James Thomason ran a simulation of the economics of a one gigawatt datacentre facility, at current build costs and current rental rates, and found it loses money in 98% of outcomes. That number has been circulating, usually stripped of the assumptions that produce it, which is a shame because the assumptions are the whole argument.
I want to work through why this kind of result is so counterintuitive to people who come from software, because I am one of them and I got it wrong first.
Software intuitions do not transfer
In software, the marginal cost of one more user is approximately zero, and the dominant risk is that nobody wants the thing. You optimise for finding demand, because once you find it the economics take care of themselves.
Capital-intensive infrastructure inverts both halves. The marginal cost of serving one more GPU-hour is not zero, it is electricity plus cooling plus the amortised cost of hardware that is depreciating whether or not anyone rents it. And the dominant risk is not that nobody wants it. The dominant risk is that everybody wants it, builds for it simultaneously, and the resulting supply collapses the price you can charge.
That second point is the one that does the damage in the simulation. You are not betting that demand for AI compute exists. You are betting on what the rental rate will be in year four, after everyone else’s capacity also comes online.
Three assumptions doing the heavy lifting
Depreciation schedule. GPUs are assumed to have a useful economic life, and that assumption is load-bearing. If an H100-class part is competitive for five years, the maths is survivable. If a new generation halves the cost per FLOP every two years, the rental rate you can charge for last generation’s silicon falls off a cliff long before the hardware physically fails. The accounting life and the economic life are different numbers, and only one of them pays the loan.
Utilisation. A facility running at 95% booked is a completely different business from one running at 60%. Spot and on-demand capacity is lumpy, training runs are bursty, and the gap between nameplate capacity and billed capacity is where margins disappear.
Power price and availability. A gigawatt is not an abstraction, it is a grid interconnect that takes years to secure and a price that is not fixed for the life of the asset.
Why this matters even if the number is wrong
I am not equipped to referee whether 98% is the right figure. The model is one person’s, the inputs are contestable, and the honest answer is that nobody knows what 2030 rental rates look like.
What is useful is the shape of the risk. The money being committed to AI infrastructure right now is being underwritten on assumptions about residual hardware value and future pricing, not on whether the demand exists. Those are very different bets, and the public conversation keeps treating them as the same one.
Which is also why the financing structures are worth watching more closely than the model announcements. When a chip vendor’s CEO reportedly has to step in and shield lenders from losses to get a large debt deal across the line, that tells you something about how the lenders themselves price this risk.
Source: The gigawatt question, via TLDR.