Category: Technology | Published: 2026-10-13
The Bank of England AI warning raises a question that can get lost behind impressive product demonstrations: how financially sustainable is the technology supporting your business?
AI services may deliver useful results while the companies supplying them face enormous costs. Data centres, computing equipment, electricity and research all need funding, often before anyone knows how much customers will ultimately pay.
The Bank's concern is that more of this expansion is being financed through borrowing. If expected returns disappoint, the consequences could reach beyond technology shareholders to lenders, investors and other businesses connected to the sector.
For organisations buying AI tools, this is not an instruction to stop using them. It is a reason to assess the provider as carefully as the product, measure the value of the service and avoid becoming dependent on arrangements that would be difficult to replace.
The Bank's Financial Policy Committee discussed AI financing at its meeting on 25 September 2026. The committee monitors risks to financial stability, rather than judging whether a particular chatbot is useful.
Its published meeting record describes increasing debt, limited visibility into some financing structures and circular arrangements that could amplify losses if expectations are not met.
This goes further than a warning about falling technology shares. Lenders and other investors are helping finance the infrastructure behind AI, so their returns also depend on the sector turning anticipated demand into profitable business.
The warning identifies a vulnerability. It does not mean every AI company will fail, or that losses across the financial system are inevitable.
A company can raise money by selling shares or by borrowing. Both involve risk, but borrowing introduces repayment obligations that do not disappear simply because sales fall short of a forecast.
AI infrastructure requires substantial spending. A data centre needs premises, power connections, cooling and computing equipment. Model developers also pay for research and for the processing needed whenever customers use their services.
Those expenses can arrive well before the income needed to cover them. If customer growth is slower than expected, pricing becomes more competitive or equipment needs replacing earlier than planned, the business still has to meet its financing commitments.
That is why useful technology and financial resilience are separate questions. A service can attract customers and produce impressive results without generating enough profit to support the scale of investment behind it.
The figures cited by the Bank indicate the scale of the funding challenge. Morgan Stanley estimated that global AI-related debt issuance had reached around US$450 billion by early September 2026. That was more than twice the amount issued during the whole of 2025.
JPMorgan analysts estimated that around US$4.1 trillion of AI-related investment in infrastructure and equipment would be financed through debt between 2026 and 2030.
These numbers need careful interpretation. They concern the wider AI sector, not the borrowing of one company. The US$450 billion figure is an estimate of debt issuance, while the US$4.1 trillion figure is a forecast covering several years.
Neither should be read as a bill already due or as a guaranteed account of what investment will eventually take place. The longer-term outcome depends on demand, financing conditions and the pace of infrastructure development.
Circular financing arrangements can make the situation harder to assess. One example is a company investing in an AI developer that then spends the money on cloud capacity or equipment supplied by that same investor.
Such an arrangement is not necessarily improper. It can help a developer obtain funding and the infrastructure it needs to operate.
The difficulty is understanding how much growth comes from independent customers and how much depends on funding circulating between connected companies. An infrastructure supplier may gain revenue from the developer's spending while also relying on that developer becoming financially successful.
If new funding becomes harder to obtain, both sides can come under pressure. Other companies using their services may then face consequences even though they had no direct involvement in the financing deal.
The Bank's concern about these connections is therefore about visibility and the potential for losses to spread, not a blanket allegation that circular investment is fraudulent.
The source article also discusses figures reportedly set out in Anthropic's IPO prospectus. An initial public offering, or IPO, is a process through which a company offers shares to public investors.
According to those reported figures, Anthropic generated nearly US$4.6 billion in revenue in 2025 while recording an operating loss of more than US$8 billion.
Its reported net loss was much larger, at approximately US$42 billion. However, roughly US$34 billion of that total related to accounting charges associated with financing arrangements, rather than expenditure on running the business.
That distinction matters. Presenting the entire net loss as cash spent on everyday operations would give readers a misleading picture. Operating losses, accounting charges and cash flows answer different financial questions.
The prospectus reportedly included US$518 billion in future cloud and infrastructure commitments. These represent commitments over the coming years, not a single year's bill. They illustrate the scale of the planned spending, but should not be treated as a precise prediction of the company's future financial position.
Competition creates another challenge. As more models become capable of completing useful work, customers may be able to obtain similar results for less money. That can put pressure on the prices providers need to recover their investment.
Open-weight models are part of this picture. They make the parameters learned during training available for others to use or adapt, allowing organisations to run a model themselves or choose a hosting provider.
However, access to a model does not remove the cost of operating it. Hardware, hosting, technical support, security and integration still need to be considered.
A lower headline price is therefore not automatically a lower overall cost. The fair comparison is the cost of completing a task reliably, including staff supervision and time spent correcting mistakes.
The Bank of England's warning makes supplier resilience a practical purchasing issue. Before putting an AI service into an important workflow, ask what would happen if the provider raised prices, changed ownership or withdrew the product.
Useful questions include:
- What service-continuity commitments are included in the agreement?
- How much notice will you receive before significant pricing or product changes?
- Can you export business information in a usable format?
- Which underlying model and cloud providers does the service depend on?
- What would moving to another provider involve?
The last two questions are particularly important when buying through an intermediary. A smaller service may rely on several larger providers, creating dependencies that are not obvious from its own branding.
This does not mean avoiding newer suppliers. It means matching the amount of dependence to the importance of the work being supported.
An exit plan is more useful when it has been tested. Keep accessible copies of business information, document integrations and understand which parts of a workflow depend on a specific provider.
Try a realistic task with an alternative service before you urgently need one. Check whether the output, data handling and integration requirements are acceptable, rather than assuming that any AI model can be substituted without changes.
A provider switch may involve retraining staff, changing instructions, reviewing security arrangements and checking results again. Allowing for those costs makes decisions about supplier dependence more realistic.
Financial uncertainty in the AI sector does not mean customers cannot benefit from its products. The important test is whether the service delivers worthwhile results in everyday use.
Start with a limited trial and a clearly defined task. Measure speed, accuracy and the staff time required to check the output. Include subscription costs, setup work and ongoing support in the comparison.
Then ask whether the benefits would still justify the service if introductory pricing ended or usage charges increased. A saving that depends on permanently cheap access may be less dependable than it first appears.
For help assessing suitable tools and introducing them into business workflows, our AI Services page explains how we can support a practical approach.
The Bank of England AI warning is best understood as a reminder to separate technological progress from financial assumptions. Rising debt and complicated funding relationships can create risks even when the underlying products are genuinely useful.
For businesses, a sensible response is to choose suppliers carefully, preserve the ability to move and invest according to measured benefits. You do not need to predict the future of the entire AI industry to make those decisions well.
Photo: Chris Downer, City of London: The Bank of England , via Wikimedia Commons, licensed under CC BY-SA 2.0 . Archive photograph taken in January 2005, reproduced without editing.