Three Ways to Get a Data Center

In September 2025, Nebius, a company that rents graphics processing units, or GPUs, to artificial intelligence developers, announced a five-year agreement to supply Microsoft with capacity from a new data center in Vineland, New Jersey, in a contract it valued at about 17.4 billion dollars. Microsoft, one of the largest builders of data centers in the world, had chosen in this case to buy capacity from someone else. That choice introduces the subject of this module: every hyperscaler decides, site by site, whether to build, to lease or to partner, and the mix says a great deal about how it manages risk.

The first option is to self-build. The company buys or options land, contracts for power, designs the facility and owns the result. Microsoft's Fairwater campus in Wisconsin and Meta's long line of company-owned campuses are examples. Self-building gives the company the most control over design, cost and timing, and it keeps the asset on the balance sheet, where it shows up in capital expenditure. It also puts the full construction risk on the company and ties up capital for years before a building produces anything.

The second option is to lease from a developer. Companies such as Digital Realty, Equinix and many private developers build data centers and lease them, often under long-term contracts signed before construction begins. The hyperscaler usually supplies its own servers and networking. As the previous module explained, the building then appears as a lease rather than as capex, and the commitment shows up in the lease footnote of the annual report, including leases signed but not yet commenced. Leasing trades some control and some long-term cost for speed and a lighter balance sheet.

The third option is to partner or to buy capacity as a service. This is what the Nebius agreement represents. A neocloud, meaning a company specializing in renting GPU capacity, such as Nebius or CoreWeave, finances and operates the facility and the chips, and the hyperscaler pays for access. Meta signed a capacity contract with CoreWeave in September 2025, and CoreWeave announced an expanded agreement with Meta in April 2026. For the hyperscaler, these arrangements add capacity quickly and move much of the construction and equipment risk onto the partner, which in turn often borrows against the hyperscaler's contract to fund the build.

Some arrangements sit between these categories. Meta's Hyperion joint venture with funds managed by Blue Owl Capital, announced in October 2025, keeps Meta involved in construction and management while outside investors hold most of the ownership. Oracle's role at the Stargate site in Abilene, Texas, is another hybrid: Crusoe developed the campus, and Oracle operates cloud capacity inside it for OpenAI. The labels build, lease and partner describe a spectrum rather than three boxes.

Why would a company choose one route over another? Speed is one reason; leasing and partnering can deliver capacity sooner than a greenfield build. Balance sheet treatment is another; spreading capital across partners and investors keeps reported capex lower than it would otherwise be. Flexibility is a third. A contract with a neocloud for a fixed term can be allowed to lapse, while an owned building cannot easily be un-built. Faraday's read is that the rise of leasing and partnering since 2024 reflects, at least in part, the hyperscalers' own uncertainty about how long demand for artificial intelligence capacity will stay at its current pace.

For anyone reading a local market, the route matters as much as the company. A self-built campus usually means a direct relationship between the hyperscaler and the utility, the county and the state. A leased campus means a developer carries that relationship, and the hyperscaler's name may not appear in local filings at all. Faraday's read is that this is sometimes deliberate, with developers filing under project code names until the tenant chooses to be named. A neocloud campus may involve three parties: the developer, the neocloud and the ultimate user. Knowing which route is in play tells you whose filings to read and whose financial health the project depends on.

The practical habit is to label every site you follow with its route and its counterparties. When the route is unclear, record that too. Later developments often clarify it, and the label keeps you from assuming that a hyperscaler owns something it only rents.