In August, Core Scientific paid $444.3 million for a company that owned about 40 acres next door to its data center in Muskogee, Oklahoma. The deal included a substation and agreements covering up to 440 megawatts of utility power. The seller also stood to earn a bonus.
If another 40 megawatts of firm electrical capacity became available before December 31, the seller would collect an extra $40 million. At the end of September, that condition was still open.
I’m used to hearing miners talk about power in cents per kilowatt-hour. Shave a penny off the price and thousands of machines get cheaper to run, hour after hour. The Oklahoma deal put a different question on the table.
If you have an expensive cluster of AI servers ready to install, how much would you pay to turn them on sooner?
Core Scientific’s November 5 filing showed just 32 bitcoin left on September 30. It had sold 3,374 BTC over nine months, including coins it had accumulated in earlier years. Meanwhile, one AI infrastructure customer accounted for practically the entire business. That customer had already tried to buy the company.
What, exactly, had a Bitcoin miner built that an AI company wanted to own?
Core Scientific began 2026 with 2,537 BTC. Over the next nine months, it mined about 869, giving it roughly 3,406 bitcoin to work with. It sold 3,374 and finished September with 32, worth about $2.7 million. The sales raised $286.8 million to help cover planned capital spending and other cash needs.
The filing for the March quarter caught that decision in progress. By then it had sold 2,385 BTC and still held 547. Investors reading the November results were seeing the cumulative effect of a treasury policy that had been underway since the beginning of the year.
Consider the choice facing a miner that holds its own production. Keeping the coins gives shareholders exposure to a future rise in bitcoin. It also leaves the company needing dollars for equipment, contractors, and construction. Management can borrow those dollars, raise them from shareholders, or sell an asset it already owns. Keeping every coin can be an expensive expression of conviction when the interest payments come due before the building opens.
Core Scientific chose to sell. That decision reveals management’s financing priorities. Turning it into a prediction about bitcoin’s price would require knowing what management expected the coins to earn relative to the alternatives. A company can be optimistic about an asset and still decide that its next dollar belongs in a project with a paying customer.
In the third quarter of 2025, mining its own bitcoin generated $57.4 million in revenue. Renting out data-center space and power for intensive computing, a service called colocation, generated $15 million. A year later, Core Scientific still operated its own miners at two sites. Its accounts for the September quarter show how little of the company those machines now represented.
The self-mining operation recorded about $19.8 million in costs against $12.3 million of revenue. Those costs include depreciation on the machines as well as the electricity they used. Across the same three months, the colocation business earned about 60 cents of gross profit per revenue dollar. An executive deciding what to do with the next available building had good reason to take a call from an AI tenant.
Larry Ellison wanted more chips. Congress wants answers about a possible intelligence operation. Both went straight to the same man.
Colocation accounted for 92.5% of Core Scientific’s $232.9 million in quarterly revenue. The bill had two large line items. Core Scientific passed $57.6 million in electricity costs through to the customer without a markup. Another $157.3 million came from license fees for the facilities. The remainder covered maintenance and other services.
How does an AI landlord make money if it supplies electricity at cost? Its customer pays for a place where hundreds of racks can draw that power reliably. The license fee captures some of the value of having the building, the connection, and the people who can keep the site running. If electricity prices rise, the pass-through can lift reported sales while adding little to gross profit.
A Bitcoin mining chip is very good at calculating the hashes that give its operator a chance to earn bitcoin. The chips a customer uses to train an AI model have different jobs, and the network linking them matters enormously. In a securities filing, CoreWeave estimated that one possible cluster of 32,000 GPUs could require around 600 miles of fiber cable. Those connections let many machines work on parts of the same job. The power distribution and cooling also have to support densely packed racks.
A parcel with high-capacity power already in service can spare a developer part of the wait for a grid connection. The International Energy Agency estimates that building a new transmission line in an advanced economy can take four to eight years. Existing power can move a site up a tenant’s list. Whether the site stays there depends on the cost of adding the cooling, fiber, and reliability the customer requires.
In the first quarter of 2026, Core Scientific wrote down $266.5 million of mining equipment and infrastructure as mining economics deteriorated. By August, it was prepared to pay that premium for additional power in Muskogee. A building full of machines can lose value while the substation outside attracts a new bidder. That possibility barely registers if you value the entire site by how much bitcoin it produces.
Someone who already knew the buildings had done the calculation from the other side. CoreWeave worked out what it might save if it owned the place. Then it put that number in a takeover offer.