If the AI bubble bursts, AI won't go away. The tools built on top of it might, though, and the prices we've all got used to almost certainly will. If your business runs on software made with one of those tools, it's worth a few minutes now.
The BBC ran a good piece this weekend on the firms feeding those bubble fears. You've probably never heard of them. They're called neoclouds, and they run huge warehouses of AI computers that they lease to the likes of Anthropic, OpenAI and Meta.
What's actually happened
The short version is debt. These firms buy the hardware up front, then borrow against the chips and against long contracts with big customers. The BBC reports that CoreWeave, the biggest of them, has $35.1bn of debt. Its shares have lost a quarter of their value in a month, and several of its peers have done about the same or worse.
Gartner predicts that by 2030, more than 80% of neoclouds will no longer operate independently. That's not the same as 80% going bust, as being bought counts too. It's also only a forecast. CoreWeave's boss says the market is acting "irrationally", and analysts quoted by the BBC say demand for computing power still exceeds supply.
So nobody knows, and I'm certainly not going to pretend I do. We build software. We don't pick stocks. But the question that interests me is what happens further down the chain if it does go wrong, because that's where small businesses are.
The tools at the end of the chain
Think of it as a stack. At the bottom are the people with the computers. Above them are the companies with the big AI models. On top are the tools most people actually use: AI app builders and coding assistants, many of which pay somebody else for the AI underneath.
That top layer is the one that feels any squeeze last and has the least room to absorb it. When the cost underneath goes up, they've got three choices. Charge more, give you less, or be bought by somebody who'll decide that for them.
We've already seen a version of this, and no bubble had to burst. In June, GitHub moved Copilot to usage-based billing, based on the tokens you use. Its reasoning was refreshingly plain: the old model was "no longer sustainable", because a quick question and a multi-hour coding session could cost the user the same.
That's a sensible change. It's also a sign of where things are going. The flat monthly price was never really the price of what you were using, and as agents do more of the work, more of the real cost lands on the customer.
If I had to guess at a burst, I'd expect fewer tools, some of them merged or quietly closed, and the survivors charging closer to what it really costs. I can't see AI-assisted software going anywhere. It's far too useful. When the dot-com crash came, the internet carried on, and so will this.
If your app was built on one of these tools
The worry isn't that your app stops working overnight. It's that the thing it depends on changes its price or its terms, or disappears, while you're busy running your business. These are the questions I'd ask this week:
- Have you got the code? A proper copy, somewhere that's yours. Lovable, to its credit, lets you sync to GitHub and deploy elsewhere. Do that today rather than meaning to.
- Could it run somewhere else? Having the code doesn't prove it. Does anyone know how to start it up away from the tool that made it?
- Where's your data? Whose account holds the database, and when was it last backed up and restored?
- Does the app itself call an AI service? If it does, you're paying by usage in production, and that bill can move. Know what it costs you per customer.
- Could anyone change it without the tool? If the only way to fix a bug is one particular chat window, that's a risk worth knowing about.
None of that is a reason to panic, or to stop. Building with these tools is a brilliant way to find out what you want. Just make sure you could walk away if you had to. If you can't tell, that's what our code audit is for: £250 + VAT, and you get a plain English answer.
Why firms like ours are looking at their own servers
We use AI every day, and we're not about to stop. But I don't love the idea of a software company's whole way of working depending on the price list of one or two suppliers. So we're looking at bringing some of that work in house, on our own specialised servers.
A couple of years ago that would have been a daft idea. Now it's a real option. Qwen3.6-27B, for example, is released under the Apache 2.0 licence and aimed squarely at coding, and one comparison of local coding models rates it as the strongest general pick for a single 24GB graphics card.
Let's be honest about the downsides. That same comparison says hosted models often give better speed, context and reliability for demanding work, and I believe it. The hardware isn't cheap either. An RTX 5090 is around £4,300 in the UK at the moment, because everyone else wants the same memory.
So I don't see it as either/or. Use the best hosted models for the hardest thinking, and use your own kit for the steady, everyday work, and for anything a client would rather not send to a third party. Then a change in somebody's pricing is an annoyance, not a crisis.
There's a client benefit in that too. Your code stays with people you can ring, and if you want that for your own business, we can set up private AI on hardware you own. My own notes on the server side, and what you actually need to buy, are over on my site.
If you've got an AI-built app the business has started to rely on, and you're not sure how much of it you really own, get it checked or just give us a shout. Better to find out now than when the invoice changes.
Cheers, Rob
