AI Is Financing Its Own Customers. Genius or a Dangerous Loop?
Imagine selling a product, then investing in your customer so they can buy more of it. Your sales grow. Your customer expands. Investors see two businesses moving ahead, and the value of your stake may rise as well. You have helped finance both the demand for your product and the evidence that your industry is booming.
Something like this is happening across AI. Chipmakers invest in model developers. Model developers buy computing power from cloud providers. Cloud providers buy chips. Contracts and financial guarantees help fund the data centres and electricity behind the whole operation.
At first glance, it looks absurd: a company gives money to a customer, then celebrates when some of that money comes back as revenue. Look again, and there is a serious industrial strategy. The industry is trying to build the infrastructure, products and customer demand for a new market at the same time.
The uncomfortable possibility is that this could be both financially dangerous and extraordinarily successful.
Follow the money
In September 2025, NVIDIA and OpenAI announced a letter of intent covering at least 10 gigawatts of NVIDIA systems, with NVIDIA intending to invest up to $100 billion as capacity was deployed. In February 2026, OpenAI announced new funding that included $30 billion from NVIDIA, alongside investments from Amazon and SoftBank.
These are dated announcements, not figures to add together. The original $100 billion was a proposed investment ceiling; the later $30 billion was an announced contribution to a funding round. Neither figure, by itself, tells us how much cash has already been spent on chips.
The Anthropic arrangement makes the overlapping roles particularly clear. Anthropic committed to buying $30 billion of Azure computing capacity and contracting additional capacity using NVIDIA systems. In the same announcement, NVIDIA and Microsoft committed to invest up to $10 billion and $5 billion, respectively, in Anthropic. The companies supplying the infrastructure are also helping finance the customer.
The connections extend to financial guarantees. NVIDIA’s quarterly filing describes $36 billion of commitments to buy computing capacity from selected AI cloud providers as of July 2026. It also describes an August agreement supporting specified lease and power payments for an OpenAI site, with guarantees capped at $105 billion and subject to conditions and phased deployment. That is a maximum potential obligation under the agreement, rather than money already paid. But it shows how far the chipmaker’s involvement can reach beyond selling hardware.
Why an expensive valuation can become useful
Normally, we think of a valuation as the market’s estimate of what a company is worth. Here, it can also help determine what the company becomes.
A high valuation allows an AI developer to raise more money while selling a smaller share of the business. That money buys computing power, pays researchers and funds products. Large contracts then help cloud providers finance data centres, while long-term electricity purchases can support new power projects. If the resulting products attract paying customers, the investment starts to justify the confidence that financed it.
There is a practical reason to coordinate all this. An AI company needs servers before it can serve customers at scale. The data centre needs electricity before it can operate. The power supplier needs confidence in future demand before committing capital. Each business is waiting for another to move.
The International Energy Agency’s 2025 base case projected that electricity generation serving data centres would more than double between 2024 and 2030. Whatever the eventual total, the funding problem reaches well beyond chips. It includes land, cooling, grid connections, generation and years of research.
In that sense, the apparently circular financing can solve a real problem. A high valuation helps raise capital; capital builds capacity; capacity enables better products; and useful products can earn revenue. The danger begins when the market treats every step in that sequence as guaranteed.
The prize is a lead that compounds—and customers who stay
There are two reasons an AI company might rationally spend far ahead of current profits.
The first is that better AI can help build the next generation of AI. More computing power allows more experiments and larger training runs. Better models can then help researchers write software, debug systems and analyse results, potentially accelerating the next round of development.
There is already evidence of this inside the labs. OpenAI says early versions of GPT-5.3-Codex helped its teams debug training, manage deployment and analyse evaluations. This documents AI helping humans develop better AI. It does not establish that models can independently design and build their own successors without human direction. Even so, the strategic implication is substantial: a technical lead might make the next technical lead easier to obtain.
From that perspective, waiting for comfortable margins could mean losing the race. A company may accept enormous spending today because it believes that falling behind would cost it access to tomorrow’s market. Compute alone cannot guarantee victory—research quality, algorithms, data and engineering still matter—but access to it can determine how many ideas a lab gets to test.
The second reason is the cost of switching once AI becomes part of how a company works. Changing a chatbot is easy. Replacing the system connected to customer records, internal documents, approvals, compliance checks and everyday operations can be much harder. Employees need retraining. Integrations need rebuilding. The new system needs testing before anyone trusts it with important decisions.
The commercial prize is therefore bigger than selling the best model this month. It is becoming part of the customer’s working environment. A rival might offer cheaper answers, yet still struggle to displace a system that the organisation has spent years adopting. That is a reason to pursue market share aggressively, although common standards and the use of several model providers could weaken this advantage.
These two forces strengthen the optimistic case: investment could buy both faster technical progress and more durable customer relationships.
A strong strategy can still produce a dangerous market
Now take a simple, hypothetical transaction. A chipmaker invests $1 billion in a developer, which spends money on cloud capacity. The cloud provider buys chips from the investor. The hardware sale can be perfectly legitimate: real equipment changes hands and useful computing happens.
But the chipmaker has also put capital at risk. Recording revenue does not erase the cost of manufacturing the equipment or the possibility that its investment loses value. Somewhere along the chain, customers must eventually pay enough to cover the research, hardware, electricity, financing and replacement costs.
The problem is that several businesses may appear to confirm one another’s success while depending on the same forecast. The developer expects paying users. The cloud provider expects the developer to keep spending. The chipmaker expects both to expand. Investors value all three on the assumption that this growth continues.
If revenue disappoints, the effects can arrive together: lower company valuations, weaker demand for rented computing power, fewer chip orders and tighter financing. Guarantees that made expansion easier can become obligations just when cash is harder to find. A portfolio spread across several AI companies may contain much less diversification than its list of names suggests.
What the old bubbles tell us
The subprime mortgage comparison is useful because it shows how shared assumptions can hide shared risk. Rising house prices supported more lending; more lending supported house prices. When the cycle reversed, losses travelled through institutions that were more closely connected than many investors understood. The Federal Reserve’s history and the Financial Crisis Inquiry Commission’s report describe how debt, fragile funding and financial links amplified the collapse.
Today’s AI financing has different foundations. Equity investors can absorb losses without triggering the repayment demands associated with debt, and several major technology firms have substantial businesses outside AI. Calling the whole structure “another subprime” skips those differences. The comparison becomes more relevant as borrowing and guarantees spread the consequences of weak AI demand to lenders and infrastructure owners.
The dot-com infrastructure boom offers a closer warning. The internet transformed everyday life, but that did not rescue every investment made in its name. Telecom companies built more capacity than the near-term market could profitably use. SBC’s testimony filed with the SEC described excess capacity, falling prices and reduced investor returns after the spending surge.
AI adds a further possibility: the technology being financed can help perform the research and engineering needed to improve itself. That could accelerate the payoff. It also makes the race more intense, encouraging each participant to spend heavily for fear that a competitor will establish an enduring lead.
The fact that AI may change our lives cannot, on its own, settle the bubble question. The internet changed our lives too.
The strongest case for AI—and the hardest objection
My view is that AI will become an ordinary part of much of what we do. It will increasingly sit inside the tools we use to write, program, search, design and make decisions. Its most valuable applications may barely look like “AI applications” at all. They may simply feel like services that understand more, require less manual work and solve problems that previously took too much time.
That is a serious reason to build ahead of demand. Waiting until every use is obvious may leave a company supplying a market somebody else already owns.
Yet the opposing argument deserves equal force: making things easier to produce does not automatically make them worth producing. A model can generate another app, article, advertisement or presentation. It cannot guarantee that anyone needs it. If falling production costs mainly produce more interchangeable content and software, the industry may create a vast supply of material with little additional willingness to pay for it.
This is where the criticism of “AI slop” has economic substance. The bottleneck can move from producing something to deciding what deserves to exist: identifying a real problem, understanding the customer, exercising judgment and building trust. Human creativity and product sense remain valuable precisely because generating a plausible answer or a working prototype is getting cheaper.
There are measurable benefits already. Research on customer support agents found an average productivity improvement of roughly 14% from AI assistance. But gains vary by task and setting. A METR experiment using early-2025 tools found that a small group of experienced developers took longer on familiar projects when allowed to use AI. That result is specific to those tools and conditions; it is a reminder to measure the outcome rather than infer it from the amount of code produced.
A scientific breakthrough or an impressive demonstration can show what a model is capable of. A repeatable, profitable application must also work reliably, at an acceptable total cost, for customers who keep paying. There may be thousands of useful applications rather than one universal “killer app.” Their combined economics still have to support the infrastructure being built.
The bet the market is making
The AI industry is spending as though the technical lead and customer relationships secured today will become exceptionally valuable tomorrow. Circular financing helps it move faster. It also means a disappointment can damage several parts of the industry at once.
I believe the technological change will be profound. I am less convinced that all the companies financing it will capture enough of the value to justify their prices. Better models could make AI indispensable while competition makes it cheap. Customers could prosper while infrastructure investors struggle.
That is the test for this boom: does the next dollar of investment create something customers value enough to keep paying for? If it does, the financing loop can help build an industry large enough to repay the extraordinary confidence placed in it. If it mainly buys more capacity to produce more of the same, rising valuations are postponing the reckoning.
AI can change the world and still leave some of the people who financed it with enormous losses. Both outcomes can belong to the same revolution.
Figures above are dated company disclosures, including announced commitments and conditional guarantees. The analysis presents an investment thesis and its risks, rather than a recommendation to buy or short a particular security
.
Originally published on Substack on 26 September 2026.