Bubble or bust is the AI question of the month. There is a comforting answer to it. Even if it overshoots, something valuable will be left behind, the way the railways survived the mania and the internet survived the crash. That answer assumes tomorrow's AI will be a larger version of today's, financed on a path that simply continues. I think the more dangerous assumption is the straight line itself. The assets securing much of today's AI debt are GPUs and data centres financed on the premise that today's giant models will still be the dominant architecture when the debt matures. If a cheaper, smaller, more specialised architecture wins instead, the technology succeeds and the collateral fails at the same moment. That is not a demand problem. It is a straight line drawn through a bend.
AI does not have to fail for the AI investment boom to end badly
There is a comforting story about artificial intelligence. It may be expensive. It may consume disconcerting quantities of electricity and water. Its leading companies may lose billions while making commitments measured in hundreds of billions. But this is the price of building the next great general-purpose technology. The railway mania left us the railways. The dot-com crash left us the internet. Even if investors lose money, society will inherit something valuable.
I think much of that story is true. Artificial intelligence is real. A large and valuable industry will emerge from the current frenzy. But it does not follow that the companies leading the race will survive, or that the infrastructure now being financed will retain its value. The industry can succeed while its financial architecture fails.
Indeed, success may help to cause the failure.
The wrong question
Most discussion of an AI bubble asks whether demand will justify the investment. Will businesses pay enough for AI services? Will consumers continue subscribing? Will productivity gains become sufficiently large and sufficiently visible? These are sensible questions, but they assume that tomorrow’s AI will resemble today’s, only larger.
That assumption is doing an extraordinary amount of work.
The present investment cycle is based on ever larger general-purpose models, trained and operated in vast data centres containing enormous numbers of advanced GPUs. The financial commitments often extend many years into the future. Yet the technology develops in months, not years. The useful life of the financing and the competitive life of the asset are moving on radically different clocks.
The question is therefore not simply whether AI will generate enough demand. It is whether that demand will require the particular infrastructure being built today for long enough to repay the money borrowed to finance it.
A real industry and a financial bubble
The dot-com comparison is useful precisely because the internet was not an illusion. It transformed commerce, communication and society. What proved illusory was the belief that every early participant deserved its valuation and that every piece of infrastructure would earn the return assumed when it was financed.
The AI boom is potentially more consequential because it combines several cycles at once. It is an equity valuation boom, a physical infrastructure boom and, increasingly, a credit boom. The Bank for International Settlements has drawn attention to the shift from financing AI investment out of corporate cash flows towards debt and private credit. Its later work describes both on-balance-sheet and off-balance-sheet structures which strengthen the links between hyperscalers, banks and non-bank investors.
The numbers are formidable. The five largest hyperscalers are expected to spend more than $1 trillion on AI-related capital expenditure across 2025 and 2026. The Bank of England’s July 2026 Financial Stability Report cites an estimate that $240 billion of hyperscaler investment requirements in 2026 will be financed through investment-grade credit issuance. For comparison, the six largest US banks would typically issue between $150 billion and $170 billion of senior debt in a year.
This does not by itself prove that there is a bubble. It does show that the consequences of being wrong are escaping the venture capital portfolios in which young technology companies traditionally fail.
The circular machine
The financing structure is difficult to see because each transaction can be rational when viewed alone.
An investor provides capital to an AI laboratory. The laboratory commits to buy computing capacity. A cloud or data-centre company uses that commitment to raise debt and build infrastructure. It buys chips from a semiconductor company. The semiconductor company invests in, lends to or supports the AI laboratory. The resulting demand validates the chipmaker’s revenues, the infrastructure company’s backlog and the laboratory’s valuation.
Money moves around the circle and appears at several points as evidence of independent economic value. Yet the ultimate source of repayment may remain the same: the future ability of a small number of AI companies to sell services at margins sufficient to meet their compute commitments.
There is nothing necessarily improper about any of this. The contracts are real and the facilities are built. But the structure is reflexive. Confidence creates the financial capacity which appears to validate the confidence.
Oracle now illustrates the concentration vividly. S&P Global Ratings downgraded Oracle to BBB minus, the lowest investment-grade rung, and identified OpenAI as a key credit risk. It estimates that OpenAI accounts for roughly half of Oracle’s $638 billion of remaining performance obligations. A future revenue backlog can look like diversification across years while remaining profoundly concentrated in one counterparty.
The same risk is dispersed across special-purpose vehicles, data-centre developers, equipment leases, private-credit funds, utilities and suppliers. Legal separation does not necessarily create economic independence. A portfolio containing loans to a data-centre owner, a power project, a cloud provider and a chip-leasing vehicle may contain four borrowers but only one underlying bet.
The collateral problem
The traditional answer to this concern is that much of the lending is secured. If the customer fails, the lender still has the GPUs and the data centre.
That answer may prove dangerously complacent.
A GPU can remain physically functional long after it has ceased to be economically competitive. Each new generation may offer greater processing capacity, more memory and materially better performance per unit of energy. New model architectures may require different combinations of hardware. Smaller models, specialised chips, sparse computation and distributed inference may alter the amount and type of centralised capacity required.
The critical distinction is between physical life, accounting life and economic life. They are not the same.
CoreWeave’s public filings assign a six-year useful life to computing equipment. Its 2025 annual report also acknowledges that failure to keep pace with evolving AI technology requirements could materially affect its business and financial condition. Six years may be a defensible accounting estimate. It is nevertheless a remarkably long period in a market in which the competitive frontier can move several times within a single year.
Older GPUs will not instantly become doorstops. They may be used for inference, fine-tuning and less demanding tasks. But lenders do not recover their money from the fact that an asset can still be switched on. They recover it from the cash the asset can earn or the price at which it can be sold.
In a default caused by ordinary managerial failure, another customer might take the capacity. In a default caused by a technological change, many owners could seek to remarket similar equipment at precisely the moment customers prefer the new architecture. The event causing the default would also destroy the collateral value.
That is the mechanism which deserves much more attention. Technological progress normally supports an industry’s creditworthiness. Here, sufficiently rapid progress could impair the assets securing its debt.
The smaller model challenge
The dominant vision of AI is a universal machine, capable of discussing Shakespeare, Picasso, particle physics, tax law and credit risk. That is an impressive research achievement. It may not be the most sensible production architecture for every task.
Most organisations need bounded capabilities. A bank considering operational resilience does not require the whole of recorded human culture to be activated for every enquiry. It needs deep knowledge of regulation, its own systems and dependencies, reliable evidence, good reasoning, transparent sources and a clear understanding of when the matter should be escalated.
The alternative may be networks of smaller specialist models. A routing layer sends a question to the relevant expertise. Those models use controlled data, deterministic tools and institutional memory. Separate models challenge or verify the result. A large general-purpose model is used only when the question genuinely requires novel cross-domain synthesis.
This is not science fiction. Research on model routing, mixtures of experts and small-model agent systems is already advancing. It reflects a simple economic principle: use the cheapest and most controllable capability able to perform the task, with escalation where necessary.
The important phrase is not simply “small language model”. It is “networked specialist intelligence”. The system may collectively know a great deal without every component knowing everything.
If that architecture succeeds, AI demand could continue growing while the demand for today’s configuration of giant models and centralised GPU estates falls well short of current projections. AI would not have failed. It would have evolved.
Physical limits will accelerate the change
Energy and water are not secondary public-relations problems. They are constraints on the industrial model.
The International Energy Agency projects that global electricity use by data centres could roughly double to about 945 terawatt-hours by 2030, slightly more than Japan’s current annual electricity consumption. AI-focused facilities are growing faster still. Those demands will increasingly compete with homes, industry, transport electrification and the decarbonisation of the wider economy.
Technological efficiency will improve. Chips will perform more calculations per watt. Cooling systems will use less potable water. Computation will move closer to users and towards locations with available power. But efficiency does not guarantee lower aggregate consumption. If the cost of an inference falls by a factor of ten and usage rises by a factor of twenty, total demand still increases.
Political choices will therefore intrude. Communities will ask why electricity grids, water systems and public support should be organised around models whose benefits and ownership are concentrated elsewhere. Planning consent, grid connections, water rights and energy prices may become part of the effective cost of AI long before the technology reaches any theoretical limit.
Those pressures favour smaller, more efficient and more distributed systems. Once again, the response to the constraint may be good for AI and bad for assets financed on the assumption that scale would continue in its present form.
From private fortunes to public pensions
OpenAI and Anthropic remain unrated private companies. Their bankers are reportedly seeking investment-grade ratings in connection with prospective public offerings. An IPO would give each company a large cushion of permanent capital and could make further borrowing easier. That would reduce immediate liquidity risk.
It would also transfer risk.
Today, much of the direct equity exposure belongs to venture funds, strategic investors and sovereign institutions able to understand that they are making an exceptional bet. After flotation, the companies could enter mainstream indices. Pension funds and ordinary savers would acquire exposure through global and US trackers, often without making any affirmative decision about either company.
Nor is their existing exposure small. The Bank of England notes that rising equity prices have been driven by a narrow group of AI-related companies, increasing concentration in global indices. Its December 2025 report calculated that the investment-grade corporate issuers most closely associated with AI already represented 14.5 per cent of a major US investment-grade bond index, more than the largest conventional sector.
A successful flotation would be presented as evidence that the risks were receding. It might instead complete a transition from concentrated private risk to widely distributed public risk. Pension money would support the valuation. The valuation would support the rating. The rating would support more debt. The debt would finance more infrastructure.
The circle would turn once more.
What would the failure look like?
It need not begin with bankruptcy. Most crises begin when the next provider of finance declines to endorse the assumptions made by the last one.
Perhaps enterprise margins prove disappointing. Perhaps a new architecture needs much less centralised compute. Perhaps public resistance delays data-centre construction. Perhaps one leading laboratory loses market confidence and cannot complete its next capital raise. Lenders then question long-term purchase commitments, infrastructure providers lose the basis for their projected revenues and collateral values are revised.
Private-credit funds may discover that apparently diverse exposures depend upon the same handful of customers. Banks may have limited direct loans but substantial commitments through revolving facilities, construction finance, fund leverage, derivatives and underwriting positions. Insurers and pension funds may hold the longer-dated debt. Public equity markets will reprice the hyperscalers, chipmakers and infrastructure providers together.
The blast zone would not be confined to technology specialists. It would reach retirement savings, power projects, construction, corporate borrowing costs and any sector competing for capital with the AI build-out.
There are important firebreaks. Microsoft, Alphabet, Amazon and Meta have real businesses and immense cash-generating capacity. Data-centre buildings and power connections retain some value. Equity will absorb losses before senior debt. The global financial system is much larger than it was during the dot-com crash.
But none of those facts answers the underlying question. How much of the apparent diversification is ultimately dependent upon the same projected pool of AI revenues, and how much of the collateral will retain its assumed value if the architecture changes?
Follow the economic dependency
Boards, regulators and risk committees should stop asking only about direct exposure to named AI companies. They should ask where the cash needed to service every related obligation ultimately comes from.
A proper economic-dependency map would look through special-purpose vehicles, funds, leases and guarantees. It would connect data centres to their tenants, power projects to their off-takers, chip finance to utilisation assumptions and pension portfolios to both equity and credit exposures. It would also test what happens when several apparently separate protections fail together.
The stress should not be limited to an OpenAI default. It should combine a severe fall in private valuations, closure of the IPO market, renegotiation of compute contracts, a sharp reduction in GPU residual values, data-centre vacancies, private-credit fund restrictions and downgrades of adjacent investment-grade issuers.
That may sound extreme. A stress test which merely confirms the assumptions embedded in the base case is not a stress test.
The central lesson is uncomfortable. We are financing long-lived infrastructure against a short-lived conception of how artificial intelligence will work. We are treating contracts as independent sources of value when many depend upon the same capital-hungry counterparties. We may soon use pension savings to validate valuations which then support another round of borrowing.
Artificial intelligence may fulfil much of its promise. It may become cheaper, smaller, more specialised and more useful. It may also require far less of the infrastructure now being constructed for it.
That would be a triumph of innovation. It might also cause the default.
Richard Anderson
Founder, AndersonRisk
This essay was first published on 19 September 2026 on my Substack.