numbers.
In April 2024 the Bitcoin code cut the reward for guessing in half, as it was written to do. A megawatt of power, a million watts, enough to run ten thousand 100-watt bulbs at once, that had earned a miner about $2.6 million a year earned about $1 million by this fall.
In the same two years the artificial intelligence boom made a connected megawatt the scarcest input in computing. AI computers run day and night on very large amounts of electricity, and a big new user of power must wait its turn to be connected to the grid. In Texas, the grid operator, ERCOT, was tracking about 410 gigawatts of requests from large new users in March 2026 (a gigawatt is a thousand megawatts), about 87% of them for data centers. That is about fifty times the site power of all 23 AI leases in this inquiry. A megawatt that is already connected lets a tenant skip the line.
AI companies, and the big cloud companies that rent computing over the internet, are now signing 10 to 25 year leases with former miners. A lease here rents the miner's site and its power connection; the tenant brings its own computers. The median rent is about $1.86 million a year for each megawatt of computing load, the power that actually reaches the computers. A site needs about 1.4 megawatts from the grid for each of those, because cooling and electrical losses use the rest, so per megawatt of a site's power the rent is 1.86 / 1.4 = about $1.3 million. That is a third more than mining grosses on the same megawatt, about twice what mining keeps after paying for electricity, and nearly three times what it keeps once its machines are replaced too. Google, Amazon, Meta or a top-rated company usually stands behind the rent, promising to pay it if the tenant does not.
David Teece, an economist who studies who profits from new technology, explained in 1986 why this happens. When a new technology is easy to copy, much of its profit goes to whoever owns the scarce assets it needs to reach the world. For AI in 2026 the chips are for sale to anyone with the money (IREN, a former miner that now runs AI computers for Microsoft, bought the ones for that contract from Dell), so chips alone are no lasting advantage. A site with power already connected is harder to come by. We estimate that the owner of the powered site collects in rent about a fifth of what the computers in it take in. On IREN's figures, rent at that level would equal about three quarters of what the computers earn once the chips are paid for.
The switch is what economists call a real option: the right, but not the duty, to make an investment now or later, one that cannot be undone once made. This one has a clock on it. On our illustrative numbers, turning a megawatt of a mining site's power into a site ready for an AI tenant costs about $7 million (about $10 million for each megawatt that reaches the computers). Mining buildings are cheap, about $0.7 million to $1 million a megawatt, and can be switched off whenever power is dear. AI computers need steady power and a far costlier building, so what carries over is mostly the power itself. The conversion pays back in about seven and a half years from the start of construction, while mining the same megawatt after the next halving, due around April 2028, keeps almost nothing once its power and replacement machines are paid for. To beat the lease, mining would need Bitcoin near $93,000 today, against about $83,500 at the end of September, and near $186,000 after the halving, if nothing else changes. A miner that stays keeps the opposite bet, a call on Bitcoin's price; a doubling by the halving would only restore today's thin margin.
The stock market has already noticed. Over the past year, the miners that have signed AI leases or left mining have nearly stopped moving with Bitcoin and move with chip stocks instead; the ones still committed to mining at scale, CleanSpark and Riot Platforms, two publicly traded Bitcoin miners, keep much of their tie. In other words, investors already price the converts as landlords.
We set down two tests, to be scored on October 31, 2027. The first asks whether the rent holds as more former miners offer their sites: if the scarce asset is really scarce, new AI leases of up to 15 years signed over the next year will price at a median of at least $1.7 million a year for each megawatt of computing load. We leave out longer leases, because their built-in yearly raises lift their average. The second asks whether the market keeps pricing these companies by what they do now: IREN, which is leaving mining this year, should see its share price stop following Bitcoin's, while CleanSpark, which mines until its AI rent begins, should keep its tie.
This inquiry began with two questions I asked our AI research desk on October 7, while we were finishing Inquiry No. 14 on Coinbase, a crypto exchange. I was curious whether "the crypto party and the AI party are correlated", and I said I was "very intrigued about further developing AI connectivity to crypto, Coinbase, Circle, CoreWeave". (Coinbase was the subject of No. 14, and CoreWeave is the first family in the story below; Circle, the company behind the dollar-linked coin USDC, we leave for another day.) No. 14 answered the first question with prices: Bitcoin and the AI stocks have moved together since 2020 mostly because both ride the same tide of appetite for risk, and once that tide is taken out, very little is left. This inquiry follows the second question past the prices, to the sites. The closest published work is a working paper by Wei Wei and Yun Wan on the stock market's reaction to miners' AI deals, and Galaxy Research's report of December 2024, Bitcoin Miners Powering the AI Revolution; they and others are credited by name in the section called "What is new here, and what is not."
My question treated crypto and AI as two parties at the same party. They turn out to share an origin, and the story has two founding families. They met early.
In September 2017 a company called The Atlantic Crypto Corporation was formed to mine cryptocurrency, "primarily Ethereum", in the words of its later prospectus. Ethereum is a second network, with its own coin, Ether, and the detail matters. Bitcoin is mined on chips built for that one job and good for nothing else. Ether was mined on graphics processors, the chips designed for video games, which turned out to be the chips artificial intelligence runs on. So the company's mining machines were, half by design and half by luck, AI computers waiting for work.
The design came first. Ethereum was proposed by Vitalik Buterin, then a nineteen-year-old programmer, in a white paper released on November 27, 2013, and it went live in July 2015. Its founders had watched Bitcoin mining fall to companies running chips built for nothing else, chips the white paper called "thousands of times more efficient" at the task, and they wanted Ethereum's mining to stay open to people with ordinary computers. So they designed its mining puzzle to be "memory hard": solving it meant reading a large block of data from memory over and over, which gives a special-purpose chip little edge and suits the graphics card in a gaming computer. Special chips for Ether were eventually built, but graphics cards could still mine it until the end. The other half was luck. What made graphics cards good at Ether's puzzle, thousands of simple calculations at once and very fast memory, is also what training an AI model needs, and the image-recognition program of 2012 that started the modern wave of AI had been trained on two graphics cards made by NVIDIA. Ethereum chose the chip for fairness; AI chose it for speed.
The company renamed itself CoreWeave in December 2019, and in 2020 it began renting out its processors' computing time to other companies, while it went on mining. That proved to be good timing, because Ethereum's builders had always meant to retire mining. They planned to replace it with a system in which holders who lock up their own Ether, instead of miners burning electricity, take turns confirming transactions, and they wrote a "difficulty bomb" into the code to make mining steadily harder and push the change along. The change came on September 15, 2022, in an upgrade called the Merge, which cut Ethereum's energy use by about 99.95%, and CoreWeave closed its mining business that month. Like the Bitcoin miners and their halving, as we will see, the Ether miners had years of warning. "[P]rior to 2022, we had limited revenue, most of which was derived from our crypto mining offerings, which we have discontinued," its prospectus said.
What CoreWeave sells now is time on chips of that kind, to companies that build and run AI models: mostly under contracts of several years that the customer pays for whether it uses the time or not, and some by the hour. By the end of 2024 its 32 data centers were running more than 250,000 graphics processors, most of them NVIDIA's. Microsoft was about two thirds of its revenue in 2025, and OpenAI and Meta have signed contracts with it worth tens of billions of dollars. In 2025 the former Ether miner reported $5.1 billion of revenue from that business, and at the end of June 2026 its customers had signed contracts to pay it about $104 billion more in the years ahead, about twenty times its 2025 revenue.
One of CoreWeave's founders has joked that its early machines ended up in his grandfather's garage in New Jersey. As it grew, it put some of its graphics processors in other people's buildings, and from 2019 thousands of them lived with a Bitcoin miner. That miner was Core Scientific, which mined Bitcoin on its own machines and also housed other companies' machines for a fee, a business the industry calls hosting. "From 2019 to 2022, Core Scientific hosted thousands of CoreWeave's GPUs [graphics processors] in its data centers," Core Scientific said later. It did not say what those chips were computing, and in those years CoreWeave did both: it mined Ether throughout, and from 2020 it also rented out computing time. Then came the crypto winter of 2022, a long slump in crypto prices. By late November Bitcoin's price had fallen 65% for the year, the network's computing power had more than doubled in two years, so each machine's share of the coins Bitcoin pays out had fallen by more than half, power cost more, and some hosting customers had not paid parts of their bills, by Core Scientific's own count. On December 21, 2022 it filed for bankruptcy in federal court in Texas. It emerged in January 2024, three months before the halving, a Bitcoin miner with a great deal of electricity and a business that was about to earn half as much from it.
Here is how the arrangement works, because everything that follows repeats it. Core Scientific is the landlord. It owns the buildings, the connection to the electric grid and the cooling. CoreWeave is the tenant. It brings its own computers, mostly NVIDIA chips, plugs them in, and pays rent for the space and the power.
In March 2024 CoreWeave came back as a tenant, and on June 3 it signed 12-year contracts to put about 200 megawatts of AI computers into Core Scientific's buildings, worth more than $3.5 billion in rent. By February 2025 the contracts had grown to about 590 megawatts of computing load and $10.2 billion.
Then the tenant tried to buy the landlord. On the same June 3 it made an unsolicited offer of $5.75 a share in cash for the whole company. Three days later Core Scientific said no, calling it an offer that "significantly undervalues the Company." In July 2025 CoreWeave tried again, this time with an offer paid in CoreWeave's own shares rather than cash, valued at about $9.0 billion when it was announced, or $20.40 a share. CoreWeave's stated reason was the "immediate elimination of over $10 billion of cumulative future lease overhead," which in plain terms means it would stop paying that much in rent and related costs over the years. On October 30, 2025 Core Scientific's shareholders voted the offer down: 203 million shares against, 21 million for. The tenant had wanted to own its landlord, and the landlord's owners chose to keep the rent. Neither the filings nor the vote count say why.
Why would a tenant rent when it could own? Because of time. A building is the easy part. A connection to the electric grid is the hard part, and it cannot be hurried: a company that wants a large new connection joins a queue, and the equipment that makes the connection, such as the large transformers that step the grid's high voltage down for a site, can take as long as four years to arrive, according to analysts at PwC. CoreWeave rented because renting put its computers to work years sooner, and the price was modest. Core Scientific's rent of about $1.44 million a year for each megawatt of computing load is roughly 14% to 21% of what CoreWeave earns from a megawatt, which works out to between $6.9 million and $10.3 million. Under the contracts CoreWeave also pays up to $1.5 million a megawatt of the fit-out, credited against its rent, so what the rent buys, for most of its 12 years, is the powered site.
Core Scientific, then, is almost purely a landlord now. In the second quarter of 2026, 83.2% of its revenue came from AI and computing, and it says it intends to convert "every megawatt in our portfolio" within three years. The CoreWeave contract is take-or-pay, meaning CoreWeave pays whether or not it uses the space; the tenant pays the power and utilities; and the rent averages about $850 million a year for 12 years. That is a steady, bond-like income. The price of the steadiness is that the landlord gave up the upside: if computing becomes more valuable, the tenant keeps the difference.
Those two families, the miner that became a cloud and the miner that became its landlord, are the whole story in miniature. Everything that follows is the second family growing, from one landlord to eight: Core Scientific, TeraWulf, Applied Digital, Cipher, Hut 8, CleanSpark, Riot and Galaxy, all former miners except Applied Digital, which hosted other people's miners.
To see why the miners were willing to become landlords, you have to see what happened to their income, and it happened on a schedule written into Bitcoin in 2009.
About every ten minutes, Bitcoin's miners race to guess a number that fits a rule. The first to find one seals the next block, a batch of recent payments added to Bitcoin's public record, the blockchain, and is paid a fixed number of new coins, which the industry calls the subsidy, plus the fees users attach. The code cuts that subsidy in half every 210,000 blocks, about every four years. At block 840,000, sealed at 9 minutes and 27 seconds past midnight, coordinated universal time, on April 20, 2024, the subsidy fell from 6.25 bitcoins to 3.125, which at that day's price of about $64,000 a bitcoin meant from about $400,000 a block to about $200,000. The miners knew the date for years, because anyone who could read the code could count the blocks. It was the rare famine announced in advance, like the dream Joseph read for Pharaoh, seven years of great abundance followed by seven of famine, with the difference that Pharaoh was told how long each season would last and the miners were told the block number. Nobody could claim surprise.
The industry measures its income as hashprice, the dollars earned each day by a petahash per second of computing (a thousand trillion guesses a second). One current machine, Bitmain's S21+ Hyd., makes 358 trillion guesses a second, about a third of a petahash. At a hashprice of $105 it earned about $38 a day; at $39.87, about $14. Bitmain is a maker of mining machines. Luxor's index put hashprice at about $105 the day before the halving, about $49 on April 29, and $39.87 on September 28 of this year.
What matters to an owner of power is what that means per megawatt-hour, a megawatt used for an hour, the unit in which power is bought. That puts mining, the power bill and the AI rent on one scale. The arithmetic is short, and we take it step by step. A machine's efficiency, e, is measured in joules per terahash. A terahash is a trillion guesses. A joule is a small amount of energy: a 100-watt bulb uses 100 joules every second. A machine rated at 15 joules per terahash spends 15 joules for every trillion guesses, and the lower the number, the better the machine. A petahash is a thousand terahashes, so a petahash every second at e joules each uses 1,000 x e joules a second, which is e kilowatts. Over a day's 24 hours that is 24 x e kilowatt-hours. Dividing the day's dollars by the day's kilowatt-hours gives dollars per kilowatt-hour, and a megawatt-hour is 1,000 kilowatt-hours, hence the 1,000:
revenue per MWh = hashprice x 1,000 / (24 x e)
where hashprice is in dollars per petahash per second per day and e is the machine's joules per terahash. A fleet, all of a miner's machines together, of good modern machines runs at about 15 joules per terahash. The newest on sale does 8.9, and the network's average in mid-2024 was about 28. Before the halving, at $105, that fleet earned 105 x 1,000 / (24 x 15) = about $292 a megawatt-hour. A megawatt-year is a megawatt used all year, 8,760 hours, so $292 x 8,760 = about $2.6 million. At this September's price it earns 39.87 x 1,000 / (24 x 15) = $110.75 a megawatt-hour, or about $970,000 a year. That is 62% less. Miners pay roughly $28 to $54 a megawatt-hour for power, so $110.75 leaves a thin margin before the machines themselves are paid for. Same machines, same power, same work: we hold the fleet at 15 throughout so the comparison is fair.
Where did the 62% go? We can split it from the blockchain itself, Bitcoin's public record of every block. Comparing the day before the halving with September 28, the reward per block fell to 0.447 of its old size, from about 7.04 bitcoins to about 3.15, fees included (the halving, plus fees that shrank from 11% of the reward to under 1%). The network's difficulty, a setting Bitcoin's code raises as more machines join, so that blocks still come about every ten minutes and the same coins are shared among more guessers, rose, so each machine's share fell to 0.651 of what it had been. And Bitcoin's price rose by 31%. Multiplied together, 0.447 x 0.651 x 1.307 = 0.380, which takes hashprice from about $105 to about $40. The price helped. It did not help nearly enough.
The next halving falls at block 1,050,000. At Bitcoin's target of one block every ten minutes, that is about 553 days from October 8, which puts it around mid-April 2028, an estimate rather than a date. At today's price and difficulty, it would take the same fleet from about $111 a megawatt-hour to about $56, and its yearly revenue from about $970,000 to about $488,000. Even before then, the miners' own reports show revenue not far above the cost of power. TeraWulf, a miner in New York State, took in about $79 for each megawatt-hour in the second quarter (its machines run at 17.6 joules per terahash, less efficient than our fleet) and paid about $49 for the power, leaving about $30 before its machines were paid for. CleanSpark's mining took in about $86 a megawatt-hour.
In 1986 David Teece asked a question that every inventor has asked in the dark: why do the people who make the breakthrough so often fail to get rich from it? His examples were EMI, whose engineer Godfrey Hounsfield shared a Nobel Prize for the CAT scanner while General Electric ended up with the market, and RC Cola, which introduced the first diet cola and watched Coca-Cola and Pepsi follow almost at once. His answer was that who profits depends on how easily the innovation can be copied and on who owns the complementary assets it needs to reach customers: the factories, the distribution, the sales force. When an innovation is easy to copy and the assets it needs are scarce and specialized, the owners of those assets take a large share.
I should say that David is a friend of more than thirty years, and I value the friendship more with each passing day as I notice how many of these inquiries end up citing him. It is not by design. His work is simply that important. He always joked about his record-breaking number of citations, and now I know why.
The oldest example of the same lesson is a gold rush. Gold was found in California on January 24, 1848, and by 1849 thousands of people were heading west to dig for it. Most made a little money, and only a few struck it rich. Some of the richest were storekeepers. Sam Brannan, who owned the only store between San Francisco and the gold fields, bought up the picks, shovels and pans he could find, and is reported to have paid 20 cents each for pans that he sold to prospectors for $15 apiece. Levi Strauss, who sold supplies to the same crowd, is another name the histories give. The miners in this inquiry are literally miners, and the rush they joined had its own shovel sellers. So does the new one. The shovel is the chip, and chips are for sale to anyone with the money. The claim, in the gold-rush sense, is a site with power already connected, and that is harder to buy.
I know the other side of this from my own family office. The Rosenzweig family office, proudly one of the smallest in the world, decided to acquire four Regus franchise centers as part of its investment portfolio. Regus rents offices, desks and meeting rooms by the month in shared buildings, and a franchisee runs a center under its name. We like everything about that business except paying the rent. The centers are still in their early years, and so far the rent has come to nearly 90% of what they take in and about half of everything we spend. I wish we owned the buildings. We would have bought them when commercial real estate was cheap and put the Regus centers inside. We were the operators with the concept and the system; the landlords owned the scarce complementary asset, a good location, and Teece's answer is that the owner of that asset takes a large share of the profit.
But the story is not so simple, because there is option value in what we did. Each landlord put up more than $1 million of its own money to build out a beautiful, AI-era center for us, and we repay it over the ten years of the lease, inside the rent. Had we bought the buildings, we would have paid for all of that build-out ourselves, plus the price of the buildings, plus a great deal more time, effort and borrowed capital, and we would have been in a different business, real estate with a Regus on the ground floor, with its own debts and its own empty-building risk. And a lease is an option: when it ends, we can walk away, and an owner cannot. The AI leases in this inquiry bundle the same things. The landlord's $10 million per megawatt is repaid through the rent, so part of what looks like a toll on scarce power is really the repayment of a construction loan, and the tenant that rents instead of building keeps its capital for chips and its freedom to leave. A landlord is paid for taking the risks the tenant declines, and the miners who became landlords took them with lenders willing to finance the buildings because a strong tenant's rent stood behind the debt.
Our Library, Vista's collection of models from economics and finance, writes Teece's model as a bargain. Let Pi be the profit the innovation earns once it is combined with the asset it needs, d_I what the innovator could earn without that particular asset owner, and d_A what the asset owner could earn without the innovator. Splitting what is left after each side's fallback, with w, a number between 0 and 1 for how much of the gain the innovator wins at the bargaining table (we avoid the usual letter, beta, because beta means something else later), gives the owner's share, 1 - w:
V_A = d_A + (1 - w) x (Pi - d_I - d_A)
where V_A is what the owner of the complementary asset keeps. The owner always gets its fallback, plus a share of the surplus that the combination creates. Scarcity sets the share. When powered sites are scarce, the tenant has nowhere else to go soon (d_I is small) and the owner can hold out for more (1 - w is large). Both raise V_A. We cannot measure d_I or w directly, so we do not solve the equation. Instead we read what owners actually get in the leases and ask whether it looks like a large share of Pi.
Now put the AI boom into those letters. The innovator is whoever has AI computers to run and needs a place to plug them in. Sometimes that is an AI lab, a company that builds AI models, such as OpenAI or Anthropic. Sometimes it is a cloud company, one that rents computing power to others over the internet, such as CoreWeave or Amazon Web Services. Either way, it has the chips, or the money to buy them, and what it lacks is power. The complementary asset is a site with power already connected to the grid, cooling, and fiber-optic lines for data, ready within a year or two. The industry counts that power in megawatts of computing load, or IT load (IT for information technology: the servers, storage and network), the power that actually reaches the servers; a site's gross power, what it draws from the grid once cooling and electrical losses are counted, is about 1.4 times as large, so computing load is about 70% of gross power (1 / 1.4 is about 0.7).
Pi is a profit, but the figure we can see is revenue, so we start there and take out the chips below. IREN owns both the building and the chips and sells the computing itself. Its five-year contract to supply Microsoft with AI computing, on about $5.8 billion of its own NVIDIA chips in Childress, Texas, is worth about $9.7 billion for 200 megawatts of computing load, so $9.7 billion / 200 megawatts / 5 years = $9.7 million per megawatt-year. CoreWeave's second-quarter revenue of $2.6 billion is about $10.3 billion a year. Divided by the 1,500 megawatts it ran at the end of June, that is $6.9 million per megawatt-year; divided by the roughly 1,000 it ran at the start, $10.3 million. It does not say which kind of megawatt it counts. d_A, the asset owner's fallback, is what the megawatt earns mining Bitcoin: about $1 million a year gross for each megawatt of the site's power, or about $1.4 million for the 1.4 megawatts that one megawatt of computing load needs (0.97 / 0.7), and less after power. d_I, the innovator's fallback, is to find or build power somewhere else, which takes time; how much time is one of the numbers we are missing (see "The data that we're missing"). Galaxy Digital, a crypto financial firm, argued in 2024 that miners with long-lead power equipment, equipment that must be ordered long in advance, had a head start.
What the landlords actually get is in the leases. Divide the total rent that each publicly traded former miner or crypto host (a company that housed other people's mining machines) has disclosed that an AI lease promises over its first term by its megawatts of computing load and its initial years, the lease's first term, before any options to renew:
rent per MW-year = contract value / IT MW / initial years
Across the 23 distinct leases we found, signed from June 2024 to August 2026, the answer runs from $1.24 million (Riot's first, small retrofit, a mining building refitted for AI, for AMD, the chip maker) to $2.67 million (Galaxy's second phase for CoreWeave), with a median of $1.86 million. That rent is 1.86 / 9.7 = about 19% of what IREN earns from Microsoft per megawatt, and 18% to 27% of CoreWeave's revenue per megawatt. A building, a substation (the equipment that brings the grid's high voltage down for the site) and a grid connection collect about a fifth of the revenue of the most coveted computers in the world.
Revenue flatters the computers' side, though, because the chips are so expensive. IREN expects about 85% of its Microsoft revenue to be left after power and running costs, before the cost of the chips is counted. The chips cost $5.8 billion / 200 megawatts = $29 million per megawatt, nearly three times what the building costs, and a five-year contract has to pay them back. What is left after the chips is 9.7 x 0.85 - 29 / 5 = about $2.45 million per megawatt-year. IREN owns its building, so that sum pays for both the building and the chips. A company with IREN's economics that rented its building instead, at the median rent, would hand the landlord 1.86 / 2.45 = about three quarters of it.
Now put the landlord on the same footing. A landlord spends about $10 million to build each megawatt of computing load, the middle of Hut 8's and Cipher's $9 million to $11 million. That money has a cost: the notes, bonds sold to investors, that finance these buildings pay about 6% to 9.9%, so charge the landlord 10% a year, $1 million. The rent of $1.86 million less that $1 million charge leaves the landlord about $0.9 million, and the $2.45 million the computers earn, less the rent, leaves the tenant about $0.6 million, so the landlord keeps about three fifths of the $1.5 million between them. That is a different measure from the three quarters above, which was rent as a share of the computers' profit; here we split the profit that remains once each side's capital is paid for. We charge the tenant nothing for the money tied up in its chips. Charged as the landlord is, 10% a year on chips worth $14.5 million on average, the tenant would show a loss of about $0.9 million, so the landlord's share would be larger still. The power owner collects a fifth of the revenue and the larger share of the profit. Teece would not be surprised.
Teece's model also says when that share should fall: when the asset stops being scarce. Is that a possibility? Yes, and it is moving in both directions. The rules for connecting are being rewritten. Texas's grid operator adopted a new batch process for large new users on June 2, 2026, with allocations due in spring 2027 and a final transmission plan in fall 2027. In June 2026 the federal regulator ordered all six of the other regional grid operators to justify or reform how they connect large users, and one regional operator, the Southwest Power Pool, already offers a 90-day process for large users that come with their own power plant. But the equipment is the real bottleneck. GE Vernova, a large maker of gas turbines, reported that its backlog and reserved production slots grew from 100 to 116 gigawatts in the second quarter of 2026; it expects to build 20 gigawatts a year this quarter, 24 in 2028 and 30 in 2030. Large transformers can take as long as four years to arrive. So the scarcity is real today and could ease later this decade, as turbines and transformers arrive and connection rules change. If utilities connect new large users faster, or gas-fired generators and high-voltage lines arrive, d_I rises and the landlord's share shrinks. That is the first of our tests.
The leases have a family resemblance. They are colocation leases: the landlord supplies the building, the connection to the grid and the cooling, and the tenant brings its own computers. The tenants are AI labs (Anthropic; Riot's "frontier AI lab", which Bloomberg reported to be Anthropic), cloud companies (CoreWeave; AWS, which is Amazon Web Services, Amazon's cloud business), Meta, Core42 (part of G42, an AI group) and Fluidstack, a middleman that rents computing to AI labs. The table shows eight of the 23. "Contract value" is the total rent over the first term. The details tell you what the tenants are paying for.
| Landlord, tenant (date) | Contract value | IT MW | Years | Rent per IT MW-year |
|---|---|---|---|---|
| Core Scientific, CoreWeave (2024 to 2025) | $10.2 billion | 590 | 12 | $1.44 million |
| TeraWulf, Core42 (December 2024) | more than $1 billion | 60 | 10 | $1.72 million |
| Applied Digital, CoreWeave (2025) | $11.0 billion | 400 | 15 | $1.83 million |
| Cipher Digital, AWS (November 2025) | $5.5 billion | 216 | 15 | $1.70 million |
| Hut 8, Fluidstack (December 2025) | $7.0 billion | 245 | 15 | $1.90 million |
| TeraWulf, Anthropic (July 2026) | $19.0 billion | 401 | 20 | $2.37 million |
| CleanSpark, Meta (July 2026) | $6.6 billion | 175 | 20 | $1.89 million |
| Riot, a frontier AI lab (August 2026) | $9.1 billion | 191 | 20 | $2.38 million |
| Median of 23 distinct leases | $1.86 million |
The landlord brings the building and the connection to the grid; the tenant brings the chips and, in most leases, pays the electricity bill. The landlord builds the empty shell, the power, the cooling and the network, at about $6 million (TeraWulf's first building, for Core42) to $15.4 million (Galaxy's Helios, its former mining campus in Texas, second phase) per megawatt of computing load on new builds. A mining site costs about $0.7 million to $1 million a megawatt (CoinShares' estimate), so an AI building costs roughly ten times as much. Riot's retrofit of a mining building cost about $3.6 million, which shows what reusing a mining building can save, and at Core Scientific CoreWeave paid for most of the work. In most leases the tenant pays for the power itself, and in many (the industry calls them triple net) it also pays the building's running costs. Cipher's lease with Amazon, Hut 8's and CleanSpark's are of this kind. So the landlord's income is close to its rent. On the large leases, the companies that disclose it expect average annual operating income, what is left of the rent after the landlord's own costs, of about $1.6 million to $2.2 million per megawatt of computing load, about the same as the rent.
And the tenants come with guarantors, larger companies that promise to pay the rent if the tenant does not. That matters because a landlord borrows billions to build, and its lenders want the rent to keep coming for 15 years. Google stands behind Fluidstack's leases at Cipher, TeraWulf and Hut 8: it promises up to $1.73 billion of Cipher's roughly $3.8 billion of rent and about $3.2 billion of TeraWulf's roughly $7 billion, and for all the rent at Hut 8's River Bend campus. At Cipher and TeraWulf it also took warrants in the landlords, rights to buy their shares for a cent each: at Cipher, about 5.4% of the company. Amazon guarantees the rent on Cipher's two AWS leases. A subsidiary of Meta is CleanSpark's tenant, with Meta's limited guarantee. Hut 8's second campus is leased to a technology company rated AA- or better by the credit agencies: among the safest borrowers, a few steps below the top grade of AAA. A miner used to sell its output into the most volatile market in finance. Now it collects rent from some of the strongest balance sheets in the world, for 10 to 25 years. That is a different business.
The trend in the rents matters for the test. The two leases signed in 2024 averaged about $1.6 million. The nine signed in 2025 had a median of about $1.9 million, and the twelve counted in 2026, as many more miners came to market with sites, about $1.9 million again. So far, more supply has not brought the rent down. Comparing leases of the same length, the picture is softer: 15-year leases signed this year have a median of about $1.76 million, against $1.87 million for those signed in 2025, about 6% lower. The 2026 median is held up by 20-year leases, and three of the five highest rents belong to 20 and 25-year leases. Many leases raise the rent about 3% a year, so the longer the lease, the more raised years go into its average. At the same starting rent, a 20-year lease averages about 8% more a year than a 15-year one.
So why has every miner not switched? Because switching is expensive, takes years, and cannot be undone, which is exactly the situation real options were invented to describe.
Why not invest as soon as a project is worth more than it costs? Because money spent on something that cannot be undone ends the chance to wait and see. If prices fall next year, a firm that waited can stay out; a firm that built cannot. That chance is worth something, and spending gives it up, so the project must beat its cost by a margin. Our Library's entry on the option to wait starts from Robert McDonald and Daniel Siegel (1986), who showed that for reasonable figures the project should be worth about twice its cost, and more the more uncertain its value. Later work, notably Steven Grenadier's in 2002, showed that when rivals can take the opportunity away, the margin a firm should wait for shrinks toward zero, because a rival may take the opportunity while it waits. Two things are unusual about the miners' version. The value of staying put falls by half on a known date. And the option to switch may not stay open, because the rent is high only while powered megawatts are scarce.
Here is the switch for one megawatt of a site's gross power, with every assumption stated. Think of it as a thought experiment with real prices; we set it out as cash and a break-even price, not as a formal option value. Converting costs $10 million per megawatt of computing load, the middle of Hut 8's and Cipher's $9 million to $11 million; a site's computing load is about 70% of its gross power (the same 1.4 as before, turned around), so that is $7 million per gross megawatt, spent over two years. The lease then pays operating income of about $1.9 million per computing megawatt, the median rent of $1.86 million, all of it income because the tenant pays the costs, or $1.3 million per gross megawatt, for 15 years. Staying in mining earns the September hashprice at 15 joules per terahash, pays $40 a megawatt-hour for power, the middle of the $28 to $54 the miners report, and replaces its machines every four years. Bitmain's S21+ hydro costs $2,864 for 5.37 kilowatts, so $2,864 / 5.37 = about $533 a kilowatt, $533,000 for a megawatt of machines, or about $133,000 a year over four years. And a miner switches its machines off rather than mine at a loss.
Cumulative cash is the running total of cash in less cash out. Mining's column stops growing after the April 2028 halving, because the halving leaves it almost nothing.
| Years from October 2026 | Cumulative cash, stay mining | Cumulative cash, convert |
|---|---|---|
| 2 | $0.73 million | minus $7.00 million |
| 4 | $0.74 million | minus $4.40 million |
| 6 | $0.75 million | minus $1.79 million |
| 8 | $0.75 million | $0.81 million |
| 12 | $0.75 million | $6.02 million |
| 16 | $0.75 million | $11.23 million |
Converting overtakes staying in the eighth year, undiscounted, and pays back its cost about seven and a half years after construction starts. A dollar that arrives years from now is worth less than a dollar today, because today's dollar could be earning interest meanwhile. Discounting at 8% shrinks each future dollar by 8% for every year of waiting, so a dollar due in ten years counts as about 46 cents today. Counted that way, and after paying for the building, converting is worth about $3.3 million per gross megawatt, against about $0.7 million for mining; at 12%, about $1.2 million against $0.6 million. Converting wins either way. The line for mining goes flat after April 2028 because at $40 power and today's price the halving leaves almost nothing over for the machines.
By McDonald and Siegel's rule, though, the switch is not an easy call. At 8%, what the lease brings in, in today's money, is about 1.5 times the building's cost per gross megawatt ($9.55 million against $6.24 million), short of the "about twice" the rule asks. At 12% it is about 1.2 times. The miners go ahead for the two reasons above: rivals can take the tenants, and the halving cuts the value of waiting on a known date.
Two qualifications keep this honest. Bitcoin's price could rise, which would lift mining, and the more miners leave, the lower the network's difficulty falls and the more each remaining machine earns. Bitcoin's design pays the miners who stay a little more for every one who leaves. Exits help the stayers. We can ask what price would make mining as good as the lease. If the lease's income must also pay a 10% annual charge on the $7 million it cost (the notes, bonds sold to investors, that financed these conversions pay about 6% to 9.9%), it leaves $1.3 million less $0.7 million, or $0.6 million a year per gross megawatt. Add the $133,000 a year for new machines to the $0.6 million, and spread the $735,000 over a year's 8,760 hours: about $84 a megawatt-hour on top of the $40 for power. Run the hashprice formula backward, (40 + 84) x 24 x 15 / 1,000, and that needs a hashprice of about $44.6. With nothing else changed, hashprice moves in step with Bitcoin's price, so that is Bitcoin at 83,479 x 44.6 / 39.87 = about $93,400 today, against about $83,500 on September 28, and about $186,000 after the halving, about twice, because the halving halves the coins.
Those who stay the course are making a bet of their own, and it is an option too. A miner that keeps its machines running holds, in effect, a call on Bitcoin's price: if the price rises, its revenue rises at once, and if the price falls, it can switch the machines off and lose only what they cost. The converter sells that call for a fixed rent. So the question is what Bitcoin's price would have to do for the stayer to come out ahead, and that can be priced.
Here is the same megawatt after the April 2028 halving at four prices, with difficulty and fees held where they are today, an assumption that flatters mining. At September's $83,479, a megawatt mining after the halving keeps nothing once power and machines are paid for. If Bitcoin doubles, to about $167,000, mining keeps about $0.49 million a year, which is today's thin margin: a doubling only undoes the halving. Standard Chartered's Geoff Kendrick, in December 2025, cut his bank's target for the end of 2028 from $500,000 to $300,000 and moved $500,000 to 2030. At $300,000, mining keeps about $1.27 million a year, almost exactly the lease's $1.30 million, with no $7 million building to pay for; at $500,000, about $2.44 million. Mining matches the lease net of its capital charge at about $186,000 and the lease's whole income at about $305,000. Run the cash out seventeen years with Bitcoin at $300,000 from the halving on, and staying is worth more in today's money than converting, about $6.3 million against $3.3 million at 8%, though converting ends ahead in plain dollars by year 17, because the halving of 2032 cuts the stayer's income again.
The catch is the word "unchanged." When mining pays, machines pour in and difficulty rises, so hashprice does not keep pace with Bitcoin's price. From the day before the 2024 halving to September 28 of this year, Bitcoin rose 31% and hashprice fell 62%. The stayer's bet on the price is a bet that the price outruns the machines, and history says the machines are fast. What a stayer keeps for certain is the coins it has already mined, and that is a different asset: a miner that holds thousands of bitcoin gains from a rise in the price whether or not its mining margin does.
I know this choice. In 1992, a Northwestern junior with B grades and a fear of not being hired, I sat on the fourth floor of the library during a blizzard and wondered what I would do if Price Waterhouse said no. "My Plan B: a nebulous and obscured set of thoughts related to monetizing scholarship," I wrote later in The Scholars' Treasure. Then, on a single day in 1993, Price Waterhouse offered me the job, and I realized "that both my Plan A and my Plan B were viable options." My Plan B became the company I founded in 1994. The miners have the same day in front of them, with a difference I envy: their Plan B comes with a guarantor that has a AA credit rating.
Teece's later work on dynamic capabilities asks whether a company can rebuild its advantage when its market shifts, and our Library gives two measures for it. One is the redeployment index, the share of a company's assets that moved; the other is simpler and fits here. The renewal share is:
N = revenue from businesses entered or rebuilt in the last five years / total revenue
Q2 is the second quarter, April to June. "Recurring" means rent that comes every month, as against one-time work, and "fit-out" is the one-time work of fitting out a building's inside for the tenant's equipment, which the tenant pays for.
| Company (latest filed quarter) | AI or computing revenue as a share of total |
|---|---|
| CoreWeave (Q2 2026) | 100% |
| Core Scientific (Q2 2026) | 83.2% |
| Applied Digital (quarter to August 2026) | 76.8% |
| TeraWulf (Q2 2026) | 71.3% |
| IREN (April to June 2026) | 51.4% |
| Riot Platforms (Q2 2026) | 13.3% (2.8% recurring) |
| Hut 8 (Q2 2026) | 7.0% |
| Cipher Digital (Q2 2026) | 0% (Amazon's rent began in August) |
| CleanSpark (quarter to June 2026) | 0% |
| Keel Infrastructure, formerly Bitfarms (Q2 2026) | 0% |
The table stands in for N. It counts all AI and computing revenue, so some of it is one-time work rather than rent. By the letter of N, CoreWeave's share would be lower, since its cloud began just over five years ago; we show 100% because all its revenue now comes from the business it built in place of mining. Applied Digital's 76.8% is mostly fit-out work it does for its tenants (rent alone is about 23%), and Riot's 13.3% includes fit-out too.
The table is a race photograph, and each runner has a story. Core Scientific now says it intends to convert "every megawatt in our portfolio" within three years, and in July signed leases with AMD, the chip maker, and an unnamed cloud company that Core Scientific says could bring in more than $14 billion over their first terms. IREN, which bought its first 248 AI chips in August 2023, is decommissioning its miners and aims to substantially finish by the end of this year; it supplies AI computing on its own chips instead of renting out buildings, and says recent three-year contracts bring in more than $20 million of revenue per megawatt of computing load. IREN does not say whether that is per year, though its claim of a two-year payback suggests it is. TeraWulf, whose CEO said in December 2024 that it would likely move its 250 megawatts of mining to AI computing, now mines with 145 megawatts against 839 megawatts of contracted computing load, of which about 102 were earning rent in July. Bitfarms said in February it would become Keel Infrastructure, its CEO, Ben Gagnon, saying "We are no longer a Bitcoin company"; the name took effect April 1, and by June 29 it had stopped mining at every American site without yet having signed a lease. It is still negotiating. Galaxy Digital, the crypto financial firm, stopped mining at its Helios campus in Texas at the end of the first quarter of 2025 to build for CoreWeave.
Two are keeping both plans. Hut 8 is leasing campuses to AI tenants while its majority-owned American Bitcoin keeps adding machines, and says its "current financials remain heavily dependent on the price of Bitcoin." And Riot, whose activist investor Starboard Value, a fund that buys a stake in a company to press its board for changes, called its first AMD lease "a small proof of concept deal" in February, signed a 20-year, $9.1 billion lease in August with what it called "a leading frontier AI lab", which Bloomberg reported to be Anthropic. CleanSpark will keep mining at Sandersville, Georgia, until the power is needed for Meta's subsidiary, whose rent is targeted to begin on November 30, 2027.
The stock market has been reading the same table. A stock's beta to Bitcoin says how far it tends to move when Bitcoin moves. A beta of 1 means that when Bitcoin rises 10% in a week, the stock tends to rise about 10% too; a beta of 0 means it does not follow Bitcoin at all. We measured each beta with a regression, the standard statistical way of finding the line that best fits two sets of numbers. That sentence reminds me of my four years in Northwestern's MMSS program, where we ran regressions like these numerous times a day, and for that we had to trudge across campus to the Vogelback Computing Center, often in a blizzard. Now I say "Hey Claude, run this regression" from the comfort of baby Solomon's nursery, when my wife isn't looking. Here that is the company's weekly price changes against Bitcoin's and, at the same time, against a fund that holds the big chip makers' shares (the VanEck Semiconductor fund, SMH). Doing both at once separates a stock that follows Bitcoin from one that follows the chip makers, since Bitcoin and chip stocks often move together (Inquiry No. 14). The standard error is the measurement's margin of doubt: the true beta could easily be 0.2 above or below the figure shown, and a figure a little below zero means no tie. We ran it for 2021 to 2023, before any of these leases, and for the past 52 weeks:
| Company | Bitcoin beta, 2021 to 2023 | Bitcoin beta, last 52 weeks | Its standard error | Chip beta, last 52 weeks |
|---|---|---|---|---|
| Core Scientific | minus 0.19 | 0.01 | 0.16 | 1.23 |
| Applied Digital | 1.03 | 0.09 | 0.23 | 1.61 |
| TeraWulf | 0.59 | minus 0.07 | 0.20 | 1.32 |
| Cipher Digital | 0.75 | 0.20 | 0.30 | 1.33 |
| Hut 8 | 1.20 | 0.17 | 0.21 | 1.34 |
| Keel (Bitfarms) | 1.06 | 0.25 | 0.28 | 1.38 |
| Riot Platforms | 1.10 | 0.48 | 0.21 | 1.15 |
| IREN | 1.00 | 0.51 | 0.23 | 1.54 |
| CleanSpark | 0.77 | 0.61 | 0.26 | 0.98 |
The companies that have signed AI leases or left mining, Core Scientific, Applied Digital, TeraWulf, Cipher, Hut 8 and Keel, now have little or no measurable tie to Bitcoin, even where the rent has not started; Hut 8 still mines through American Bitcoin and has lost most of its tie anyway. The ones still committed to mining at scale keep more of theirs: CleanSpark's barely moved, from 0.77 to 0.61, and Riot's halved but stays near 0.5. Each of the two also holds more than 11,000 bitcoin, worth roughly $1 billion at September's price, so its value moves with Bitcoin whatever it does with its buildings. IREN keeps its tie too, for now, because its miners were still running for most of the year. We leave out Galaxy, whose crypto trading dwarfs its Texas campus (it sold over $8 billion of digital assets last quarter); its beta of about 1.0 says nothing about the buildings. Core Scientific's figure for 2021 to 2023 rests on only 49 weeks of prices, because its stock was listed only in January 2022 and left the Nasdaq exchange when the company went bankrupt at the end of that year, and a year of weekly returns leaves a standard error of about 0.2 on each of these numbers, so only large differences mean much. The pattern is still plain enough to test. The market has done its own sorting.
What does the market pay for a megawatt under contract? A rough answer divides each company's enterprise value by its megawatts of AI computing load under signed leases, though many of the buildings are not yet built. Enterprise value is what it would cost to buy the whole company: the value of all its shares at today's price, plus its debts, less its cash, since a buyer takes on the debts and keeps the cash.
First, a yardstick. A 15-year lease paying $1.86 million a year per megawatt is worth 1.86 x 8.559 = about $15.9 million at an 8% discount rate. The 8.559 is what $1 a year for 15 years is worth today at 8%: about 93 cents for the first year's dollar, 86 cents for the second, and so on down to about 32 cents for the fifteenth, which add up to $8.56. That is about $5.9 million more than the $10 million it costs to build. Which yardstick applies depends on whether the building is paid for. Once the $10 million is borrowed and spent, it sits in enterprise value as debt, and the yardstick is $15.9 million. Before that, it is the $5.9 million left over.
| Company | Enterprise value | Contracted IT MW | Enterprise value per contracted MW |
|---|---|---|---|
| Core Scientific | $7.65 billion | 1,119 | $6.8 million |
| Applied Digital | $10.60 billion | 1,410 | $7.5 million |
| TeraWulf | $9.84 billion | 839 | $11.7 million |
| Hut 8 | $11.43 billion | 949 | $12.0 million |
| Galaxy | $6.81 billion | 526 | $12.9 million |
| Cipher Digital | $10.76 billion | 493 | $21.8 million |
| CleanSpark | $4.54 billion | 175 | $25.9 million |
| Riot Platforms | $7.33 billion | 241 | $30.4 million |
These are rough. The debt is as of each company's last filed quarter, several raised large sums since (CleanSpark's debt is counted as of June, before the $2.276 billion of notes it sold on September 25 to build for Meta), Galaxy and Riot hold other businesses inside the same value, and who pays for the build differs by deal.
The pattern worth noticing is at the bottom of the table. The companies with the most power not yet under lease, Riot with about a gigawatt, 1,000 megawatts, at Corsicana under a nonbinding letter of intent, a written plan to agree that binds neither side, and CleanSpark with Texas sites of several hundred megawatts, carry the most value per contracted megawatt. That is telling, though less than it looks. Their mining businesses, and in both cases more than 11,000 bitcoin, sit inside those values too; take the bitcoin out at September's price and Riot falls to about $26.5 million per contracted megawatt and CleanSpark to about $20.1 million, still well above the yardstick. CleanSpark mines on roughly 670 megawatts (42.6 exahashes a second at 15.8 joules per terahash), and our illustration values a megawatt of mining at under $1 million, so its mining may be worth a few hundred million dollars, part of the roughly $0.7 billion by which it still sits above $15.9 million per contracted megawatt once its bitcoin is taken out. One reading fits what is left: the market may already be paying for leases that have not been signed, which is a bet on our first test, that the rent for a powered megawatt holds.
Each of these could go wrong. Everything here rests on AI companies continuing to want sites, and that demand has a cycle. Inquiry No. 9 asked how long an AI chip earns its keep. The building will still be useful when today's chips are scrap, which comforts the landlord. The tenant must keep paying for 15 years whether or not its chips still earn their keep.
The tenant's own books are worth a look, because the landlord's rent depends on them. CoreWeave's second quarter of 2026 shows a business that earns a lot before its chips wear out and loses money after. Revenue was $2,575 million, up 112% from a year earlier, and adjusted EBITDA, its profit before the cost of the chips and interest, was $1,510 million, a 59% margin. But it recorded an operating loss of $49 million, paid $640 million in net interest, and lost $626 million, with about $35 billion of debt on its balance sheet and $6.4 billion spent on equipment in the quarter alone. Rent is not what decides this. Whether its chips earn back their cost before they wear out, and whether customers such as Microsoft, about two thirds of its revenue in 2025, keep paying, is. That is the tenant's risk, and a landlord shares it when it depends on one tenant.
Several leases rest on one intermediary, Fluidstack. At Cipher and TeraWulf, Google's guarantee of its rent has a dollar ceiling: if Fluidstack could not pay, Google would cover a bit under half the rent at each, and the landlord would need a new tenant for the rest. Core Scientific's biggest tenant is CoreWeave, and Core Scientific restated its results, reissued its past accounts with corrections, this year after finding it had kept mining assets it had already committed to demolish on its books at full value instead of writing them down, cutting their recorded value; its auditors found a material weakness in its controls. The builds are financed with secured notes, bonds backed by the buildings and their leases, at about 6% to 9.9%, which is fine while rents arrive on time and painful when a building is late, because interest is owed from the day the money is borrowed but rent begins only when the building is handed over. And the neighbors have a say. Matteo Benetton, Giovanni Compiani and Adair Morse found that crypto mining in upstate New York raised what small businesses and households paid for power by about $92 million and $204 million a year, and data centers are now meeting the same question: whether their demand raises the neighbors' bills.
The question has a growing literature, and the closest work deserves its credit first. Wei Wei and Yun Wan, in a working paper this summer, studied 24 first disclosures of AI deals by 13 listed miners and found that miners converting operating sites earned an abnormal return, the rise in the share price beyond what the market as a whole did, of about 23% in the three days around the news, against 4% to 9% for the whole sample, though their most cautious test cannot rule out chance; they call the gap between what a kilowatt-hour earns in AI and in mining the "Joule premium" (doi.org/10.2139/ssrn.7128321). A kilowatt-hour, the unit on a power bill, is 3.6 million joules. Galaxy Digital, a crypto financial firm that was converting its own Texas mining campus at the time, argued in December 2024 that miners with land, water, fiber and power equipment that must be ordered long in advance could convert, and that investors valued miners at 6 to 12 times a year's operating earnings, against 20 to 25 times for companies that run data centers, the buildings full of computers that serve the internet. In October 2024 CoinShares coined the nickname "mullet miners" for the hybrids, AI in front and mining in back, and put AI hosting, housing other companies' AI computers, at about $1.5 million per megawatt a year. Bernstein, a Wall Street research firm, was reported by DL News in October 2024 to value miners at $2 million to $4 million per megawatt, meaning the price of the company divided by its megawatts, against $30 million to $50 million for data centers, and to estimate that reusing a miner's power connection cuts a wait for the grid of up to four years by about three quarters. VanEck, an investment firm, modeled a fifth of miners' capacity moving to AI by 2027. Starboard Value, an investment fund pressing Riot's board for changes, wrote to it in February 2026 and cited four mining companies' AI deals, announced from August to December 2025, averaging about $1.8 million per megawatt of computing load. Alex de Vries noted in 2023, after Ethereum stopped paying miners, that "The devices previously used to mine Ethereum can still be repurposed", while Bitcoin's could not. And Luxor, whose Hashrate Index tracks what mining earns, already publishes what a megawatt-hour of mining earns each week.
The facts of the switch are public, and others have told it well: Wei and Wan measured the stock market's first reaction, Galaxy, CoinShares and Bernstein described and priced the conversions as they happened, Starboard Value averaged the rents on four companies' leases, Luxor publishes mining's revenue per megawatt-hour every week, and Voica, Panait and Iacob, three researchers writing in the journal Energies in January 2026, noted that when mining becomes AI, "the only reusable advantage" is access to electrical capacity. What we add is all 23 leases set beside mining's revenue on the same megawatt, with the arithmetic shown; Teece's model of profiting from innovation used to explain the size of the rent, a fifth of the revenue and the larger share of the profit; the switch set out as a real option whose payoff steps down on a known date, with its break-even Bitcoin price before and after the next halving and the stayer's own option priced at a bank's forecast; and the market's betas set beside each company's commitment to mining. We have not seen those in one place, or with dated tests attached.
For anyone who owns these companies, the first lesson is to value the megawatt before the hashrate, a miner's total guessing speed. A miner with power and a signed investment-grade lease, backed by a company the credit agencies rate a safe borrower, is a landlord with a long, rising income stream and a construction project to finish; a miner without one is a Bitcoin fund with machines. The betas say the market already makes that distinction, and the price per contracted megawatt suggests it may already be paying in advance for leases not yet signed.
The asymmetry runs both ways. A signed lease with a guarantor turns a volatile income into a dependable one, which is why some of the construction debt now carries an investment-grade rating (Moody's, one of the credit rating agencies, rates Hut 8's Beacon Point notes Baa2, one step above the lowest investment grade). But a company whose price already assumes its next gigawatt will lease at today's rents is exposed to exactly the thing Teece's model says can change, the scarcity of power.
Does any of this change if Bitcoin doubles? Less than you might think. A doubling by the halving restores today's thin mining margin and no more. Mining beats the lease, net of the lease's capital charge, only above about $186,000 after the halving, and matches the lease's whole income only near $305,000, both with difficulty unchanged, which history says it will not stay. A pure landlord has sold that upside for a fixed rent. A miner that keeps mining keeps it, at the price of the rent it forgoes.
What if you believe in the AI boom and are bullish on Bitcoin too? Then the asset to look for is a powered megawatt that is not yet leased, held by a company that also keeps its coins. An unleased megawatt is a two-way option: lease it if the rent holds, mine it if Bitcoin spikes. In our table, Riot, with about a gigawatt at Corsicana and more than 11,000 bitcoin, and CleanSpark, with its Texas sites and about 13,700 bitcoin, hold that option, and Hut 8 keeps a foot in both camps through American Bitcoin. The high values the market puts on Riot's and CleanSpark's contracted megawatts may be the price of that option as much as a bet on unsigned leases. The pure landlords, Core Scientific, Applied Digital, TeraWulf and Cipher, have sold the Bitcoin call for 12 to 20 years of rent, and IREN sold its miners. Could the miners who kept the megawatt redeem the old gold miners if Bitcoin jumps? A few could, which is what the gold rush taught as well: a few struck it rich, most made a little, and the ones who did best owned what everyone else needed. If the price spikes, the machines will follow it, and the lasting gain goes to the company that owns cheap power, good machines and the coins it already holds. None of this is a recommendation. It is a map of who holds which option.
If the scarcity eases, who gains and who loses? The tenants, such as Amazon, Meta and Alphabet, would pay less rent, though rent is a small part of what AI computing costs. The landlords priced for leases they have not yet signed have the furthest to fall, as the price table shows; landlords with long guaranteed leases already signed, such as Core Scientific and Applied Digital, are protected for 12 to 15 years. The makers of the cure, gas turbines, transformers and grid equipment, are paid in both worlds: while power is scarce their order books are full, and an easing is those orders being delivered. Owners of existing power plants that sell into the market are paid by scarcity, because new supply lowers the price of power, while regulated utilities, which earn a return on what they build, are helped by the build-out. That is a map of roles, not a recommendation, and it says nothing about what any of them is worth.
The idea travels, too. Wherever a stranded industrial site holds a large grid connection, a closed smelter, a mill or a retired power plant, the same arithmetic applies, and the same question about how long the scarcity lasts. For shareholders of the tenants, such as Amazon, Meta and Alphabet, part of the rent is a price paid to start sooner, and its trend is a measure of how tight power remains.
And the next halving is a signpost with a date. Around April 2028, every megawatt still mining will take in about half as much and, after power and machines, keep almost nothing, unless Bitcoin's price has risen to meet it.
The people who negotiated these leases and built these sites know what the filings cannot say, and finding and asking them is what Vista is really about. We would most like to hear from these.
In general, what did tenants pay for, the megawatts, the date the power was ready, the land or the building, and how did the rent move as more miners offered sites?
The answer that would change the view"Rents are already falling; tenants now have five sites to choose from for every one they need."
In general, what was the walk-away price, how many months of delay would a tenant pay to avoid, and how much of the rent was paying for speed?
The answer that would change the view"We paid for speed, and the speed premium ends when the new gas plants and transmission lines arrive."
In general, how long does a large new load wait to connect today, and how much of the miners' advantage is a place in that queue that others cannot buy?
The answer that would change the view"Interconnection reform will cut the wait to about a year, and the head start will disappear."
In general, what of the mining site survived the conversion, and what did it cost per megawatt compared with building from bare land?
The answer that would change the view"Almost nothing but the substation survives; the miners' sites save months, not millions."
Every conversation would draw on past, general experience only: no current employees of the companies named, nothing confidential, and every Advisor screened for conflicts before a word is said. For a client engagement, Vista would combine this inquiry with those interviews: two or three Advisors, structured conversations, and one written brief that says where they agree, where they split, and what would change the answer.
We will score this inquiry on October 31, 2027, from the companies' filings and weekly closing prices through September 30, 2027, and publish the result whether it flatters us or not.
What the outcomes would mean. If both tests pass, the powered megawatt is the asset, and the market knows it. If only the rent test passes, the megawatt holds its value but the market has not yet sorted the miners by what they do. If only the market test passes, investors have sorted the miners but the rents are falling, which would say the scarcity was a passing thing. If neither passes, the scarcity was passing, and the market never priced it.
The bet. That a powered, connected megawatt stays scarce long enough for the miners who own one to collect a rent a third above what mining grosses on the same megawatt, about twice what mining keeps after power, and nearly three times what it keeps after power and new machines, for 10 to 25 years, from tenants backed by the strongest companies in technology.
The payoff. On our illustration, a converted megawatt is worth about $3.3 million more than it cost at an 8% discount rate, against about $0.7 million for mining it; a contracted megawatt is worth about $15.9 million against a $10 million build. If rents fall toward mining's level, the payoff shrinks to little or nothing over the cost of the building, and the companies priced for leases not yet signed have the furthest to fall.
Our read. The evidence is on the landlords' side today: rents have held as more sites came to market, though 15-year leases signed this year came in about 6% below last year's; the tenants are strong; and the halving makes the alternative worse on a known date. The market has already re-sorted the miners by what they do. The other side of the bet belongs to the miner that keeps a megawatt unleased and holds its coins: it has sold nothing, and it is paid only if Bitcoin's price outruns the machines. By McDonald and Siegel's rule the switch is a closer call than the headline suggests, at about one and a half times its cost against the "about twice" the rule asks, and my own family office's Regus centers are a reminder that the owner of the scarce asset has a hard-won advantage and that owning is a different business.
What settles it, and when. The rents on the next year's leases, the speed of grid interconnection, and the next halving, around April 2028, together with Bitcoin's price when it comes, measured against two marks at today's difficulty: about $186,000, where mining after the halving matches the lease net of its capital charge, and about $305,000, where it matches the lease's whole income. We will score the two tests on October 31, 2027, after talking first to the people who negotiated these leases and the planners who decide how fast new power arrives.
In Graham Greene's novel The Power and the Glory, a priest keeps his calling after the world has taken away his church. The phrase comes from the doxology, the line of praise that later manuscripts add to the Lord's Prayer in Matthew 6:13: "For Yours is the kingdom and the power and the glory, forever." The miners have done the reverse. They are giving up the calling and keeping the power. The glory, for now, belongs to whoever has the electricity. Unless Bitcoin's price spikes, in which case the ones who kept a megawatt unleased will keep their calling too, and the power besides.
Because the most valuable thing a miner owns turned out to be its connection to the electric grid. In April 2024 Bitcoin's code cut the coins it pays miners in half, and the revenue of a modern mining megawatt fell from about $2.6 million a year to about $1 million. In the same two years the AI boom made a powered, connected megawatt scarce: in Texas alone about 410 gigawatts of large new users were waiting to connect in March 2026. AI labs and cloud companies now pay former miners to plug their computers into sites that already have power.
Across 23 distinct leases signed from June 2024 to August 2026 by publicly traded former miners and hosts, the rent runs from $1.24 million to $2.67 million a year for each megawatt of computing load, with a median of $1.86 million. A site needs about 1.4 megawatts from the grid for each megawatt that reaches the computers, so per megawatt of site power the rent is about $1.3 million: a third more than mining grosses on the same megawatt, about twice what mining keeps after paying for power, and nearly three times what it keeps after replacing its machines too.
Hashprice, the dollars a petahash per second of computing earns each day, fell from about $105 the day before the halving to $39.87 on September 28, 2026, a drop of 62%. The reward per block fell to 0.447 of its old size, difficulty rose so each machine's share fell to 0.651 of what it had been, and Bitcoin's price rose 31%; multiplied together, 0.380. At 15 joules per terahash that is $110.75 per megawatt-hour, against power costs of $28 to $54. The next halving, around April 2028, would cut it to about $56 at today's price and difficulty.
On our illustration, converting a megawatt of a site's gross power costs about $7 million over two years and then earns about $1.3 million a year for 15 years; it pays back about seven and a half years after construction starts and is worth about $3.3 million at an 8% discount rate, against about $0.7 million for staying in mining. By McDonald and Siegel's rule for irreversible investments the margin is thinner than it looks, about 1.5 times the cost against the "about twice" the rule asks, and the miners go ahead because rivals can take the tenants and the halving cuts the value of waiting on a known date. Mining would need Bitcoin near $93,000 today and near $186,000 after the halving to match the lease, net of its capital charge, with difficulty unchanged.
A miner that keeps mining holds a call on Bitcoin's price, and the converter sells that call for a fixed rent. After the April 2028 halving, with difficulty and fees unchanged, a doubling of the price to about $167,000 only restores today's thin margin of about $0.49 million per gross megawatt a year. At Standard Chartered's end of 2028 target of $300,000 mining would keep about $1.27 million, about the lease's whole income, with no building to pay for; at $500,000, about $2.44 million. The catch is that difficulty does not stay unchanged when mining pays: from April 2024 to September 2026 Bitcoin rose 31% and hashprice fell 62%.
Yes. Regressing each company's weekly return on Bitcoin and on chip stocks together, the miners that have signed AI leases or left mining, Core Scientific, Applied Digital, TeraWulf, Cipher, Hut 8 and Keel, now have little or no measurable tie to Bitcoin, while CleanSpark and Riot, still mining at scale and each holding more than 11,000 bitcoin, keep more of theirs. The inquiry's second test asks whether IREN's Bitcoin beta falls below 0.2 as it leaves mining while CleanSpark's stays above 0.4, scored on October 31, 2027.
Every figure in this inquiry is computed in code from published sources. Mining revenue comes from Luxor's hashprice and the formula above, cross-checked against block data; the halving's split comes from blocks before April 20, 2024 and before September 28, 2026. The rents divide each lease's disclosed initial-term value by its megawatts of computing load and its initial years; extensions are excluded, escalators are included where the filings say so, and tranches, parts of one deal signed in stages, and later-year follow-ons are left out of the median to avoid counting a lease twice. The switch is an illustration with the assumptions stated in the text; the stayer's bet holds difficulty and fees at their September 2026 levels and takes Standard Chartered's targets from its December 9, 2025 note as reported by CoinDesk. Betas are from weekly returns on Friday closes, regressed on Bitcoin and on the VanEck Semiconductor fund together. Enterprise values are from a licensed market data feed as of October 7, 2026. The Regus percentages come from the two centers' own books for January to September 2026. The calculations are in model.py, betas.py and valuation.py, and the facts behind each input, with links to the filings, are kept with the inquiry's working files.
| Input | Value | Source |
|---|---|---|
| Hashprice before the halving and on September 28, 2026 | $105 and $39.87 per PH/s per day | Luxor Hashrate Index; cross-checked against block data (mempool.space) and Coinbase Exchange prices |
| Fleet efficiency | 15 joules per terahash | Bitmain S21+ Hyd. and S21 Pro listings; IREN's stated fleet |
| Power cost | $40 per MWh (filed range $28 to $54) | TeraWulf, CleanSpark and peer filings |
| Machine cost and life | $2,864 for 5.37 kW; replaced every four years | Bitmain shop, October 7, 2026 |
| Computing load as a share of site power | 0.70 (1.4 gross MW per IT MW) | Filings: 0.65 to 0.83; Hut 8 River Bend 1.35 PUE |
| Conversion cost | $10 million per IT MW ($7 million per gross MW), over two years | Hut 8 and Cipher guidance, $9 million to $11 million per IT MW |
| Lease income | $1.86 million per IT MW-year for 15 years | Median of 23 distinct leases; Hut 8's stated NOI |
| Capital charge on the building | 10% a year | Assumption; the builds' notes pay about 6% to 9.9% |
| Bitcoin price | $83,479 on September 28, 2026 | Coinbase Exchange |
| Next halving | Block 1,050,000, about 553 days from October 8, 2026 | Bitcoin's subsidy rule; 600 seconds a block |
| IREN and Microsoft | $9.7 billion, 200 IT MW, five years; chips about $29 million per MW; 85% margin before chips | IREN 8-K of November 3, 2025 and investor deck |
| Standard Chartered targets | $300,000 for end of 2028; $500,000 for 2030 | Geoff Kendrick, December 9, 2025, as reported by CoinDesk |
| Betas | Weekly log returns on Friday closes, 2021 to 2023 and the 52 weeks to October 2, 2026, on Bitcoin and SMH together | Licensed market data (aggregates only) and Coinbase Exchange |
| Enterprise values | As of October 7, 2026; debt as of the latest filed quarter | Licensed market data feed and filings |
| Regus centers | Rent nearly 90% of revenue and about half of all costs, four centers in their early years | The family office's own books, 2026 |
The model (model.py), its output and the facts files behind each input are kept with the inquiry's working files; the arithmetic in the text was checked against them.
This inquiry is research, not investment advice. It states what the filings and models show under labeled assumptions; nothing here is a recommendation to buy, sell or hold any security, and the decision belongs to the reader. As of October 4, 2026, Russ W. Rosenzweig, Vista's founder, owns shares of Riot Platforms, CoreWeave, Microsoft, Amazon, Alphabet, Meta Platforms, AMD, Coinbase, NVIDIA, Dell Technologies and GE Vernova, and some Ether, and holds options on CoreWeave, Microsoft, Amazon, Alphabet, Meta Platforms, AMD, Coinbase, NVIDIA and the iShares Bitcoin Trust; he owns no shares of IREN, Core Scientific, TeraWulf, Cipher Digital, Hut 8, CleanSpark, Keel Infrastructure, Galaxy Digital, Applied Digital or Circle. The Rosenzweig family office also holds an interest in two franchised Regus centers, described in the text. Holdings through mutual funds and exchange-traded funds are not counted. How this inquiry was made: written by Russ W. Rosenzweig with Vista's AI research desk. It began with two questions I asked the desk on October 7, 2026, while finishing Inquiry No. 14; the facts were gathered from the companies' filings, the models were built and every figure computed in code, each public source was checked against its original, and the draft was fact-checked before I read it. I read it and listened to it before it was published. Why I write with AI. Why I publish these inquiries.