A bitcoin miner just signed a $9.8 billion lease and never mentioned bitcoin

The tenant is anonymous, and that is the strangest thing about Hut 8’s new Beacon Point lease. The company put out a press release on Monday about a 15-year, $9.8 billion deal and would not name the company writing the checks, beyond calling it “high-investment-grade.” When a customer commits to a decade and a half of payments at that scale and stays behind a curtain, the anonymity is doing work. Somebody very large wants 704 megawatts of Texas power and would prefer you not know how badly it needs it.

That 704 MW figure is the whole story. This is the second lease at Hut 8’s Beacon Point campus in Nueces County, and it goes to the same tenant that signed Phase 1 back in May. They took 352 MW then, and now they have taken the other 352 MW, filling the entire one-gigawatt site. The campus is fully commercialized against its 1,000 MW of utility capacity, which Hut 8 secured through an interconnection agreement with AEP Texas before any of the leasing happened. Lock the power first, sell the compute second. That sequencing is the entire business now, and I think it is the only part of the AI infrastructure story that is hard to copy.

The numbers read like a REIT that wandered into a semiconductor conference. The new lease carries a 3.0% annual base rent escalator and three five-year renewal options, and it is expected to throw off roughly $655 million a year in net operating income once it stabilizes. At the campus level, base-term contract value now sits at $19.6 billion, average annual NOI at $1.31 billion, and if every renewal option gets exercised, the number climbs to $50.2 billion. Across the whole portfolio, Hut 8 is now holding 949 MW of contracted AI capacity backed by 1,330 MW of utility capacity, $26.6 billion in aggregate base-term value, and average annual NOI north of $1.75 billion, all of it leased to or backstopped by investment-grade counterparties. Shares opened about 5% higher premarket and kept climbing through the session, trading up more than 11% around $102 by midday. The stock has nearly doubled this year, though it is still off the $124 it touched in late May.

What “352 MW built on NVIDIA’s architecture” actually means

I want to slow down on the engineering, because the most interesting line in the whole release is the one nobody outside the infrastructure world will notice. Hut 8 went back and redesigned the first data hall around NVIDIA’s architecture and pulled 57% more capacity out of the exact same land and the exact same utility contract. Same dirt, same interconnect, 57% more compute. When you are power-constrained, and every single company building these things is power-constrained, squeezing more density out of a fixed feed is worth more than going out and buying new land.

To understand how that is even possible, you have to know what these buildings are now. The second Beacon Point hall is going up as an “AI factory” on NVIDIA’s DSX reference architecture, and DSX is worth explaining because it quietly rewrote how data centers get built. NVIDIA introduced the Omniverse DSX Blueprint at GTC 2025 as a reference design for gigawatt-class facilities, scalable from around 100 MW up into the multi-gigawatt range, validated at its own AI Factory Research Center in Virginia and already underpinning monsters like the 1.2 GW Stargate site in Abilene. The pitch is that a company can build one of these from a blueprint instead of inventing the whole stack from scratch, which is exactly what a former bitcoin miner needs.

Two pieces of DSX matter for what Hut 8 is doing. The first is a configuration NVIDIA calls DSX Boost, which tunes power management and workload distribution to deliver either roughly 30% lower power draw or about 30% more GPU density per megawatt. Stack that kind of density gain on top of a full liquid-cooling redesign and a 57% capacity uplift from the same footprint stops sounding like marketing. The second is the digital twin. The DSX approach runs the entire facility as a physically accurate simulation in NVIDIA’s Omniverse before anyone pours concrete, and Jacobs, the firm leading engineering and construction at Beacon Point, is building exactly that kind of data center digital twin to model the critical assets ahead of commissioning. You debug the building in software first. For a project with a Q1 2027 energization deadline, that is the difference between hitting the date and eating a year of penalties.

Then there is the heat. A rack of current-generation NVIDIA silicon, the GB300 NVL72 class of machine, throws off so much thermal load that air cooling is simply off the table. Hut 8’s cooling and power partner here is Vertiv, whose OneCore reference architecture for the DSX blueprint runs high-density liquid cooling at over 140 kilowatts per rack. For context, a normal enterprise rack from a few years ago lived around 5 to 10 kilowatts. We are talking about an order-of-magnitude jump in heat per square foot, which is why the redesign around NVIDIA’s architecture, the density gains, and the liquid plumbing are all the same story told from different angles. And because this sits in South Texas, Hut 8 has said it is using closed-loop cooling to hold down water consumption, which is the kind of detail that reads like a footnote until a county commissioner asks where the water is going.

Vertiv being the counterparty is the part that made me sit up, by the way. NVIDIA’s DSX ecosystem lists Vertiv as one of the partners providing simulation-ready cooling and power for these blueprints, and Vertiv is also the company physically delivering that at Beacon Point. Hut 8 is not improvising an AI factory. It is assembling one from the exact reference parts NVIDIA is handing the entire industry, which is efficient and also tells you how commoditized the “build” layer is becoming. The scarce thing was never the blueprint.

From a wooden hut at Bletchley Park to a Miami power broker

The name is a Turing reference. Hut 8 was the building at Bletchley Park where Alan Turing’s team broke the German naval Enigma, and the company took the name when it was founded in October 2017 by Marc van der Chijs as a Canadian bitcoin miner. The early story could not be further from a $50 billion contract pipeline: an electricity supply agreement in Medicine Hat, Alberta, a $100 million construction commitment, a TSX listing in 2018 that made it one of the first publicly traded miners, and a stubborn HODL strategy that turned it into one of the largest corporate holders of self-mined bitcoin anywhere. For years, the whole thesis was cheap Alberta power pointed at ASICs, coins accumulating on the balance sheet instead of being sold at production.

The company that signed Monday’s lease got there through a couple of hard turns. In November 2023, the Canadian miner merged with US Bitcoin Corp, redomiciled to the United States, moved its headquarters to Miami, and put Asher Genoot in charge. Then in March 2025 it spun up American Bitcoin, a mining-and-treasury vehicle that Hut 8 owns roughly 80% of, founded alongside Eric Trump and Donald Trump Jr. and listed separately on Nasdaq as ABTC. That structure let Hut 8 push the pure bitcoin exposure into a subsidiary and start recasting the parent as something else entirely, an “energy infrastructure platform” in its own words, that happens to have come up through crypto. The politics of that particular cap table is not my lane, so I will leave the American Bitcoin ownership story to the people who cover it properly.

Why a miner becomes a landlord

The pivot is not subtle, and it is not unique to Hut 8. Bitcoin mining economics have been ugly for a while, with BTC trading below many miners’ estimated production costs for months and something like a fifth of the industry underwater. What these companies were quietly sitting on, though, was the one thing the AI buildout cannot conjure on demand: grid-connected power, land, cooling, and a signed interconnect. An AI tenant does not care that you used to mine bitcoin. It cares that you have a thousand megawatts ready to energize and the operational muscle to build to spec. So the miners flipped the model. Bitcoin goes in the back as the flexible, opportunistic load, and AI moves to the front as the contracted, higher-margin tenant.

Hut 8’s first proof of that model was not Beacon Point. It was River Bend in Louisiana, signed in December 2025: 245 MW, a $7 billion base-term value, and a detail that matters more than the headline: the lease payments are backstopped by Google. That backstop is the mechanism that makes this entire sector function, and it is worth understanding because it is doing quiet financial engineering. Most miners have no investment-grade credit of their own. So when one of them signs a lease with a smaller AI operator, a hyperscaler guarantees the payments, and the miner can then walk that contract into the debt markets as collateral for senior-secured notes at a rate that would otherwise be impossible. The lease stops being a revenue line and becomes a financing instrument.

Beacon Point is a cleaner version of the same idea, because the tenant here is high-investment-grade directly, with no guarantor in the middle. And the lease terms are where the real protection lives. These are triple-net, take-or-pay agreements. Triple-net means the tenant, not Hut 8, covers the taxes, the insurance, and the maintenance on the building. Take-or-pay means the tenant owes the money whether or not it ever spins up a single GPU. Hut 8 gets paid for the capacity, not the usage, which turns a data center into something much closer to a toll road with a fifteen-year contract. Genoot’s line in the release is that the real test of a power-first approach is what partners commit against it, and a tenant doubling its footprint on the same site is about as loud a vote as you get.

The rest of the field is running the same play

Hut 8 is not even the biggest of these deals, which tells you how strange this market has become. IREN, which holds zero bitcoin by choice, signed a $9.7 billion deal with Microsoft for 76,000 NVIDIA GB300 GPUs at its Childress, Texas campus, plus a separate $3.4 billion cloud arrangement with NVIDIA and a partnership to build up to 5 gigawatts of DSX-aligned infrastructure over time. TeraWulf, which is exiting mining more or less entirely, has locked in something like $12.8 billion of contracted AI revenue, headlined by a 20-year lease with Anthropic at its Hawesville, Kentucky site worth roughly $19 billion, more than the company’s own market cap when it was signed. Cipher landed a 15-year AWS contract worth $5.5 billion. Applied Digital, CleanSpark, MARA, and Riot are all somewhere on the same curve.

Core Scientific went furthest and then hit a wall worth remembering. It had roughly $10 billion of contracted revenue through CoreWeave and about 1.3 gigawatts of contracted power, and then its own shareholders voted down CoreWeave’s offer to acquire the whole company. That is the cautionary note underneath the euphoria. When the tenant tries to buy the landlord and the landlord’s own investors say no, you get a reminder that these tightly coupled arrangements can come apart at the seams, and that a signed contract is not the same as a closed deal.

Zoom out and the sector has announced more than $70 billion in cumulative AI and HPC contracts, and analysts at CoinShares expect AI and high-performance computing to make up around 70% of listed-miner revenue by the end of this year. The market has already repriced accordingly. A basket of these stocks has run up 50 to 70% in 2026 even as bitcoin itself slid into the low teens. Nobody is valuing these as mining companies anymore. They are being valued as power-backed data center developers that still happen to hold some coins.

What it means for markets

The signal underneath all of this is that the scarce input in AI stopped being the chip somewhere in the last year and became the electricity to run it. NVIDIA can ship GPUs faster than the grid can deliver megawatts, and that inversion is why a former Alberta bitcoin miner can suddenly hold a $26.6 billion contract book. The companies that spent the crypto boom fighting over cheap power and building industrial-scale sites turned out to be holding options on the exact bottleneck the hyperscalers now face. This is the same story I keep coming back to when I write about who actually controls the compute, the models, and the off switch: whoever controls a construction-ready interconnect controls the deal.

The constraint is getting political, too, not just physical. New York already drew a hard regulatory line, and I wrote about what the 50 MW threshold in its data center moratorium means for anyone trying to build at this scale. Texas is friendlier, which is a big part of why Nueces County has a gigawatt campus going up in it, and Long Island does not. That regulatory gap is becoming a real variable in where the compute physically lands. It also runs straight into the economics, because the incentive party that helped build a lot of these campuses is ending, and the subsidies that made greenfield power cheap are getting clawed back jurisdiction by jurisdiction. If you want the wider view on how those incentives shake out, I broke down the international data center policy picture country by country separately.

The second thing worth watching is the debt. Building to hyperscaler spec is brutally capital-intensive, and the sector has taken on infrastructure-scale leverage to fund it. Convertible notes, senior secured notes backed by those backstopped leases, and interest expense at some of these names are jumping tenfold in a single quarter. Hut 8’s own structure looks conservative next to some peers, and the take-or-pay terms give it real downside protection, but the whole edifice only holds if the tenants are solvent in year 14 and the AI demand curve keeps bending the way everyone has priced in.

Which is the thing I am not going to fully untangle here, because it deserves its own post. A 15-year lease is worth $9.8 billion only if the tenant is still there and still paying in 2041, and “investment-grade” is a rating, not a promise, especially when the counterparty is almost certainly building this to chase AI demand that nobody has actually watched hold at this scale for fifteen straight years. Lock in a decade and a half of power at today’s assumptions, and you are making a very long bet on a very young workload. Hut 8 has built a quietly clever machine here, and I would still love to know who signed this. Ask me again when the 10-K makes them say it.