Is the AI bubble bursting?

Chinese AI shares have fallen for twelve weeks running, most organisations that own AI hardware report that it sits idle at least half the time, and the companies that rent that hardware out are being asked whether they can repay what they borrowed. The main laboratories are in a different position, and so is the business that uses the tools rather than builds them.

What has fallen

The question is being asked because of what happened to share prices this week. On 30 October China’s CSI 300 index of blue-chip shares reached a one-year low and closed out its twelfth consecutive weekly fall, the longest run of weekly losses on record. The index is down 14 per cent over those twelve weeks and 7.6 per cent in October alone, and the selling has been concentrated in the companies that supply the AI build-out. Eoptolink, which makes the optical modules that connect servers inside a data centre, lost 46 per cent of its value in a month.

In the same week Alphabet reported that Google Cloud’s revenue had doubled in a year, and its shares fell by more than twelve per cent in a day. The complaint in both markets is the same one. Investors have stopped treating AI capital expenditure as an investment with a return at the end of it and started treating it as a permanent cost. One UBS analyst described the market as now pricing peak AI spending as a lasting burden on earnings rather than a temporary one.

The idle hardware

Behind the share prices sits a question of utilisation. A survey published by VentureBeat on 23 October, pooled from 385 respondents at organisations of 100 or more staff between July and September, found that 79 per cent of those running AI in production said their GPUs were busy half the time or less. The figures are self-reported and the sample is self-selected, so the precise number should be held loosely, but it agrees with everything else that can be seen. MIT Technology Review reported in 2025 that as much as 80 per cent of the newly built computing capacity in China’s data-centre rush was going unused. Gartner expects 80 per cent of the independent GPU cloud providers, the companies that buy chips and rent them out by the hour, to have ceased independent operation by 2030, and Bernstein estimates that only 35 per cent of the data-centre capacity announced in North America will be built. On 29 October Elon Musk predicted that more than 200 of those providers face consolidation or closure, and the finance director of one of them agreed publicly that a wave of mergers was coming.

Why half-used hardware does not pay

A third-party AI data centre is a simple business underneath. A company borrows to buy chips, servers, networking and power, then rents the chips out by the hour. The chips depreciate whether or not anyone is using them, a newer generation arrives every year or two, and over the life of the loan the rate an older chip can command falls. The loan is repaid from rental income across the useful life of the hardware, which is somewhere between three and five years, after which the chip is worth a fraction of what was paid for it.

That arithmetic only closes if the chips are busy. Hardware that is used half the time has to earn its whole cost in half the hours, and at the rates being charged this autumn it covers the power, the staff and the interest and leaves little or nothing towards the hardware itself. A provider that cannot reach high utilisation is therefore not a business having a bad year. It is a business whose model does not close, and it stays open only for as long as its lenders and its chip supplier are willing to carry it. Nvidia’s latest quarterly filing shows up to $164.5 billion in guarantees and commitments tied to data centres running its chips, which gives some idea of how much of the sector is currently being carried.

Why the main laboratories are different

None of this applies in the same way to the handful of companies that build the frontier models. Three things separate them.

Their hardware is full. Alphabet told investors on 29 October that it was working to bring more capacity online to meet demand, and Anthropic’s registration statement, filed on 26 October ahead of its public listing, says the company has asked its cloud partners for more computing capacity than they were willing to provide. Their constraint is supply. The third-party provider’s constraint is demand, which is the opposite problem.

The service they sell is profitable. Anthropic’s filing shows revenue of $16.3 billion for the first half of 2026 against a cost of revenue of $2.4 billion, a gross margin of about 85 per cent, and an operating profit in the second quarter. Google Cloud reported an operating margin of 36 per cent on revenue that had doubled. The losses that have made the headlines over the past two years sit in research and development, which is where the cost of training the next model is recorded, and in capital expenditure. Anthropic spent $9.7 billion on research and development in six months. Alphabet spent $51 billion on capital expenditure in a single quarter. Serving the model that already exists pays. Building the next one is what consumes the money, and that is a choice about growth rather than a flaw in the business.

Their position is difficult to assault. The export controls imposed since 2022 have split the market in two. Nvidia now reports no data-centre revenue from China at all, and analysts estimate that Huawei supplies around half of China’s AI chips with hardware that still trails the American frontier. A Chinese laboratory builds on weaker silicon for a domestic market in which the state is the main source of capital and paying demand is thin. The American laboratories were first to market with the models that businesses now depend on, they sell into the largest pool of paying demand in the world, and the controls ensure that their nearest rivals build on inferior hardware.

Full demand, profitable inference and a protected position mean that the main laboratories will keep growing, and will keep absorbing whichever parts of the digital economy can be run through a model. Whether they justify the valuations placed on them is a separate question, and not one this piece needs to answer.

Where the demand that lasts will come from

So part of the market is a bubble in the strict sense. Capital has been raised to build capacity that demand has not yet asked for, and some of that capital will be lost. The same happened with the railways and with fibre-optic cable, and in both cases the technology outlived the companies that overbuilt it. The rapid growth of the past three years will not continue indefinitely because it cannot. Share prices and data-centre announcements have run well ahead of the use being made of the technology.

The demand that lasts comes from somewhere less dramatic. It comes from the hours a business saves. A company of fifteen people that takes two days out of its month-end, stops paying for a reporting tool it uses a tenth of, or answers customer enquiries in an hour rather than a day has found a return that does not depend on anyone’s share price. There are more than five million businesses of that kind in the United Kingdom. Most of them have not yet made the change, and the time, cost and efficiency they can recover is where the sustainable growth in demand will come from.

What this means for a smaller business

AI is, in the end, more capable software. It is software that can read, write, sort and decide within limits that are now well understood, and it can be built into the systems a business already runs. The question for an owner is whether there is work in the business that this kind of software could now do, and in almost every business we have looked at the answer is yes.

The advantage goes to whoever engages first, and it is a disproportionate advantage for a reason that has little to do with the technology. A smaller business carries less structure. It has no procurement committee, no legacy platform contract and no department whose purpose is to protect its own existence. It can change how it works in a month where a large organisation takes two years, and for the next few years the tools are cheap, because the laboratories are competing for usage and cutting their prices with each release. That window will not stay open indefinitely. The point at which a market consolidates is the point at which prices firm up, and the businesses that already have the efficiency are the ones that bought it at the bottom.

What EGM offers

EGM Software builds bespoke systems for sole traders, small and medium-sized businesses and larger organisations at a fixed fee, and the client owns what is built. Where AI belongs in a system we put it there, local or cloud, and where ordinary software does the job we say so.

EGM Consulting offers the Position Review, a fixed-fee look at how a business runs now that ends in a written decision on what should change, including where software or AI would recover time or cost and where it would not. It commits the client to nothing further.

EGM Education and Training teaches staff to use AI tools well, in schools and in businesses, delivered only by trainers with demonstrated experience in education or instructional design.

If you would like to know what the technology could recover in your own business, please contact us at enquiries@egmfamilyholdings.com.

Written by the people who do the work at EGM. All Insights · Write to the company