The AI bubble doesn’t need a thousand companies to fail.
It may only need one.
Different logos. Different balance sheets. But how many genuinely different bets? The danger is not just that AI disappoints. It is that a handful of buyers, and one shared view of computing demand, underpin too much of the boom.
What happens when the cost of intelligence falls faster than the cost of its infrastructure?
The AI industry looks like a collection of separate bets. There are model developers, cloud operators, chip suppliers and specialist infrastructure businesses. Follow their commercial and investment relationships, however, and the boundaries start to blur.
Microsoft is both an investor in OpenAI and a cloud provider to it. Google, Amazon and Microsoft have both equity and cloud relationships with Anthropic. These are documented connections, not just similarities between businesses. Microsoft’s agreement; Anthropic’s S-1.
That does not make the relationships inherently unsound. Strategic investment can finance infrastructure that customers genuinely need. But it does make one question unavoidable: how much apparently diversified exposure depends on the same small group of buyers remaining viable?
This is a concentration-risk story, with a technological disruption risk underneath it. A major buyer’s distress could transmit losses through the network; a breakthrough that changes computing economics could force several investments to be repriced together. Neither possibility establishes that collapse is imminent.
Your investment. Your customer. Your risk.
The financing logic is straightforward. A technology company invests in an AI developer. The developer purchases computing capacity. Cloud providers build capacity, hardware suppliers sell equipment, and the prospect of future demand supports more investment. If the buyer weakens, the same supplier may face both an investment loss and reduced commercial demand. That is the double exposure to examine.
The network is not confined to one cloud. OpenAI’s relationship with Oracle brings another infrastructure provider into the picture, while the Stargate build-out identifies NVIDIA hardware. CoreWeave separately discloses OpenAI contracts and its own dependence on NVIDIA technology. OpenAI’s Stargate announcement; CoreWeave’s 2025 Form 10-K.
The point is not that every dollar goes around in a closed circle. It is that a growing network of contracts is not the same as a diversified base of paying end customers. The diagram makes those selected relationships visible without pretending to measure the probability of loss.
A supplier can also be an investor
Selected commercial and equity links. Select a company to isolate its connections.
Different logos do not necessarily mean different risks. Arrows show direction, not size; selected links are not a complete ownership or supply-chain map. Evidence & definitions ↗
The scale is enormous. The dependencies are narrow.
Start with OpenAI. In its October 2025 partnership announcement, Microsoft said OpenAI had contracted to purchase an incremental $250bn of Azure services. The same announcement described Microsoft’s equity stake in OpenAI. This is a dated announcement, not a claim that the entire amount remains outstanding today. Microsoft’s October 2025 agreement.
Now consider Anthropic. Its September 2026 S-1 discloses $114.5bn of remaining obligations under the relevant Google arrangements, at least $110bn of AWS spending commitments, and at least $44.9bn of Azure commitments. The Google period runs from October 2026 to July 2033; AWS from April 2026 to April 2036; and Azure from November 2025 to December 2032. Each agreement has its own conditions. Anthropic’s S-1.
These are not three annual revenue figures, and adding them does not produce a useful annual spending number. They are a way of seeing the scale and duration of one buyer’s commitments to several counterparties.
The distinction matters because today’s contract is supported by tomorrow’s ability to pay. If demand, financing or competitive position disappoints, the question is not simply whether computing is useful. It is whether the contracted capacity remains economically supportable on the agreed terms.
Tomorrow’s compute, contracted today
Anthropic’s disclosed cloud commitments, US$bn
Source: Anthropic S-1, September 29, 2026. Amounts shown without aggregation.
A broad industry can still have a narrow customer base.
CoreWeave makes the concentration tangible. Microsoft accounted for approximately 67% of its FY2025 revenue. At year-end, CoreWeave reported $60.7bn of remaining performance obligations against roughly $5.1bn of annual revenue. The former is contracted future business, not cash already collected. CoreWeave’s 2025 Form 10-K.
This does not mean that 67% of CoreWeave’s revenue came from OpenAI. The disclosed customer is Microsoft, and attributing that share to an underlying lab would require separate evidence. It does mean that one buyer’s decisions matter disproportionately to this particular business.
A diversified cloud company and a specialist infrastructure provider are not interchangeable balance sheets. Their ability to absorb weaker demand, delayed payments or lower utilisation can differ substantially. Concentration tells us where to look; it does not tell us that a default is inevitable.
One customer. Two-thirds of revenue.
CoreWeave’s revenue by customer, FY2025
Source: CoreWeave 2025 Form 10-K. Microsoft’s share is approximately 67%; the remainder is derived.
What if OpenAI or Anthropic cannot meet expectations?
Imagine a major lab losing access to affordable financing, suffering a sustained competitive setback, or finding that customer revenue cannot support its infrastructure commitments. The first question would be about the lab. The next questions would be about everybody connected to it.
- Cloud providers: which commitments remain collectible, and which capacity plans need to be reconsidered?
- Infrastructure owners: which data centres still have an economically viable tenant, and how easily could that capacity be redeployed?
- Hardware suppliers: which planned equipment orders survive a change in the buyer’s financing or demand?
- Investors and lenders: where could investment losses coincide with weaker customer revenues or asset values?
Those are channels to stress-test, not a forecast of simultaneous insolvency. Some counterparties could absorb the shock, find new customers or adjust spending. Others might have less room. The danger is a cascading repricing of expectations, rather than an assumption that every connected company must collapse together.
The second trigger could be success, not failure.
Now imagine a different shock: a new model, training method or inference technique delivers comparable capability with substantially less computing. That could be excellent news for users. For an infrastructure investment priced around expensive computation and sustained scarcity, the implications could be less comfortable.
“Diffusion replaces transformers” is not a reliable shorthand for “GPUs become worthless.” NVIDIA’s own research explores diffusion-based language generation using transformer language models and GPU infrastructure. The categories are not mutually exclusive. NVIDIA’s diffusion-language-model research.
The relevant issue is therefore not the name of the next architecture. It is whether innovation changes the amount, type, location or price of the computing that customers need. An infrastructure asset can remain technically useful while earning less than its financing assumed.
More usage does not automatically mean more computing.
Consider an illustrative case. Computing required per task falls by 50%, while task volume rises by 50%. Aggregate computing demand becomes 0.5 × 1.5 = 0.75, or 25% less than before. Task volume would have to double merely to keep aggregate computing demand unchanged.
Cheaper AI could stimulate enough new demand to exceed that threshold. It could also shift work towards more demanding tasks. Both are reasons to test the assumptions, not reasons to assume that every existing infrastructure commitment will earn an adequate return. Use the controls below to explore that trade-off.
Cheaper intelligence. More demand. Enough?
How efficiency gains and task growth combine to change computing demand. This is not a forecast.
Holds workload mix and other factors constant. Does not model prices, margins, training demand, hardware mix or contract renegotiation.
Better technology can challenge infrastructure economics. Compute index = (1 − efficiency gain) × (1 + task-volume growth) × 100.
Falling compute per task does not imply the disappearance of GPUs.
The market does not need to wait for bankruptcy.
A financial correction can begin before an operational failure. Investors need only become less confident that future payments, utilisation, margins and asset values will match expectations. When several investments depend on the same buyers and assumptions, those expectations can reset together.
This is why “AI will keep improving” is not a sufficient investment argument. The harder question is how much of the financial structure depends on a handful of companies being right about demand, financing and technology at the same time.
The evidence does not tell us that the bubble will collapse tomorrow, or that the entire industry shares the same fragility. It tells us to distinguish the breadth of company names from the breadth of underlying economic exposure. Strong demand, adaptable infrastructure and lower costs could also support a more resilient outcome.
The next shock would not have to prove that AI is useless. It could simply reveal that too many companies financed the same future and called it diversification.
The vulnerability is not that AI stops working.
It is that the economics change before the contracts do.
A case for examining connected exposures and testing assumptions, not a prediction of insolvency or a recommendation to buy or sell securities.