Power, Sites and the Refinancing Calendar

Start with the physical constraint, because it now binds harder than the chips. A cluster of graphics processing units, or GPUs, needs electricity delivered at scale, continuously, to a site with cooling, fibre and permits. Chips can be ordered. Power has to be connected, and in many markets the connection takes years. An operator that holds contracts and chips but no energised site holds a promise it cannot yet keep.

Crusoe's development at Abilene, Texas shows both the advantage and the fragility of controlling a site. Crusoe built the campus on land at the Lancium Clean Campus, with a planned capacity of about 1.2 gigawatts across eight buildings, for Oracle, which serves OpenAI under the Stargate program announced in January 2025. The first buildings came into service in 2025. Crusoe's announcements describe on-site natural gas generation alongside the connection to the grid run by the Electric Reliability Council of Texas, or ERCOT. On-site generation is an answer to the interconnection queue: it lets an operator start sooner than the grid alone would allow.

Then the plan moved. In March 2026, Oracle and OpenAI ended plans for a further expansion at the site, according to Bloomberg, while the existing campus continued operating. Later that month Crusoe announced a separate 900 megawatt campus in Abilene to support Microsoft, including an on-site power plant, bringing its projected capacity across the location to about 2.1 gigawatts. One customer's change of plan became another customer's capacity. That is the value of controlling power: a site with energy attached can be re-let. A cluster built for one customer's specification is harder to move.

The general principle follows. In this model, power and land are the durable assets and chips are the perishable ones. An operator that owns or controls energised sites holds something that retains value across chip generations and across customers. An operator that leases space from others holds only the perishable layer, plus a lease obligation. CoreWeave's attempt to buy Core Scientific in 2025 was, in these terms, an attempt to move down the stack toward the durable asset. The failed shareholder vote left it relying on leases and new builds instead.

Now the second constraint: refinancing. The debt that funds a fleet is sized to the contracts that fleet serves. When those contracts are renewed at the same or better terms, the operator refinances easily. When they end, the operator must find new customers at whatever the market pays for older hardware, and borrow against those new, often weaker, contracts. The investment-grade ratings on CoreWeave's March 2026 facility show how cheap money can be against a strong contract. They say nothing about the cost of money against a weak one.

Refinancing risk also has a calendar. Bonds and loans issued in 2024 and 2025 mature on fixed dates. If those dates arrive when credit markets are tight, or when the market for older chips is soft, the operator may have to accept expensive terms, sell assets or raise equity at a low price. The dates are in the filings. A reader should map them against the contract end dates and against the expected arrival of new chip generations, because the three together describe the stress points.

There is a link between the two constraints. Lenders lend more readily against a fleet that sits in a secure, energised site with a long lease or outright ownership, because a stranded chip in a building with no power is worth very little. Power security therefore lowers the cost of capital, and the cost of capital decides how cheaply the operator can bid for the next contract. Site control compounds.

Faraday's read is that the neoclouds most likely to last are the ones converting chip-era profits into control of power and land, because power is the input whose scarcity is lasting longest. Watch where each operator's sites come from, who owns them, and how long the leases run.