Jim Chanos, whose firm has previously placed bets against Tesla, Wirecard and the Chinese property sector, said in an interview with The Observer, that he was betting against “most of the neoclouds” – smaller companies that let out AI computing capacity – because they are “valued as tech growth stocks when, in fact, they’re highly leveraged landlords”.
Many neoclouds are former bitcoin mining companies that have borrowed heavily in order to turn to the business of buying AI chips and land for data centres, before renting the computational power out to companies such as OpenAI. Neoclouds including CoreWeave, Nebius and TeraWulf have all become targets for short sellers betting that demand for the assets they lease will fall.
I’m always wary of companies that change their business line. Yes, this has its problems – Apple’s move from computers to phones worked pretty well, as did Nokia’s from gumboots to phones.
On the other hand, those who go from yesterday’s fashion today to today’s fashion tomorrow, hmm. Perhaps they’re being most agile in riding the current waves of interest. Or, perhaps, there’s nothing there except that fashion.
Risk is, that if the AI bubble pops, then these companies will be lumbered with big buildings in unattractive places ( Slough. Hayes Middx ) full of expensive kit. I used to work for a company who built and ran datacentres just as the IT economy popped in 2001/2. They nearly went bankrupt and were bailed out at the time and lo! They are doing the same thing today, but under different management.
They could convert them into homes for migrants, I suppose.
The infrastructure itself will always be useful but they may need to swap out some of the servers depending on the latest technology fad.
We’re back to the railways and even all that fibre that went in to the ground, those who funded it may take huge financial losses and go bankrupt, but the underlying technology is still quite valuable once someone has figured out how to exploit it without having to service huge debts.
Yes, much like railway mania, some companies might go to the wall, and some others buy up the assets at a good price.
The servers aren’t so much “latest fad” but faster. Which means operationally cheaper. At a certain point, it adds up to spend money on better ones to save on the operational costs.
But isn’t a warehouse full of obsolete chips not worth much, or very useful? Fibre in the ground is still fibre in the ground 10 years later, a railway ditto, but a data centre is largely the chips inside it, and they depreciate pretty quickly. I guess a data centre will retain some a value because it has the physical infrastructure already there, energy, cooling etc, but it will be a fraction of what was invested. Googling suggests the chips and IT infrastructure is about 75% of data centre cost, and that will all go to zero pretty quickly.
Yep. The land, the structure and the power supply for a data centre are all useful things which could be repurposed.
However, I don’t know about the chips. If the AI landscape changed so that only 20% of existing data centres continue operating, what happens to all the chips? How might they be used?
The chips don’t stop working just because they are last year’s spec. If the original owner goes to the wall and the assets are bought cheaply, they can still be sweated as a data centre, just no longer a top-tier one.
That depends on the planning permission system. The building could be worth a lot more than a bare plot as it is immune to NIMBYs saying it should be left undeveloped.
There’s still some value. I’ve not thought of when chips are obsolete but Iceman’s 5-7 years sounds about right. Which means quite a drop each year.
If the AI bubble bursts, it won’t be that the chips are just last year’s model. The problem will be they’re the wrong type of chip. AI workloads are big on floating point math, hence adding specialist processors which aren’t much use for much else.
Much like railway mania, there’s demand at the right price.
Something I did would probably work at DeepSeek prices but not Gemini. Unfortunately, I need Gemini to make it work.
Chips (GPUs) have a 5-7 lifetime. You can possibly extend that out a little. But that only becomes economic lifetime is sometimes still wants to pay for using them during all those years. This is not obvious given the speed of technology development
Indeed.
Another consideration is the expected length of life of the customer and their profucts
5 years is an eternity in IT.
Part of that is the energy cost. New kit consumes less power than legacy for the same work output. A time comes when the power saving makes it cost effective to turn off the old stuff and invest in new.
So, they’re doing infrastructure. Supply the computing power for someone else to play with. Which yes, doesn’t justify high valuations on a tech leader basis. It doesn’t make it not worthwhile though.
The best analogy I can come up with is building and operating shopping centres. Your strength is knowing how to do building, securing an anchor tenant to drive traffic, running the place, and maintenance. And knowing who to bribe for planning permissions 🙂
Actual retail is not your gig, that’s your tenants’ job. If you get it right, solid, dependable income. But it’s not sexy so it has to be hyped up.
Is the demand there? Maybe, at least for a while. I guess they’ll find out.
Owners of shopping centres – say, Land Securities – are boring high dividend companies. Or bust, like Intu. They’re not high growth companies like Nvidia has been…..
Agreed. As I said. The valuations probably aren’t justified.
Apple’s move from computers to phones worked pretty well
But they didn’t really move from one to the other, they spotted what 1,000s of highly paid consultants missed when it came to 3G and 4G – there was no killer app in the sense of one service that would drive mobile network business models, the killer app was the internet itself.
The shift from Macbooks to iPhones is a continuation of what Apple are good at. Both of these are computers that run UNIX. Both of them sell because of Apple really focussing on making them easy to use, nice to look at and handle (in a way that Nokia smartphones weren’t). Apple’s particular thing is computing outside the enterprise (where Microsoft are very much more enterprise).
There’s quite a lot of overlap between crypto mining and running neoclouds. You need to manage loads of servers and the software on them. I don’t think that’s the problem here. Coreweave know what they’re doing. The real problem is that we have too much capacity for the AI demand. Almost daily I’m reading things from Ed Zitron or Michael Burry about circular funding, or companies using special companies to pay for compute, so it doesn’t appear in the accounts.
Nvidia, Coreweave, Oracle etc are all making money off this AI stuff, but it’s all sitting on the shoulders of companies like OpenAI and Anthropic that are only staying alive because of investors burning money on them. So, those are nearly all Greater Fool, at least in the medium term. New investors dry up and those companies fall. Plus there’s more and more competition in the AI game (Meta, Grok, Perplexity, Google, Mistral, various Chinese ones).
In a gold rush, sell shovels?
I am reminded of the Only Fools episode, where Grandad was locked away in a Spanish police station and admitted to gun running during the Civil War.
Rodney : “You was running guns ? In the middle if a civil war ?”
Grandad : “Well that’s the best time to do it, Rodney !”
Yes. Sell shovels. Sell. Not…
Lend money to a leasing company so that they can buy your shovels and then lease them out to prospective prospectors, knowing full well that if they fail you won’t get more than your first month of lease payments or your shovel back. But still be able to book the profits for selling them because the leasing co has bought them with money you lent them.
Apple’s move from computers to phones worked pretty well, as did Nokia’s from gumboots to phones
Nokia started in the 1860s as a paper mill, moved into gumboots (rubber products, but I don’t think they made condoms, unlike LRC), and were then bought by a cable company (roughly the Finnish BICC), before going in for electronics.
At current pricing data centers are a much higher value use of big buildings full of servers than bitcoin mines. It would be a negative if they didn’t respond.