if it floats, flies or megaFLOPs

I just thought that this little bit of research I did for my own curiosity might be of interest – I’ve been saying for a while that it’s definitely a sign of something that companies which had previously been claiming to be on the path to AGI are now renting out their spare compute to companies which are actually on the path to AGI capable of credibly claiming to be expanding the frontier.

I had not checked, however, whether the simplest possible explanation for this behaviour might be true – that it’s just an extremely profitable business line to be in. And, when in an idle moment I looked up some numbers, I think it is. Surprisingly, a company run by Elon Musk is doing at least one thing that’s profitable and cash generative.

Basically, we know from press releases that SpaceX is renting out the Colossus 1 datacentre to Anthropic for a payment of $1.25bn a month. Or as we finance pros say, $15bn a year. Estimates of the running costs vary, but I think it’s very hard to justify anything more than $500m a month ($6bn a year) for electricity, security, maintenance, the whole shebang. I have not seen any estimate of its construction costs that was above $13bn (and a lot of estimates that were much lower), and let’s say for the purpose of argument that it’s got a depreciation life of five years (and so you allow $2.6bn a year for that).

In other words, SpaceX is taking in a profit of (15-6-2.6 =) $6.4bn a year for a $13bn asset, on very conservative assumptions. That’s a return on capital of 50%, nice work if you can get it. (This is the unleveraged pre tax cash-on-cash return – you can mess around with assumptions about financing costs if you like, but you are not going to get a result which turns a 50% ROI into a bad investment).

Caveats!

Obviously, the biggest one is that leasing always looks like a fantastic business because you don’t see the embedded options. Anthropic apparently has some very strong break clauses, and if those are exercised in such a way as to leave SpaceX high and dry without an external client, the ROI drops pretty fast.

Also, two longer term points. The first being Nassim Taleb’s proverb – “you often see a glut that isn’t followed by a shortage, but it’s very rare to see a shortage that isn’t followed by a glut”. A 50% ROI on data centres will last for exactly as long as there’s a global shortage of the relevant chips, and not a minute longer. (And in this context, it’s worth remembering that past estimates of the depreciation life of AI GPUs seem to have been systematically and significantly too short. As a chronicler of Brompton Bikes, I am fully aware of the business problems associated with making a product which lasts forever).

Second, and in my view still a source of concern – the Anthropic cash flow isn’t really coming from end users. Anthropic in Q2 was making a profit of $559m; if we assume that this doubles, then doubles again, then it’s still not reaching the $1.25bn a month lease payments. Don’t put any load on that envelope calculation – I don’t know and can’t check whether Anthropic accounts for the lease payments as an expense or as capex. Even if they account for it as an expense, SpaceX doesn’t have any prior debt-like claim on it, so you have to look at the cash in the overall context of Anthropic, which is of a company that still needs to raise lots of external finance all the time to keep doing what it does.

So, the cash is still coming from investors. It doesn’t make sense, in my view, to go down rabbit holes, but the fundamental fact is that this is still a shovel rush rather than a gold rush – the amount of money that actual end-user customers are prepared to pay, at present, is a lot more than it used to be, but nowhere near the levels that would be needed to support the capex.