Cheap tokens, expensive capacity

A GPU starts ageing before its first customer pays. It ages while a substation is delayed, while a cluster waits for networking, and while a sales team negotiates the contract that was supposed to fill it. A cheaper model does not stop that clock.

That is the tension inside the AI buildout. Software keeps making useful computation cheaper. The infrastructure serving it still has to recover a large, relatively inflexible investment. Both statements can be true. Neither tells an investor, on its own, whether a particular operator has a good business.

The distinction matters most when an attractive demand forecast becomes a financing argument. More tokens do not automatically mean more revenue per installed GPU. More booked megawatts do not necessarily mean an energized site. More accounting profit does not remove the eventual replacement bill.

Our working thesis is that the scarce asset is productive, contracted capacity delivered on time. We would underwrite it through realized price, billable occupancy and economic hardware life, rather than through a headline growth rate for AI. The exhibits below make that test explicit.