Google buys troubled Crusoe data center and will invest $17 billion in the campus

The AI boom has a problem that a new model or another batch of GPUs cannot fix: computing capacity has to be housed somewhere, connected to the power grid and cooled on time. This is exactly the level at which Crusoe's large project in Wyoming stalled. The company failed to secure Google as a tenant and then sold it the unfinished site. Now Google intends to complete the campus on its own and invest $17 billion in its development.

For a developer, this may look like a distant deal about concrete and substations. But the timing of new cloud capacity, the availability of accelerators and, ultimately, the price of inference depend on projects like this. If a data center is delayed by a year, a software team does not receive the promised resources on schedule, no matter how much money has been allocated to models.

Where the original plan broke down

Crusoe announced the project together with the infrastructure company Tallgrass, which is linked to Blackstone. The campus was to be gradually expanded to 10 GW, a scale that is not that of a single machine hall but of a huge energy and construction complex. However, Google, whom Crusoe was counting on as a client, was reportedly unhappy with the cost and the delivery schedule.

The scheme was a familiar one for an independent operator: the developer creates the site, brings in power and leases capacity to a major client. Imagine a product team ordering a cloud cluster for July. The provider builds a building for it in advance, buys transformers and arranges the grid connection. If equipment is delayed, the budget has grown and the power line is not ready, the cluster cannot be partially launched by one team from a terminal. The customer is left to wait, look for another site or take construction under its own control. Google chose the third option.

The purchase price has not been disclosed, so it is impossible to estimate how much Crusoe lost or earned on the sale. Another figure is known: Google plans to put $17 billion into the further development of the campus. The companies' cooperation is not ending, though. Crusoe continues to build another facility for Google near Amarillo, Texas.

Why independent operators are feeling the squeeze

Crusoe operates in several models at once: it leases space to clients with their own GPUs, rents out its own accelerators and sells compute for running AI models. The company is also involved in other giant projects. In Abilene, it is building a campus for Oracle and OpenAI in phases, and next to it is erecting a separate 900 MW site for Microsoft. The first building of the latter project is expected in mid-2027.

Such a portfolio shows demand but does not eliminate construction risks. It is necessary to obtain land, power, grid equipment, cooling systems and the support of local authorities all at once. Any single element can delay the entire facility. Additional pressure comes from residents of the regions, who fear a growing load on power grids, water consumption and the cost of utility infrastructure.

Selling the site did not cut Crusoe off from capital: in September, the company raised $3.9 billion at a valuation of $30.9 billion. So what is happening looks, for now, more like a redistribution of risk than the collapse of a business. Google gets direct control of a critical facility, while Crusoe sheds a problematic project and continues its other construction work.

What product and engineering teams should do

When planning an AI product, cloud GPUs can no longer be treated as an infinite commodity with instant delivery. Ask your provider not only for the accelerator type and the hourly price, but also for the location, the date the site comes online, the available electrical capacity and a backup region. For critical workloads, check portability between clouds in advance and separate training from inference: they have different requirements for latency and availability.

Founders should build infrastructure risk into their financial model. A lower model price does not guarantee a lower total bill if the needed cluster is delayed or becomes available only under a long-term contract. The Google–Crusoe deal is a reminder of a simple thing: AI scales not only with code. Between a request to a model and the answer to a user stand years of construction, billions of dollars and a very physical power grid.