Hut 8's $9.8B lease: your 2028 compute is already priced

Hut 8 announced on July 20, 2026 that it fully commercialized its one-gigawatt Beacon Point campus in Nueces County, Texas, with a second $9.8 billion, 15-year lease for 352 MW of IT capacity. The first data hall under that lease won't be delivered until Q2 2028. Read that again: the price of AI compute two years from now is being signed today, in triple-net real estate contracts, by people who will never meet you.

What actually happened

Per Hut 8's announcement, the same unnamed "high-investment-grade" tenant that signed Phase 1 doubled its footprint at the campus to 704 MW. The Phase 2 lease is structured on substantially the same terms — triple net, 15-year base term, 3.0% annual base rent escalator. Campus-level base-term contract value now sits at $19.6 billion, rising to as much as $50.2 billion if all three five-year renewal options are exercised.

Site energization is on track for Q1 2027. Hut 8 expects cumulative NOI of $9.8 billion on the lease, averaging roughly $655 million a year at stabilization. Reuters reported shares rose on the news; the stock has nearly doubled this year.

Three details matter more than the headline number: the term is fifteen years, the rent goes up 3% every year regardless of what happens to model efficiency, and one tenant took the whole gigawatt.

Why long-dated compute leases matter for your business

The token price on your invoice looks like a software price. It isn't. Underneath it is a stack of physical commitments — interconnection agreements, escalating rent, debt service on $4.25 billion of senior secured notes — that were locked in years before the model you're calling was trained. Software prices fall. Fifteen-year leases with 3% escalators do not.

That's the asymmetry to plan around. When a new model makes inference 40% cheaper, that saving gets absorbed against a cost base that is contractually rising. Some of it reaches you. A lot of it doesn't. Meanwhile the concentration is worth noticing: an entire gigawatt campus contracted to a single counterparty is efficient right up until that counterparty's priorities change.

So don't budget as if AI costs trend down on their own. Budget the model layer like a utility bill that can move in either direction, and build so you can act when it does. Concretely: route every model call through one abstraction so switching providers is a config change; log cost per completed task, not cost per token, so efficiency gains show up as a number you can defend; and keep a working fallback to a cheaper or open-weight model that you've actually tested, not one you assume works. We've made this case about open weights and inference silicon — it's the same discipline, viewed from the land side.

Sources: Hut 8 — Fully Commercializes 1 GW Beacon Point AI Data Center Campus, Reuters.