AI is driving a data center building boom. I have written about it extensively (see 1, 2, 3, ∞), done podcasts, and more.
It is easy, however, to get wrapped up in chatbots, image models, massive capital expenditures, soaring power demand, ASI/AGI, and the looming singularity, and not think about the financial drivers of what is happening.
But that is an error. Think of what is happening through the lens of project financing. Developers can lock in 12–15% IRRs¹ on new data centers, while stabilized hyperscale campuses² trade like long bonds³ at 4–5% cap⁴ rates. That spread is irresistible, so money floods into development.
² Data centers full of GPUs leased out to large companies like Microsoft, Google, and OpenAI.
³ Bonds that don't mature for decades. These are generally highly sensitive to interest rates.
⁴ The rate of return on the leased data center, based on the difference between operating costs, interest expense, and lease income.
But the imbalance is obvious. Once the centers are built and leased, relatively few buyers want them. The tenant credit is pristine (Amazon, Microsoft, Google), but that’s precisely the problem: the leases are too safe, too long, and too flat. Stabilized data center assets behave like annuities, not growth real estate. Only mega-pensions and sovereigns can absorb them, and even then only sparingly.
Developers and private credit increasingly don’t care. They’ve already learned how to sidestep the exit problem. The answer is a familiar word: securitization.