Anthropic's $19B TeraWulf data center lease

The headline number is staggering for a company TeraWulf's size: a 20-year lease expected to generate roughly $19 billion in contracted revenue, more than the miner's entire market capitalization of around $12 billion at the time of the announcement ^[1]. The campus that will house it does not yet exist - initial power is not expected until the second half of 2027, with the full 401 MW online by early 2028. Broken down, the lease implies about $950 million in annual revenue, or roughly $2.37 million per megawatt per year ^[2], a premium rate that reflects how much a frontier lab will pay to guarantee power years in advance.

The economics look even more lopsided from TeraWulf's side of the ledger. The company plans to invest only around $3 to $4 billion to build out the facility - less than a fifth of the lease's headline value - against last-quarter revenue of roughly $34 million ^[3]. In other words, a single tenant has just committed to pay TeraWulf, over two decades, a sum that dwarfs everything the company is today. That asymmetry is exactly why the stock reacted the way it did, and why the deal is being read as a template for an entire class of power-rich, cash-poor infrastructure owners.

TeraWulf did not start out as an AI landlord. It began as a bitcoin miner, and like its peers it found that a single long-dated lease was far more attractive than the volatile economics of coin production. The pivot is industry-wide: crypto miners sit on something Anthropic desperately needs - large, already-powered industrial sites - and by March 2026 they had sold off coins and signed more than $70 billion in AI computing contracts as they repointed that capacity toward steadier demand ^[1].

The Kentucky site itself tells the story. TeraWulf acquired the Hawesville campus, a former aluminum smelter on roughly 790 acres, in February 2026 for about $200 million ^[5]. Aluminum smelting is one of the most power-hungry industrial processes there is, which means the site came with exactly the grid interconnection a data center craves. CEO Paul Prager framed power - not chips - as the binding constraint of the AI build-out, arguing that 'you could create all the disruption in technology, manufacture more silicon and chips. You just cannot create overnight Megawatts.' ^[4]That single line explains why a lab would sign a 20-year lease on a building that will not open for another year and a half: the scarce resource is not compute hardware, it is the electricity to run it.

For Anthropic, the lease is a window into a balance sheet the public has never fully seen. The company has never turned a profit, yet it is locking in fixed compute costs stretching to 2047, ahead of an anticipated IPO ^[2]. That timing is not incidental: committing to two decades of rent gives potential public-market investors an unusually concrete look at the fixed costs underpinning a frontier lab, and raises the uncomfortable question of whether token-based revenue can actually cover them.

The concern is not hypothetical. Anthropic CEO Dario Amodei has warned that leading AI companies may need hundreds of billions of dollars in annual revenue to sustain escalating compute costs, or risk severe financial strain ^[6]. A 20-year, $19 billion commitment to a single campus is one line item in that math, and Anthropic has signed similar deals elsewhere. The lease, then, is less a real-estate transaction than a public statement about how expensive it is going to be to stay at the frontier, and how far in advance those bills now have to be committed.

Not everyone is convinced. The loudest counterpoint is definitional: $19 billion is contracted revenue spread over 20 years, not money in the bank, and it depends on a facility being built on time, financed at reasonable cost, and a single tenant staying solvent for two decades ^[2]. Skeptics note that the pattern - a pivot-stock popping on a press release for capacity that does not yet exist - has played out before, and the market's own reaction hinted at the doubt: after spiking as much as 19 to 26 percent, WULF gave back most of the move and closed up only around 4 to 5 percent ^[2].

That split runs straight through the community reaction. Retail traders treated the news as a momentum play and immediately began hunting for the next former miner to announce an AI deal, while more skeptical voices zeroed in on execution risk - the cost and feasibility of retrofitting a decommissioned aluminum smelter into a modern data center, and whether a 20-year lease survives a decade of hardware turnover. The bull case has institutional backing too: sell-side analysts raised price targets following the deal, with one shop lifting its TeraWulf target to $40 on the strength of a hyperscaler-grade tenant ^[7]. The gap between that optimism and the on-the-ground skepticism is the real story to watch as the first concrete gets poured.