Who’s paying for the AI boom?
Following the money behind artificial intelligence leads somewhere surprisingly familiar.
We tend to think of suppliers, customers and investors as separate parts of an economy. In artificial intelligence, those distinctions are becoming rather difficult to maintain.
In January 2026, Nvidia invested another $2 billion in a company called CoreWeave. Unless you’ve been following the extraordinary amount of money being spent on artificial intelligence, the name might not mean very much. CoreWeave operates data centres, the enormous buildings filled with processors that make modern AI possible. Many of those processors are manufactured by Nvidia.
Nvidia was investing in one of its own customers, which isn’t particularly unusual. Businesses have been helping their customers expand for generations, especially when doing so means selling them more products.
What caught my attention was another agreement between the two companies, announced a few months earlier. Nvidia had agreed to purchase certain computing capacity from CoreWeave if CoreWeave couldn’t sell it to anyone else. The arrangement was initially valued at $6.3 billion and extended, subject to conditions, into 2032.
In other words, Nvidia manufactures the processors, invests in the company purchasing them and has agreed to become a customer for some of the resulting computing capacity if other buyers don’t materialise.
Then I started looking at who else was involved.
In September 2025, Nvidia and OpenAI announced a proposed partnership involving an investment of up to $100 billion. By early 2026, reports were discussing a substantially smaller investment of around $30 billion. The original proposal and the subsequent investment discussions were not necessarily equivalent arrangements, and neither should be confused with equipment purchases, future computing contracts or money already borrowed. Yet when announcements like these are discussed together, it becomes surprisingly easy to mistake overlapping commitments for separate sources of economic activity.
OpenAI was also one of CoreWeave’s largest customers. By September 2025, CoreWeave had announced computing agreements with OpenAI worth up to approximately $22.4 billion. At the end of 2025, CoreWeave reported around $21.6 billion in debt and $60.7 billion in remaining performance obligations, representing contracted revenue for services it had yet to deliver.
Those numbers describe different things, of course. Debt has to be repaid, while future revenue depends on services being delivered and customers meeting their obligations. But the relationship between them interested me more than the figures themselves. CoreWeave was borrowing enormous sums to build infrastructure for companies developing artificial intelligence, while the company supplying much of the equipment was also investing in the business.
And once I began following those relationships, CoreWeave stopped looking like an isolated example.
In February 2026, SpaceX acquired Elon Musk’s xAI, bringing its artificial intelligence business into the same group as its launch and satellite operations. Nvidia had participated in xAI’s $20 billion funding round the previous month, and SpaceX subsequently began selling computing capacity to other AI developers, including Anthropic.
In October 2026, the Financial Times reported, and Reuters subsequently confirmed, that SpaceX was seeking around $40 billion to purchase Nvidia processors. The proposed financing comprised $10 billion in bank loans and $30 billion in investment-grade debt. The deal had not been finalised.
SpaceX is a very different business from CoreWeave. It launches rockets and operates satellite networks, giving it sources of revenue beyond artificial intelligence. But the relationships were beginning to look familiar. Nvidia invests in an AI developer, that developer becomes part of a larger infrastructure business, and the business seeks billions in financing to purchase more Nvidia processors.
The financial markets appeared to notice. Following the reports, the cost of insuring SpaceX’s debt against default reached a record high, its longer-dated bonds weakened and its shares fell. SpaceX had floated only four months earlier, in June, so both shareholders and lenders were responding to the same proposed borrowing.
I found myself wondering where I’d seen something like this before.
Around the turn of the millennium, telecommunications equipment manufacturers such as Lucent and Nortel were helping finance the companies buying their equipment. The arrangement made commercial sense. Customers needed expensive infrastructure, manufacturers wanted to sell it, and financing helped bring the two together. But when the telecommunications boom collapsed, some of those customers could no longer meet their obligations. The manufacturers discovered that part of the demand supporting their sales had depended on financing they themselves had helped provide.
There is no suggestion that today’s AI arrangements are improper, and the financing structures are not identical. What interests me is the question they raise. When suppliers invest in customers, customers borrow to buy more equipment, and the resulting capacity is sold to other companies within the same expanding network, how much demand is coming from outside it?
There’s another way demand can disappoint, even if nobody stops using AI.
For years, businesses were encouraged to move their infrastructure into the cloud. Why maintain expensive servers and data centres when somebody else could provide computing power whenever it was needed? For many organisations, the economics were compelling. Then businesses began examining the bills more closely. Storage costs accumulated, transferring data became expensive, and predictable workloads sometimes turned out to be cheaper on infrastructure they already owned.
Get Paul Thomas’s stories in your inbox
Join Medium for free to get updates from this writer.
The cloud didn’t fail. Companies simply became more selective about how they used it, and some brought workloads back in-house.
Artificial intelligence may face a similar adjustment. Businesses are buying subscriptions and integrating increasingly powerful language models into their operations, but not every task requires the most sophisticated model available. Classifying documents, extracting information from invoices and answering routine internal questions can often be handled by smaller models, sometimes running on relatively modest local hardware.
That doesn’t mean demand for AI computing will disappear. Cheaper computing could encourage far greater adoption. But it does mean that the price businesses are prepared to pay for individual tasks may change considerably, even while their use of artificial intelligence continues to grow.
For companies financing data centres and processors over many years, that distinction matters. Imagine borrowing billions to purchase equipment expected to generate revenue for a decade, only for a more efficient architecture to deliver comparable results at a fraction of the cost. The processors haven’t stopped working, and the debt hasn’t disappeared. It’s simply become harder to earn the revenue originally anticipated.
The telecommunications boom offers another lesson here. Companies invested heavily in fibre-optic networks because they correctly predicted extraordinary growth in Internet traffic. What many failed to anticipate was how improvements in technology and overbuilding would drive down the price of capacity.
There is another complication, one that has less to do with technology than geography.
In March 2026, Amazon Web Services reported physical damage to facilities in the United Arab Emirates and Bahrain during the conflict involving Iran. Drone strikes damaged infrastructure, disrupted power supplies and affected cloud services, prompting AWS to advise customers to move workloads to other regions.
We tend to talk about the cloud as though it exists somewhere beyond the physical world. In reality, it consists of buildings, processors, electricity supplies and telecommunications connections. Some of those connections depend on undersea cables passing through politically sensitive waters, while much of the world’s most advanced semiconductor manufacturing is concentrated in Taiwan.
A large technology company with facilities across several continents may be able to absorb a regional disruption. A smaller infrastructure provider carrying substantial debt against a limited number of facilities may find that more difficult. The service can be interrupted, but the financial obligations remain.
And that brings us back to the money.
The largest technology companies, including Microsoft, Alphabet, Amazon and Meta, finance much of their AI expansion from enormous operating cash flows. That gives them considerably more room to absorb an expensive mistake. Specialist infrastructure providers relying heavily on borrowing have less room for disappointment, and the institutions financing them may be exposed to the same underlying assumptions about future demand.
Banks, credit funds and institutional investors provide capital to these businesses, sometimes using money that ultimately belongs to pension savers and ordinary investors. That doesn’t establish that AI poses a systemic financial risk. We would need to know far more about who holds the debt, how concentrated those exposures are and whether losses could spread beyond the technology sector.
Interestingly, the International Monetary Fund has now begun examining precisely these questions.
In its October 2026 Global Financial Stability Report, the IMF examined 14 publicly disclosed AI financing transactions dating back to 2023. Between them, these arrangements involved approximately $300 billion in equity investments, $650 billion in debt financing and $1.4 trillion in commitments for future services and equipment.
Those figures describe different financial relationships, not a single pool of money being spent three times. But their scale is extraordinary, and the connections between them are perhaps more revealing than the amounts themselves.
The IMF identified what it calls wrong-way risk, a situation in which financial difficulties affecting one company could simultaneously undermine investments in that business and the revenues other companies expect to receive from it. In other words, a company struggling to meet its commitments might create problems not only for its lenders and shareholders, but also for the businesses that had been counting on its future spending.
That sounds remarkably familiar.
The IMF isn’t predicting an inevitable collapse. It acknowledges that many of the largest technology companies remain highly profitable and that current financial stability risks appear manageable. Its concern is that increasing borrowing, concentrated investment and interconnected financial commitments could make an eventual correction more difficult to contain.
And that brings us back to something the telecommunications industry discovered a quarter of a century ago. A technology can succeed, demand can continue growing and the infrastructure can remain useful, while the financial assumptions supporting its construction prove considerably less durable.
There are important differences between the two periods. Today’s largest technology companies are considerably more profitable than many of the businesses financing the telecommunications boom, and AI already has practical applications across numerous industries. Those differences matter. They may make the current expansion more resilient, even if they cannot guarantee that every investment will succeed.
Perhaps the most interesting possibility is that everyone involved is broadly correct about artificial intelligence. Nvidia may be right about the future of computing. OpenAI may be right about the demand for intelligent software. Infrastructure providers may be right that enormous amounts of computing capacity will be needed.
The IMF’s warning doesn’t settle those questions. It does, however, make the financial relationships between the companies rather harder to ignore.
The Internet succeeded spectacularly. Some of the companies that financed its arrival did not.
