Bitcoin Miners Are Quietly Becoming AI's Landlords

Key Takeaways

• The scarce asset in the AI boom is not a chip. It is a plug. Bringing new power onto the United States grid now takes a median of more than four years, and the connection queue holds more than 2,060 gigawatts of projects. A site that already has a live connection is worth a fortune.
• A dead aluminum smelter is now the site of a $19 billion AI lease. Anthropic signed a 20-year deal on a Kentucky plant that was shut in 2022 precisely because its electricity got too expensive. Its new owner is a Bitcoin miner.
• This is a sector, not a one-off. Public Bitcoin miners have signed tens of billions of dollars in Artificial Intelligence (AI) hosting contracts, and the biggest of them are on track to earn most of their money from computing rather than coins by the end of 2026.
• The 20-year leases hide two risks the Bitcoin headlines miss: who actually stands behind the rent, and what a re-energized smelter does to the power bill of the town next door.

The Smelter That Died of Its Own Power Bill

In June 2022, Century Aluminum told more than 600 workers in Hawesville, Kentucky that it was idling the plant, the largest producer of military-grade aluminum in North America. The reason was not weak demand. It was the electricity bill: the power the smelter needed had more than tripled in price, and melting metal stopped making economic sense.

An aluminum smelter is, in essence, a machine for turning cheap electricity into metal. When the electricity is no longer cheap, the machine is worthless. So the pot lines went cold, the workers got layoff notices, and the great hall went dark.

Four years later, that same electrical connection is the most valuable thing on the property. On July 6, 2026, a Bitcoin miner named TeraWulf announced a 20-year lease at the old Hawesville site with Anthropic, the AI lab behind the Claude models. The contract is expected to generate roughly $19 billion in revenue and deliver about 401 megawatts (MW) of computing power, with the first electricity flowing in the second half of 2027.

The irony is not decoration. It is the entire thesis. The power that was too expensive to make metal in 2022 is now the anchor of a multi-billion-dollar bet on AI. Nothing about the physical connection changed. What changed is what plugs into it, and how desperately that new tenant needs a socket that already works.

What Anthropic Actually Rented

The buildings are the least of what Anthropic rented. The prize is the power: a heavy, high-voltage connection to the grid of the kind a smelter needs and a data center cannot run without, already built and already energized on a site that used to melt aluminum. That connection is where the money is.

The popular story of 2026 is that AI is bottlenecked on chips. That was true in 2024. The bottleneck now is upstream of the chip: it is the wire that feeds it. As of the end of 2025, more than 2,060 gigawatts (GW) of new power generation and storage were sitting in line waiting to connect to the United States grid. The median project now waits more than four years to go from application to switch-on, roughly double the wait of the mid-2000s.

Read that again with a builder’s eyes. If you want to stand up a gigawatt-scale AI campus on a fresh field, the wait for power alone can cost you four years or more before a single chip runs. But a former smelter comes with a connection built decades ago and already energized. You are not buying real estate. You are buying a place near the front of a very long line.

That is why a lease priced per megawatt tells the real story better than the headline dollar figure. Anthropic’s $19 billion over 20 years for 401 MW works out to a contracted rate of roughly $2.4 million per megawatt, per year:

$\frac{\$19{,}000\text{M}}{401\ \text{MW} \times 20\ \text{yr}} \approx \$2.37\ \text{M per MW-year}$

Line the recent miner deals up on that same per-megawatt basis and a market rate appears:

The asterisk on IREN matters: that deal is an AI cloud contract in which IREN also operates the graphics processing units, so its per-megawatt number bundles in the compute, not just the power and the shell. The pure colocation leases, where the tenant brings its own chips and just needs powered space, cluster around $1.2 to $2.4 million per megawatt per year. Either way, the number that made all of it possible is the one no spreadsheet lists: the years of queue time the miner already paid off.

The Whole Sector Got Repriced

TeraWulf is not an outlier. It is the newest data point in a stampede. IREN struck its own $9.7 billion agreement with Microsoft, expecting roughly $1.9 billion a year in recurring revenue from a single tenant running Nvidia’s newest chips. Cipher Mining leased 300 MW to Amazon and another 168 MW to the cloud firm Fluidstack, with rent starting in the second half of 2026. The tenants are the biggest names in computing, and the announced totals already run into the tens of billions of dollars.

The tell is in the revenue mix, which has flipped almost overnight. Several of the largest miners, including IREN, Core Scientific and TeraWulf, are on track to earn around 70% of their revenue from AI and high-performance computing (HPC) by the end of 2026, up from essentially nothing in 2024. None of this happened because a group of Bitcoin companies suddenly got good at artificial intelligence. It happened because the world got power-constrained around them, and they were sitting on exactly what became scarce.

Follow the Backstop

There is a puzzle buried in the TeraWulf announcement. The company said its Anthropic lease is expected to be supported by an investment-grade (IG) credit. But Anthropic, for all its scale, is a private startup that is still deeply unprofitable and does not carry an investment-grade rating. So where does the safety come from?

It comes from the same place it came from in Cipher’s deal. When Cipher signed with Fluidstack, the arrangement was quietly backed by Google, which agreed to take over the lease if the AI startup went bust, and which also took an equity stake in the miner. A hyperscaler standing behind a smaller tenant is how a lease to an unrated startup gets dressed up as investment-grade paper.

If that structure sounds familiar, it should. It is the same move this site documented when Anthropic borrowed $36 billion and the senior debt only priced at investment grade because Broadcom, the chip partner, agreed to cover the shortfall on default. The pattern repeats one layer down the stack: a cash-rich giant lends its balance sheet to make a shakier counterparty’s obligations look safe. The miners are not just landlords. They are the visible tenants in a building whose real credit sits with Google, Microsoft and Amazon.

The Case That This Is Smart Money

It is tempting to call this a bubble and move on. That would be lazy. The scarcity is real and durable: an interconnection queue measured in years cannot reverse next quarter, so a company that already controls energized capacity holds something a richer competitor cannot simply buy its way past. And where a hyperscaler backstops the lease, the fear that a lab stops paying rent in 2028 is largely someone else’s problem. The crude short thesis, that these miners are doomed crypto husks, is wrong. Which is exactly why the real risk sits somewhere the doom crowd is not looking.

Where the Risk Actually Lives

The risk lives in the tail and in the bill.

Start with the tail. These are 20-year leases against a technology whose shape changes every couple of years. The Anthropic deal commits 401 MW to a single tenant until 2046. That is a bet that the compute economics of 2026 still make sense two decades out, and that the backstopping giant still wants the exposure. A hyperscaler guarantee protects the first few years far better than the last ten.

Then look at the equity, where the market is already flashing caution. TeraWulf’s stock spiked as much as 19% on the morning of the $19 billion announcement, then surrendered most of the gain to close up only about 5%, and by the next session it traded below where it had started before the deal was public. A $19 billion headline that leaves the stock lower two days later is a market telling you the good news was already priced, and the execution risk between now and first power in late 2027 is not. Critics have been blunter still, arguing that miners dumping Bitcoin to chase AI rent are making a historic strategic mistake.

Finally, the bill, which is where this stops being a finance story and becomes everyone’s problem. Every megawatt a former smelter re-energizes for AI is a megawatt competing for the same regional grid that heats homes and runs hospitals. When a shuttered smelter comes back not as an aluminum plant that employed more than 600 people but as a data center that needs a small fraction as many, the power draw returns while the local payroll does not. The scarcity that makes the miners valuable is the same scarcity that pushes electricity prices up for the households next door. The miners capture the upside of the constraint. Ratepayers absorb the cost of it.

That is the quiet trade underneath the loud one. A dying industry’s scrap became the AI boom’s most coveted real estate, the credit behind the rent increasingly leans on a handful of hyperscalers, and the meter kept running on everyone else. The first power at Hawesville is due in the second half of 2027. Whether the electricity that once made military-grade metal ends up making cheaper intelligence or just more expensive utility bills is the number worth watching, and it will show up on a statement long before it shows up in a press release.