Buying picks and shovels in a gold rush is an investment cliché that dates back to the dotcom bubble, when people used it as a justification for avoiding frothy internet commerce businesses like Amazon and instead concentrating on the hard physical asset companies which were providing the infrastructure of the internet. Like Enron, or Global Crossing. These days, the proverb is often applied to the AI gold rush, and in particular the companies building the data centres that power scale-obsessed frontier labs.

And the numbers look pretty persuasive. Possibly the most interesting thing in the SpaceX IPO prospectus (if not the most spectacular) was a little nugget of information about its data centre business. Because Elon Musk has a “dual monetisation strategy” — in case the whole Mars thing doesn’t work out — SpaceX had to tell us a bit about the current rental market for data centre chips.

Alphaville’s emphasis in bold below. Since this was published, more details about this deal have slipped out. It seems that Anthropic has agreed to take on more or less the entire compute of the Colossus I data centre near Memphis.

Depending on whether you believe Epoch AI or Global Data Center Hub, this cost between $7bn and $13bn to build and kit out with GPUs. It uses around 350MW of electricity, which costs about $150mn at US wholesale prices; let’s allow the same again for other costs like insurance, staffing and maintenance. And let’s be really conservative and allow a three year depreciation life for the GPUs.

That would mean that all told, the running expenses and depreciation for Colossus I will be somewhere between $3bn and $4.5bn. That means that if it is being rented out for $1.25bn a month, the return on investment is likely to be more than 100 per cent. You can do a little ready reckoner to see the effect of different assumptions, but they’re not really going to make a practical difference; this is an asset which likely cost no more than $13bn to make, and should generate $15bn of rental payments in the first year.

In other words, it’s a very profitable business, and consequently, we should guess that most other data centre projects look pretty good on the numbers right now. On the most recent earnings conference call, SpaceX’s CFO Bret Johnsen confirmed that their AI capex was seeing a payback period of less than one year. Nice work if you can get it.

But in many ways, it seems a little bit too nice, in a way that will be familiar to students of financial history. Leasing businesses often show up during technological revolutions. For example, there was a bubble in computer leasing stocks in the 1970s, famously described in “The Money Game” by Gerry Goodman.

And computer leasing has always looked like a fantastic business. Except when it doesn’t, as in the case of OPM Leasing, the company which popularised the phrase “Other People’s Money”. The trouble with computer leasing is always in the break clauses; the client is prepared to pay a premium to lease equipment because it really needs the computer services, and for some reason it can’t get them any other way.

But the client obviously doesn’t actually want to be paying out enough to build someone else a new data centre every year, so they are likely to always be looking for alternatives. And when they find another way to get the compute that they need, the lease payments dry up quite quickly. In other words, lease revenue of this kind is something you should expect to have for a good time rather than a long time.

The very fact that it’s so profitable tends to attract market entry. As Nassim Taleb puts it, you often get gluts that aren’t followed by shortages, but it’s very rare to get a shortage that’s not followed by a glut. And the historical record suggests that when the scarcity starts to be relieved, it very quickly becomes a good idea not to have leveraged too much, and not to have invested on the basis of valuations that depended on returns that couldn’t be sustained.