Skip to content
3 September 2026Global economy · AI · Financial stability

Experimental Newsroom

Careful observation · evidence · constructive criticism

Document reporting · Global economy

The AI boom can be real. Its financing can still be fragile

Wording and precision revised on 12 October 2026. The source record retains the dates stated in the article.

AI may save time at work long before its investors recover the cost of building it. The Bank for International Settlements examines how useful technology and vulnerable financing can grow together.

Originator, editorial direction & final publication responsibility
Bjørn Moe Aldema

Writing and editorial development
Experimental Newsroom human–AI collaboration

Published by Experimental Newsroom

Why this matters

  • The BIS says AI investment supported growth and financial markets in 2025, while warning that high valuations, debt and poorly disclosed financing could amplify a reversal.
  • When an investor is also a supplier, capital can return as revenue through purchase commitments. The arrangement may be legitimate, but outside readers may struggle to see how independent the demand is.
  • This is not a forecast of an imminent crash. The public-interest test is whether investors, lenders, auditors and regulators can trace who carries the risk if expected returns arrive late.

This edition is based on public documents available on 3 September 2026. It distinguishes the BIS assessment from our analysis and does not allege wrongdoing by any named company. No company, lender, worker, regulator or affected community was interviewed for this edition.

Language revised on 15 September 2026. The source record remains dated as stated in the article.

One deal can tell several stories

Consider a financing arrangement described by the Bank for International Settlements. A chip maker or cloud provider invests in an AI company. The company uses that backing to secure computing capacity from its investor. A data-centre developer may then enter the arrangement, building facilities against a long lease. Money, future sales and obligations now connect several businesses.

Each business can have a sound commercial reason to take part. For someone outside the arrangement, however, it becomes harder to see how much demand comes from paying customers and how much relies on those businesses supporting one another.

The BIS, which is owned by central banks, examines this relationship in its Annual Economic Report 2026. It describes how optimism about AI supported investment, share prices and favourable financing during 2025. It also identifies vulnerabilities: expensive shares, fierce competition, debt and private-credit arrangements whose terms can be difficult to see. Annual Economic Report 2026

If returns take longer to arrive than investors expect, those connections could carry the disappointment beyond technology companies. Suppliers, lenders and employers elsewhere could feel it too. The report describes a possible route for financial stress, rather than predicting a collapse.

What is documented — and what is not

Documented by BIS
A surge in AI infrastructure spending supported US investment and produced spillovers through global supply chains during 2025.
Documented by BIS
AI optimism also lifted stock valuations and helped keep financial conditions favourable.
Documented by BIS
Financing increasingly includes debt, private credit and arrangements in which investors in AI companies may also receive purchase commitments from them. BIS calls the reciprocal form “circular financing”.
Documented, uncertain
Studies of particular tasks often report large time savings, while estimates of economy-wide productivity gains are much smaller because adoption and integration are difficult. The size and timing of the payoff remain uncertain.
Not established
The report does not show that an AI crash is imminent, that every reciprocal arrangement is unsound, or that a named company will fail.
Our assessment
The central public-interest question is whether investors, lenders, auditors and regulators can see enough of the financing chain to judge who carries the risk if expected demand or productivity fails to arrive.

The boom begins with a plausible promise

There are substantial reasons for investing in AI. Assessing the financing starts by taking the technology's usefulness seriously.

BIS cites studies in which AI saves 20 to 50 per cent of the time spent on particular tasks. Yet the report says estimates of economy-wide productivity growth generally remain below one per cent over a long horizon. An organisation still has to redesign its work, connect systems and deal with bottlenecks before a faster task becomes a larger overall gain.

A hospital, for example, might shorten an administrative task while spending more time on verification and security. A small firm might find an AI service valuable without buying enough computing to support the scale of infrastructure investors expect. These are illustrations of the gap between usefulness and commercial return.

The timing matters as much as the eventual benefit. Investors and lenders have committed money against expectations of revenue arriving at particular points. Useful technology can still earn too little, or earn it too late, to meet those expectations.

How optimism moves beyond technology companies

Meanwhile, investment creates income well before an end customer pays for a service. Chip orders support manufacturers. Construction supports builders. Electricity contracts support utilities. Higher share prices can increase wealth and make further borrowing easier.

That gives expectations a real economic effect today. It also gives a reversal a route through the economy. Lower revenue can mean fewer equipment orders, postponed building work and more cautious lenders. Businesses outside AI may then find credit harder or more expensive to obtain.

BIS treats this as a risk worth examining. High valuations, low compensation for investment risk and limited disclosure could make a change in expectations harder to absorb.

The circular-financing question

One of the report's central concerns is circular financing: an investment relationship that also helps generate business for the investor.

A cloud provider can acquire a stake in an AI company whose spending then returns to it through a computing contract. Both can report expansion. The AI company has funding and capacity; the provider has future sales. The figures describe different parts of a connected arrangement, rather than necessarily representing separate sources of demand.

Such agreements can make expensive projects possible and give each side greater certainty. To assess their strength, an outside reader needs the terms: which purchases are conditional, who bears the losses, whether assets support several obligations, and how much business would continue without the investment link.

The revealing comparison is with demand that can stand on its own. Would customers still buy the service, at that price and scale, if the supplier were no longer helping to finance the buyer?

A data centre is also a physical promise

Even a fully financed data centre needs electricity, equipment, permits and customers. The International Energy Agency's Energy and AI report describes the sector's electricity requirements and gaps in the information available to policymakers. A financial commitment can be signed well before a grid connection is ready. International Energy Agency's Energy and AI report

Scarcity can make an operating facility valuable. For a project still under construction, the same scarcity can increase costs and postpone income. The effect depends on where the project stands and who has agreed to carry the delay.

Following the physical work therefore helps make sense of the financial figures. Announced, financed, connected and operating facilities are different groups. Recurring use by paying customers provides another measure again.

Where the risk could travel

BIS also points to private credit's growing involvement in AI and information technology. These lenders can support projects on terms different from bank loans or public-market financing. Their role can also leave less of the overall exposure visible to outsiders.

An equity investor may take the first loss when an asset falls in value. Loans, guarantees, leases and fund investments can pass some exposure to others. Separately, the Financial Stability Board has examined risks from financial firms' use of AI, including dependence on a few providers, correlated behaviour and failures of models or data. Those are distinct risks, but both accounts show why supervisors need to understand connections before stress puts them to the test. Financial Stability Board

An unsuccessful data centre need not endanger the wider financial system. The concern is a sufficiently large, connected fall in asset values: one that tightens lending, delays investment and affects jobs well beyond the original projects.

The strongest objection

Earlier technology booms offer a useful qualification. Railways and fibre networks could remain useful after investors lost money. AI infrastructure may likewise continue to support valuable services even where its original financing proves unsustainable.

That leaves two questions to judge separately. Does the infrastructure serve a useful purpose? Can its revenues meet the obligations used to build it? A productive facility can change owners after a default. Society can gain lasting infrastructure while also bearing some costs of a financial correction.

Recognising that possibility makes the analysis more precise. Confidence in the technology is compatible with close scrutiny of the contracts financing it.

What we will watch next

  1. Capital expenditure compared with realised AI revenue and cash flow. Announced spending should be separated from completed investment and from revenue earned from independent customers.
  2. Reciprocal commitments. Companies should disclose when an investor is also a major supplier and when investment is connected to future purchase obligations.
  3. Debt, leases and private-credit exposure. Long-term contracts, guarantees and asset pledges can carry risk even when they do not look like ordinary bank loans.
  4. Physical completion and use. A financed or announced data centre is not the same as a connected, operating and substantially used facility.
  5. Productivity beyond selected tasks. Time saved in a demonstration should be separated from durable gains across an organisation or economy, including verification and integration costs.
  6. Who absorbs a reversal. Regulators should be able to trace whether losses would remain with sophisticated investors or travel through funds, insurers, banks, workers, suppliers and local communities.

Who can make the financing easier to judge?

Following AI investment now means following electricity connections, construction schedules, customer purchases and credit as well as the technology itself.

The information needed is practical. Developers can show which sites are connected and occupied. Providers can distinguish independent customer demand from purchases linked to their investments. Lenders can explain where losses would fall. Auditors and regulators can test whether the disclosures make those relationships understandable.

The headline investment total cannot answer those questions. What would make the boom easier to judge is a clearer account of how the money moves, what customers use and who carries the risk while expected returns are still in the future.

Sources and limits

The BIS analysis is an institutional assessment, not an independently audited map of every AI financing arrangement. This article does not establish the risk position of any named company. Material corrections and source updates will be dated on this page.

Support independent journalism →