Edition 11
In the AI economy, investment, supply, demand and financing keep appearing inside the same set of transactions. That makes one ordinary question surprisingly hard to answer.
There was a time when the relationship between a technology company and its customer was easy to describe. One built something and the other bought it. AI is producing something more complicated. The company making the chips may invest in the company buying the chips. The cloud provider may invest in the model company that commits billions of dollars to using its cloud. The model company may buy infrastructure from companies whose valuations depend partly on demand from the model company. Investors fund enormous computing commitments because they expect AI demand to grow, and the existence of those commitments is then presented as evidence that AI demand is growing.
None of this means the demand is artificial. The demand for AI infrastructure is plainly enormous. But the flow of money has become difficult to separate from the flow of technology, and that deserves close scrutiny.
In February 2026, OpenAI announced $110 billion in new investment at a $730 billion pre-money valuation, with $50 billion from Amazon and $30 billion each from NVIDIA and SoftBank. The same announcement expanded its infrastructure relationship with Amazon, made AWS the exclusive third-party cloud provider for its Frontier enterprise platform, and secured next-generation inference capacity from NVIDIA. Anthropic offers another example. In April, Amazon announced a further $5 billion investment, with up to another $20 billion tied to commercial milestones, on top of the $8 billion it had already put in. In return Anthropic committed more than $100 billion over ten years to AWS technologies and secured up to five gigawatts of Amazon compute capacity.
These are not small strategic partnerships at the edges of the industry. They are part of its structure. And once the investor, supplier, customer, infrastructure provider and distribution partner begin appearing repeatedly inside the same set of transactions, a simple question becomes surprisingly difficult to answer.
Where does the demand begin?
The Prompt
Start with an ordinary business relationship. A company has a product, customers want it, revenue grows, and the company invests to increase capacity because demand justifies the expansion.
Now change the sequence. A supplier believes its customer’s market will become enormous. The customer needs far more infrastructure than its current cash flow can support. So the supplier invests in the customer, provides financing support, guarantees parts of the infrastructure or helps make financing available. The customer uses that capital to acquire more of the supplier’s technology.
If the market grows as expected, everybody can win. The customer gains capacity it could not otherwise build quickly enough. The supplier secures a large future buyer. Investors gain exposure to the growth. Infrastructure gets built before scarcity constrains adoption.
There is nothing new about the underlying mechanism. Industries have used vendor financing for decades. Aircraft manufacturers support customers. Telecommunications equipment companies have financed buyers. Property development depends on commitments made before buildings exist. Semiconductor manufacturing itself relies on long-term supply agreements. What is unusual about AI is the scale, the speed and the concentration of these relationships.
NVIDIA’s position illustrates it well. The company sells the scarce technology at the centre of much of the AI boom. It also invests in AI companies and infrastructure providers that consume that technology. More recently it has begun using its own balance sheet to support infrastructure development at extraordinary scale. On 17 August, NVIDIA disclosed guarantees connected to an OpenAI-linked data-centre project in Pike County, Ohio, being developed by SoftBank-owned SB Energy. The filing describes residual value guaranties covering leases for roughly 4.25 gigawatts, capped in aggregate at $105 billion and triggered only if OpenAI defaults, with an option to extend support to a further 3.8 gigawatts. NVIDIA also invested $1.5 billion in SB Energy, will be the exclusive chip provider for the site, and OpenAI, which holds a stake in SB Energy, is leasing it for twenty years.
Analysts and journalists have started examining these structures through terms such as vendor financing and circular financing. Morgan Stanley credit analysts went further and called NVIDIA’s commitments balance-sheet-as-a-service, warning that the risks are hard to quantify. NVIDIA rejects the suggestion that any of this is unhealthy circularity. Its position is straightforward: frontier AI companies are constrained by access to compute, not by technology or customer demand, and helping strong customers obtain capacity is both a strategic investment and a source of future returns. In August it also signed memoranda with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to build financing platforms aimed at mobilising more than $500 billion of third-party capital, which several analysts read as a deliberate move away from funding customers off its own balance sheet. That may prove entirely correct. But it is a promise about the future, and remains to be seen before it is believed.
The most recent example arrived on 26 August, when The Information reported that NVIDIA had agreed to buy Hugging Face, the main public repository for open-source AI models, for $12.9 billion. Business Insider reported the same week that no agreement had been signed and that the talks could still fall apart, so the figure is reported, not settled. Two details are already visible. NVIDIA was already a shareholder, having backed Hugging Face in a 2023 round that valued it at $4.5 billion. And several outlets pointed out that owning a distribution hub with its own paid hosting business would give NVIDIA somewhere to sell computing capacity it has guaranteed on behalf of customers, should those customers not consume all of it.
If that reading is right, it is the loop acquiring an exit. A supplier that has underwritten demand for its own product buys a channel through which unsold capacity can be resold.
The harder question lies elsewhere. If a supplier invests in the customer, helps finance the infrastructure, sells the technology used inside it and later points to the resulting purchases as evidence of market demand, how should an outside observer read the signal?
The transaction is real. The capacity is real. The chips are real. The money is real. But the relationships underneath them are no longer independent.
The Mirage
Markets prefer signals they can read quickly. An order is taken as evidence of demand, a long-term contract as evidence of confidence, an investment as evidence of conviction, and a large infrastructure project as evidence of consumption still to come. Each of those readings is reasonable on its own. The difficulty in the AI economy is that the same set of companies is now producing all four at the same time, often within a single announcement.
Consider the Amazon and Anthropic relationship. Amazon is an investor in Anthropic. AWS supplies Anthropic with infrastructure. Anthropic has named AWS its primary training partner and uses Amazon’s Trainium chips. Claude reaches customers through Amazon Bedrock, so Amazon benefits when Claude usage increases on AWS. Anthropic benefits from Amazon funding the infrastructure required to build and operate Claude. Amazon benefits if Anthropic becomes more valuable. And Anthropic has committed to consume more than $100 billion of AWS technology over the coming decade.
None of those facts invalidates any of the others, which is why the relationship is hard to classify. The conventional distinction between investor, supplier, customer and distributor starts to blur.
OpenAI’s infrastructure relationships are becoming similarly interconnected. Amazon and NVIDIA participated in its February funding round while also securing major commercial relationships around compute. Microsoft remains a major shareholder and cloud partner, with OpenAI continuing substantial Azure commitments even as the partnership has been amended to allow greater cloud flexibility.
This creates a temptation at both extremes. One extreme says the entire AI economy is circular and therefore somehow fictitious. That is difficult to sustain. People and enterprises are using these systems at enormous scale. Revenue exists. Compute is being consumed. Businesses are paying for services. New capabilities are reaching production. The other extreme treats every investment, infrastructure commitment and capacity reservation as independent confirmation of unstoppable final demand. That deserves more careful examination.
A market can contain very real demand while also containing financing structures that amplify how quickly capacity is built around expectations of future demand. Both can be true at once, and that is what makes this moment hard to read.
The Reality Check
The loop itself is not the problem. The question is what happens if one of its assumptions changes.
These arrangements work well when demand keeps expanding, models keep improving, compute remains scarce, customers keep paying for increasingly capable services and capital remains willing to finance the next generation of infrastructure. Every participant has reason to keep building. The model company wants more compute. The chip company wants more model activity. The cloud provider wants more workloads. The data-centre developer wants long-term tenants. The investor wants exposure to AI growth. The enterprise wants access to whatever productivity advantage AI may create. Each reinforces the next.
That reinforcement can be productive. The modern technology industry would not exist without companies investing ahead of proven demand. Networks must be built before everybody can use them. Semiconductor fabs must be financed years before the chips arrive. Cloud capacity has to exist before a customer can consume it. But reinforcement also changes the risk.
Estimates of the build-out now run into the trillions. Morgan Stanley expects the largest cloud companies to spend around $3.5 trillion between 2026 and 2028, and tallies of off-balance-sheet obligations across the major hyperscalers have climbed from roughly $1.65 trillion in July to about $3 trillion, spreading exposure into private credit, insurance and other parts of the financial system. Both the Bank for International Settlements and the International Monetary Fund have described this as a financial-stability question rather than a single company’s problem, on the grounds that when a small cluster of firms hold stakes in each other and commit to buying from each other, nobody outside the loop can easily check whether the numbers are real. Some of the concern is no longer about whether AI works. It is whether demand develops quickly enough to justify the quantity of infrastructure being financed in anticipation of it.
A data centre is not a software licence. Power connections, buildings, cooling equipment, networking hardware and financing obligations remain after expectations change. So do the guarantees, the contracts and the chips.
The AI economy is therefore making a very large bet about the future. It is betting that intelligence will become cheap enough, useful enough and widely consumed enough to justify an industrial build-out unlike anything the technology sector has attempted before. Perhaps that bet will look conservative in ten years. Perhaps demand will outrun even today’s extraordinary plans. But the size of the bet deserves to be understood separately from enthusiasm for the technology itself.
The Leadership Question
For most companies, the practical question is not whether NVIDIA’s balance sheet carries too much exposure, or whether an AI infrastructure cycle resembles an earlier technology boom. The more immediate question is what assumptions their own AI investments inherit from this system.
Enterprise leaders increasingly decide on the expectation that AI capability will keep improving while becoming cheaper and easier to access. That assumption influences architecture and staffing. It influences whether companies build capabilities themselves or consume them through an API, which applications get rewritten around AI, how much proprietary data goes into systems that depend on external providers, and what they forecast about future productivity.
Those decisions can be entirely rational. But they are being made inside an ecosystem where providers are themselves spending enormous amounts of capital to ensure that enough computing capacity exists for the future they are simultaneously predicting. That does not make the prediction wrong. It means the signal needs context.
A hundred-billion-dollar compute commitment tells us that somebody believes an enormous amount of AI will be consumed. It does not tell us who the final customer will be, what they will pay, which applications will generate durable returns, or how much of today’s infrastructure demand is preparation for tomorrow’s expected demand. Those distinctions apply when a company starts writing its own five-year assumptions around the same future.
There is a second question. How many supposedly independent strategic choices ultimately depend on the same small set of upstream assumptions? A company may use several AI applications from several vendors and believe it has diversified. Those applications may depend on the same model provider. The models may depend on the same cloud. The clouds may depend on the same chip architecture. The infrastructure may be financed around the same expectations of future AI demand.
At the user interface, the ecosystem looks competitive. Further down, it can become remarkably concentrated.
The Closing Signal
The AI economy is real, and so is the extraordinary machinery being assembled around it. Money is moving into model companies. Model companies are committing that money to compute. Compute providers are investing in the companies consuming the compute. Chip companies are supporting infrastructure that will buy their chips. Cloud companies are financing model companies whose workloads strengthen their cloud businesses.
The loop can create genuine value. It can accelerate a technological transition that might otherwise take decades. It can make scarce infrastructure available sooner. It can allow companies with extraordinary technical capability but limited current cash flow to build at a scale previously available only to the largest corporations and governments. That is the optimistic case, and it is a credible one.
The same structure also makes the AI economy harder to read. When investment, demand, financing and supply reinforce one another, no single number tells the whole story. A chip order may represent present demand, anticipated demand or infrastructure built ahead of demand. A cloud commitment may reflect genuine consumption expectations while also supporting the valuation of the company making the commitment. An investment may demonstrate confidence in a customer while helping that customer afford more of the investor’s products. All of these can be true at the same time.
The mistake would be deciding in advance that the loop proves either a bubble or an inevitability. We do not know that yet. What we can see is an economic system increasingly capable of generating its own momentum. And when a system starts reinforcing itself, one question matters more than how quickly it is growing: how much of the demand is coming from outside the loop?
The Closing Signal
The loop can build real capacity. It cannot supply its own demand.