The AI Buildout Signed 15-Year Leases Against a Price That Resets Every Quarter
Compute contracts are now being written like commercial real estate: long leases, land financing, third-party guarantees. The revenue backing them reprices in months. Here is how to read the next data center announcement for who is actually carrying that mismatch.
A 15-year lease worth $9.8 billion was signed this July for AI data center capacity in Texas. The tenant was not named, only described as investment grade.
Hold that next to the other half of the market. Model vendors are now competing on tokens per second and price, and that price has moved several times this year. The buildings have 15-year terms. The revenue that pays for them reprices every quarter.
That gap is the most interesting number in AI right now, and almost nobody quotes it.
## The contracts stopped looking like technology deals
Read the last month of compute agreements in our deal tracker and the structures are unmistakably property finance.
Hut 8 fully commercialised its 1 GW Beacon Point campus with a second 15-year, $9.8 billion IT lease. Anthropic committed about $10 billion over six years to Volta for Nvidia-powered capacity that physically sits in a 133 megawatt Bitdeer facility in Norway, which means Anthropic holds a six-year obligation to a startup that holds capacity in someone else's building. In Korea, the first 30 billion won for the Namyangju X-AI Smart Data Center went to equity and land acquisition, funded by LG CNS alongside securities and asset management firms. That is a land deal with a compute story attached.
OpenAI went the other way and bought the risk outright, committing roughly $20 billion to a self-developed 3.2 GW campus in Georgia with a long-term power and tax agreement underneath it.
## Then look at who is guaranteeing whom
The structure worth staring at is Nvidia negotiating a $250 billion financing guarantee to back OpenAI's lease and construction debt for a 10 gigawatt campus in Piketon, Ohio.
Read that slowly. The supplier is backstopping the credit of the customer who buys its chips, so that the customer can finance the building that houses them. If demand comes through, it is elegant. If it does not, the chip vendor is holding the paper on its own demand.
## The grid became the gatekeeper
There is a third party in every one of these deals that signs nothing and can stop all of it.
On August 4, Texas ordered ERCOT and the Public Utility Commission to audit every new data center project before it can connect. The interconnection queue is around 474 gigawatts, roughly 90 percent of it data centers. A queue that size is not a pipeline. It is a waiting list, and most of it will not be served.
A signed lease on a campus without a firm interconnection is a claim on a building, not on compute.
## What to do with this
**Read term length before you read deal size.** A $10 billion headline tells you almost nothing. Six years against a product that reprices quarterly is a different instrument from six years against a government programme. Ask what the tenant's revenue visibility actually is over the term.
**Check whether the guarantor sells to the borrower.** Vendor-backed financing is a real and legitimate structure, and it is also the one that quietly converts a demand risk into a balance sheet risk. When a supplier guarantees a customer's construction debt, count that capacity twice at your peril.
**Watch interconnection dates, not groundbreaking dates.** Power queues, not chips, are where this cycle will visibly break first. Texas just made that measurable in public.
None of this says the buildout is wrong. Demand may well arrive. But the industry has spent a year arguing about model quality while quietly taking on the duration risk of a utility, and the two conversations have not met yet.