Bitcoin Miners Became Landlords, and AI Signed the Lease

• This is an exit from mining rather than a diversification into AI. Core Scientific booked 83% of Q2 2026 revenue from colocation, Keel (formerly Bitfarms) is winding mining down across 2026 and 2027, and Bitdeer began dismantling a Washington mine instead of retrofitting it.
• Almost nothing above the transformer survives a conversion. SHA-256 ASICs have no path into AI workloads, and the build cost jumps from roughly $700,000 to $1 million per megawatt to $8 million to $15 million. What carries over is the interconnection agreement, the substation, the fiber, and the land.
• The financing gives away the business model. Hut 8’s $4.25 billion Beacon Point notes priced at Baa2, non-recourse and fully amortizing, before the building exists. Ratings that strong on construction-stage assets only work if the agencies are underwriting the tenant’s credit.
• Two very different trades are being reported as one. Colocation landlords hold credit-wrapped 15-year rent on a multi-decade asset. AI cloud operators hold depreciating silicon and monthly utilization risk, and Bitdeer’s entire ~$76 million AI Cloud ARR is around 4% of Hut 8’s expected annual rent NOI.
• The headline numbers deserve less weight than they get. Announced contract values are gross, undiscounted, and stretched across 15 to 20 years, and the counterparty base behind the sector’s $70 billion is a short list dominated by Anthropic, Fluidstack, CoreWeave, and AMD.

On 1 April 2026, Bitfarms stopped being Bitfarms. The company redomiciled from Ontario to Delaware, moved its principal office to lower Manhattan, and started trading as Keel Infrastructure under the ticker KEEL. Ben Gagnon described the result as a pure-play infrastructure developer and owner, and the filing lists index eligibility and standing with utility partners among the reasons for the move. A company built to run hashing machines removed both halves of its name and rewrote itself as a real estate and power business.

That is the honest version of what happened across public bitcoin mining between the April 2024 halving and now. The sector did not add artificial intelligence as a second revenue line while carrying on mining. It exited mining, kept the interconnection agreements, and started leasing them to people with better credit. Core Scientific’s second quarter of 2026shows the arithmetic plainly, with high-density colocation delivering $136.7 million of $164.2 million in revenue, more than 83% of the total, while self-mining fell to $21.5 million from $62.4 million a year earlier. The mining business that gave the company its balance sheet, its grid relationships, and its name is now a rounding error inside it.

The cost curve settled the argument before anyone made it

CoinShares put the weighted average cash cost to produce one bitcoin among listed miners at roughly $79,995 in the fourth quarter of 2025, against a coin that spent the first quarter of 2026 trading between $68,000 and $70,000, which works out to about $19,000 of loss on every coin produced before anyone touches depreciation, interest, or the cost of the next fleet upgrade. Hashprice bottomed near $28 to $30 per petahash per day in early March, a post-halving low, and at that level a miner running mid-generation hardware needs power below $0.05/kWh to stay cash-positive. The number of North American sites that clear that bar without a legacy power contract signed years ago is small, and shrinking.

Buy bitcoin and stocks. Signup to Coinbase today

Hashrate peaked around 1,160 EH/s in early October 2025 and fell to roughly 920 EH/s, with three consecutive negative difficulty adjustments, the first such streak since July 2022. Machines came off the wall because keeping them on the wall cost more than the coin they produced.

In previous drawdowns that would have been the whole story, with unprofitable operators selling rigs to better-capitalized ones and the hardware carrying on hashing somewhere cheaper. This time the site itself changed function, and the capital allocation question stopped being who should own the ASICs and became what the substation should be doing instead.

Nothing above the transformer survives the conversion

The received wisdom that mining sites are natural AI data centers gets the asset backward. A SHA-256 ASIC is a fixed-function chip with no path to matrix math, so it has no resale value into AI and no salvage value into anything else once hashprice falls below its operating cost. The racks, the immersion tanks, the ambient-air sheds, the power distribution, and in most cases the building envelope go to scrap.

Everything below the IT load survives, which is the part that takes four to seven years to replicate from scratch. The signed and energized interconnection agreement, the substation, the transmission spur, the water rights, the fiber, the local permits, and the land. Keel’s own portfolio description makes the pitch without embarrassment, leading with 2.2 gigawatts of pipeline and established grid interconnections already in place, and mentioning no hardware at all.

The capital numbers show how little of the old plant carries over, with mining infrastructure running roughly $700,000 to $1 million per megawatt against $8 million to $15 million per megawatt for AI-ready capacity. Ten to fifteen times the build cost on the same piece of dirt. Rack density explains most of that gap, because a mining hall distributes power at a handful of kilowatts per rack across a wide, cheap footprint, while Keel’s converted 18 MW Washington site is engineered for 190 kW per rack with liquid cooling and a target PUE between 1.2 and 1.3, built against Nvidia GB300 reference designs. Nvidia’s Kyber rack for Rubin Ultra, due in 2027, draws 600 kW in a single cabinet. Anything being poured today is already designed against a moving target.

In Wenatchee, Bitdeer began dismantling a mine originally deployed in 2018 rather than retrofitting around it, clearing the site for a 13 MW AI facility targeted for late 2026 into early 2027. Demolishing functional plant only makes sense on the view that the plant was never what the company owned.

Mining operations tolerate downtime by design, since curtailing during a price spike is a revenue strategy rather than a failure, and a rig that trips offline for six hours costs a few thousand dollars in foregone block reward. A 15-year AI lease inverts that completely, with service level credits, concurrent maintainability requirements, and a tenant whose training run does not care that ERCOT was paying well that afternoon. Teams that spent five years optimizing for cheap and interruptible are being asked to deliver expensive and continuous, and that is the part of the transition where execution risk sits.

The rating agencies are underwriting the tenant

Hut 8 closed $4.25 billion of senior secured notes for its Beacon Point project in June 2026, rated Baa2, fully amortizing, non-recourse to the parent, and priced 20 basis points inside its own River Bend issuance from six months earlier. J.P. Morgan led the book. That brought cumulative project-level investment-grade construction financing at Hut 8 to $7.5 billion, which is an unusual sentence to write about a company that was a bitcoin miner in 2024.

Investment-grade paper on a construction-stage data center with no operating history only prices that way if the agencies are looking straight through the sponsor to the counterparty. River Bend’s 245 MW lease sits with Fluidstack and is backed by a Google financial guarantee covering the lease payments. TeraWulf’s 20-year Anthropic lease at the Justified Data campus in Hawesville, Kentucky, worth roughly $19 billion of contracted revenue across 401 MW of critical IT load, is described in TeraWulf’s own release as expected to be supported by an investment-grade credit. What is being rated in each case is the tenant’s promise to pay, dressed in a project SPV and secured against a building the sponsor has yet to finish.

The future of finance is here. Trade crypto and more

Which determines what these businesses now earn, because a landlord with a credit-wrapped 15-year lease captures a development spread and a construction risk premium, collects contracted rent, and is largely insulated from what the tenant’s compute is worth in year eight. Hut 8’s 949 MW of contracted capacity carries $26.6 billion of base-term value and more than $1.75 billion of expected average annual NOI, and none of those numbers move with hashprice, GPU spot pricing, or model training demand. They move with delivery and with counterparty solvency.

Miners with secured HPC contracts trade around 12.3 times next-twelve-month sales against 5.9 times for pure-play miners, which is a doubling for the same megawatts on the same land under different paper. Once that spread opened, every board with an interconnection queue position had a fiduciary argument for conversion whether or not anyone in the room believed in AI.

Two different bets are being called the same pivot

Hut 8, TeraWulf, and Core Scientific are running the landlord version, where the company delivers the shell, the electrical plant, and the cooling, signs a long lease, and lets the tenant bring or finance the silicon. The rent is contracted, the asset has a multi-decade life, and the operator’s exposure to what a GPU-hour is worth in 2031 rounds to nothing.

Bitdeer is running a different bet, and its own disclosure shows the size of the gap. The company’s June 2026 operations update reports AI Cloud annualized recurring revenue of roughly $76 million at 95% utilization, across 4,248 deployed GPUs with 3,517 under external subscription. That business is growing fast and running hot, and the entire annualized run-rate still comes to about 4% of the average annual NOI Hut 8 expects from rent alone. Bitdeer owns the silicon, depreciates it against a three-to-five-year useful life, carries utilization risk every month, and reprices into whatever the GPU rental market does next, where Hut 8 owns a building and a guaranteed rent roll and carries neither.

No one has raised investment-grade paper against a GPU fleet with no contracted offtake, and no one is going to. Credit markets have already told these companies which version of the pivot they will fund at scale, which is why the largest commitments in the sector are colocation leases rather than cloud buildouts, and why the cloud operators in this group are the ones still carrying meaningful equity risk.

Keel’s own hedging reflects the same read, since Gagnon flagged GPU-as-a-Service or cloud monetization specifically at Washington, an 18 MW site he described as less than 1% of the total developable portfolio, while claiming that converting that one site could produce more net operating income than the company ever generated from bitcoin mining. Two things sit inside that. The unit economics of AI compute dwarf mining by an order of magnitude, which is true and roughly the whole story of the last two years. And the company is willing to test the higher-variance version of the business on 1% of its footprint while building the landlord version across the rest.

Selling the asset the business was built to accumulate

Public miners have taken more than 15,000 BTC off their treasuries from peak levels, and the disclosures around those sales are more revealing than the totals. Bitdeer held 150 BTC at the end of June 2026, against 1,502 a year earlier. Riot sold 1,818 BTC for $162 million in December. Core Scientific sold roughly 1,900 BTC for $175 million in January and signaled it would liquidate substantially all of what remained. Marathon, still the largest public holder at 53,822 BTC, quietly widened its treasury policy in its March 10-K to authorize sales from the entire balance sheet reserve, under pressure from a $350 million bitcoin-backed credit facility whose loan-to-value ratio had climbed to 87% as prices fell.

Two theses lived inside these companies for a decade, one being that bitcoin is a superior reserve asset worth holding through drawdowns, the other being that mining it is a viable industrial business. Both were financed by the same balance sheet, and a collateral squeeze at 87% LTV settled which one was load-bearing. The treasury was working capital the whole time.

There is a circularity here that the sector has been reluctant to say out loud, because the companies liquidating bitcoin to fund AI conversions are the same companies whose hashrate secures the ledger those coins settle on. Every megawatt converted reduces the security budget and increases the incentive for the next operator to convert, since falling difficulty temporarily improves economics for whoever stays. That feedback loop is stable at current scale and gets less comfortable if the trend runs another two years.

New Scam Prevention Course. Never make a rushed crypto decision again.

The grid number nobody at ERCOT will defend

ERCOT’s preliminary long-term forecast filed with the PUCT in April 2026 projects approximately 367,790 MW of demand in the region by 2032, against an all-time peak of 85,508 MW set on 10 August 2023. More than four times the record, inside six years.

ERCOT does not believe it either. Pablo Vegas said in the same release that the operator believes the forecast to be higher than expected future load growth, and that it looked forward to working with the commission on refining how it ascertains load information. The number is assembled from what transmission and distribution utilities report about medium and large load customers, which means data centers, crypto miners, industrial users, and oil and gas processes tell their utility what they intend to build, and the utility passes it up. Nothing in that process penalizes a developer for reserving capacity at six sites to build one, and nothing deduplicates a single project being shopped through three different interconnection requests.

That gap between requested load and real load is exactly why the miners’ asset is worth what it is worth. Interconnection queues across the major US markets are congested with speculative and optioned positions, so a signed, energized, physically delivered interconnect is scarce in a way that a queue position is not. Miners hold those because they got in line between 2018 and 2021, when the only competition for stranded and interruptible power was other miners. They are selling scarcity they acquired cheaply, in a market that now prices it as the binding constraint on the entire AI buildout.

CleanSpark’s greenfield approach confirms the logic from the other direction, since the company bought 271 acres in Austin County, Texas together with 285 MW of contracted power, with initial energization of more than 200 MW targeted for the first half of 2027. Even a greenfield build in this sector is bought power-first, with the acreage priced as an accessory to the interconnect.

Something real is being lost in the swap, and grid planners have not priced it. Mining load is genuinely interruptible, because curtailing into scarcity pricing is profitable rather than merely tolerable, which made miners one of the few large flexible loads on ERCOT and a useful buffer during summer peaks. A multi-week training run does not curtail on those terms, and the lease paying for the building says as much in its availability commitments. Megawatts are moving from the most responsive category of large load to something near the least responsive, at the same time as the aggregate is growing. Any resource adequacy assumption that treated crypto load as dispatchable demand response stops holding for a site the moment it signs an AI lease.

What breaks first

Most risk write-ups on this trade lead with an AI demand slowdown, and that is the least interesting failure mode available. A 15-year lease guaranteed by Google, or backstopped by an investment-grade credit, does not reprice because training demand hits an air pocket in 2028. The tenant absorbs that, which is the entire reason the rent was underwritten at Baa2.

The failures worth watching are narrower and closer, starting with delivery, because every one of these leases carries a commencement date and the schedule runs through transformers, high-voltage switchgear, chillers, and grid upgrades on lead times that have not improved. A conversion that slips two quarters against a fixed rent start is a materially different investment from one that lands on time, and none of the announced contract values disclose the penalty structure.

Concentration is the second, and it is underappreciated because the headline number obscures it. Something over $70 billion of announced AI and HPC contracts across the sector traces back to a short list of names, with Anthropic, Fluidstack, CoreWeave, and AMD accounting for a large share of the total. Fluidstack in particular sits on both sides of the table, as the tenant at Hut 8’s River Bend campus and as the buyer of TeraWulf’s 50.1% stake in the Abernathy joint venture. That is a thinner counterparty base than “$70 billion in contracts” suggests, and the correlation between those names is not zero.

A presentation problem runs through every figure in this piece, including the ones above. Contract values are quoted gross, undiscounted, and summed across fifteen or twenty years, which makes them incomparable to anything else on these companies’ financial statements. TeraWulf’s $19 billion works out to roughly $950 million a year, and reaches $33 billion only if Anthropic exercises both five-year extension options it is under no obligation to exercise. Core Scientific describes 1.1 gigawatts of leased capacity as more than $24 billion of potential contracted revenue and its initial AMD agreements as more than $14 billion of potential base contracted revenue, with “potential” carrying weight in both. Discount any of it at a plausible rate, against construction capex that lands in the first three years and rent that arrives across two decades, and the numbers compress hard. The headline sums are accurate and they are also the least informative available way to describe the economics, which is presumably why they are the ones in the headlines.

In crypto, one bad click is final. Learn to verify before you lose.

Irreversibility is the third. Bitdeer knocked down a building. Keel is winding its bitcoin mining business down across 2026 and 2027. Core Scientific has converted the majority of its revenue base away from mining. If bitcoin returns to $100,000 and hashprice recovers, the converted megawatts do not come back, because nobody rips out liquid cooling and a Tier-III electrical plant to reinstall ASICs at 15 kW per rack. CoinShares framed the whole transition as contingent on bitcoin recovering to $100,000 by year-end, and there is something slightly off about that framing, since a recovery would not reverse the conversions already committed. It would only slow the next wave.

Bitcoin traded near $64,000 on 6 August 2026, well below the $80,000 line CoinShares identified as the level at which hashprice keeps falling and more operators exit. The pivot has never been tested against a bull case, because it has been running for eighteen months into a worsening one.

The part that is easy to miss

CoreWeave’s proposed $9 billion acquisition of Core Scientific was terminated in October 2025 when Core Scientific shareholders voted it down, on the argument from Two Seas Capital, Gullane Capital, ISS, and Glass Lewis that the all-stock offer undervalued the company. Nine months later Core Scientific had signed an initial 530 MW across five sites with AMD, expanded its leased customer capacity to roughly 1.1 gigawatts, and reported colocation as 83% of revenue. The shareholders who blocked the sale were right, and they were right for a reason worth stating plainly, which is that the market was still valuing the company as a miner at the moment it stopped being one.

The pivot narrative has been told mostly as a story about compute, with AI displacing hashing as the more valuable use of a GPU-shaped hole in a data hall. The companies themselves tell a simpler story in their filings and their names. They stopped selling compute and started selling access to power on land they already control, at a moment when that access became the scarcest input in the largest infrastructure buildout in a generation. Keel put it in one sentence when it explained the name, saying it wanted to be the base that everything else depends on. Landlords have described themselves that way for a very long time.