When the chipmaker starts building the financing too

The economic debate about AI is usually narrowed to whether the current investment wave is a bubble or the beginning of a new industrial era. The two possibilities are not mutually exclusive, however. A technology may be genuinely significant while too many of the associated assets are built, too quickly or with financing that is too expensive. The important signal is therefore not the enthusiasm itself, but the way in which the financing of AI infrastructure is changing.

In August 2026, Nvidia announced that it was working with six major financial participants on platforms that could eventually mobilise more than $500 billion of external capital for AI infrastructure. Under the plan, Nvidia compute could become an independently financeable asset and loan collateral. For now, the announcement represents an intention to cooperate, not a portfolio of final agreements, and Nvidia is not automatically guaranteeing the full amount. In selected transactions, however, it may assume part of the risk associated with the hardware’s future residual value.

This is not the classic role of a co-borrower, but it goes far beyond a simple supplier relationship. One of the world’s most profitable hardware manufacturers, with one of the strongest market positions, is not merely selling its product; it is also helping to organise the financial system from which customers can buy more of its products. If the return on AI investment and its financeability were self-evident, this would not appear necessary at first sight. The structure therefore does not prove that the investment wave will collapse, but it shows that technological superiority alone is no longer enough: a separate financing structure must be built behind the next wave of demand.

When a rapidly ageing asset becomes collateral for a long-term loan

Two separate problems meet in AI infrastructure. The first is technological. A data centre building, its power supply and its network can be used for decades, but the economic value of the accelerators operating inside it may change over a much shorter period. The older hardware does not have to become physically unusable. It is enough for a newer system to perform the same task substantially more cheaply, faster or with less energy. The older asset still works, but may generate less future revenue and is therefore worth less as loan collateral too.

The second problem is financial. Large technology companies are increasingly using bonds, private credit, project companies and long-term capacity-purchase commitments. The Bank for International Settlements calls structures in which the loan does not necessarily appear legally on the technology company’s balance sheet “shadow debt” when the cash flow is nevertheless sustained by the company’s multi-year lease, purchase or guarantee obligations.

Long-term financing thus relies on a rapidly changing technological asset. If utilisation is high, the price of computing capacity is adequate, Nvidia’s system remains dominant and the market for used accelerators works, the structure may be sustainable. But if new hardware, a cheaper competitor or a more efficient model pushes down the economic value of the old capacity, the collateral and cash flow may weaken at the same time. A lower residual value can produce higher lending risk, higher risk can lead to more expensive financing, and more expensive financing can result in weaker returns.

Why is the investment requirement growing faster than visible revenue?

One reason for the growth in capital expenditure is a genuine capacity shortage: cloud providers report demand in several areas that exceeds their available computing capacity. The second reason is rising unit cost. With memory, accelerators, networks and energy becoming more expensive, the same additional spending does not necessarily produce the same additional capacity. The third reason is competition itself. Behind a demand-led investment there is already a customer; behind a strategic investment is the assumption that whoever fails to build capacity now may lose their market later.

This brings spending forward. Companies have to build data centres and hardware before the full end-market demand and revenue can be demonstrated. In a market where a few participants may capture most of the revenue, excessive upfront investment may also be rational for individual companies: waiting may appear more dangerous than the risk of excess capacity.

According to the plans available on August 22, 2026, the combined annual investment of Amazon, Alphabet, Microsoft and Meta is approaching $750 billion. This amount is not spent exclusively on generative AI, and it is not a completed annual figure. The scale nevertheless permits a simple stress test. If a single $740 billion investment vintage has to earn a return within four to five years at a cost of capital of 8%, it must recover $185–223 billion a year. If $50 of every $100 of revenue remains after operating costs to service the capital, this requires $371–447 billion of annual revenue. If only $30 remains, the requirement rises to $618–745 billion.

In August 2026, the annualised revenue rates reported by OpenAI and Anthropic amounted to approximately $105 billion combined. This is not profit, a completed annual figure or the revenue of the entire AI market. As a reference point, however, it shows the scale of the difference: if this revenue alone were to support the investment, it would have to grow by 3.5–7.1 times for the four-to-five-year scenario. Other sources of revenue may also contribute, including cloud services, enterprise software, advertising efficiency and internal productivity gains. The stress is therefore not captured by a simple ratio between capital expenditure and the revenue of the two AI labs, but by the system’s need to turn a multiple of currently visible revenue generation into durable cash flow that is also sufficient to service capital. If similar investment vintages follow one another, the capital-recovery requirement will accumulate too.

Apple points to a different business path

Behind current data-centre investment is the assumption that a significant share of future AI use will remain a centralised, paid computing service. Training large models and performing the most complex tasks may indeed require data centres. Part of the everyday use of completed models, however, may move onto the user’s own device.

Apple’s M5 Max already has the memory and computing performance required to run local large language models. The higher memory figures for the M7 are media reports for now and therefore cannot be treated as specifications for a finished product. The direction is nevertheless important: if models become more efficient and personal computers more powerful, more tasks can run locally, without recurring cloud charges and with more direct control over data.

Apple’s model does not imply a future without the cloud. Processing may take place locally by default, with only the more complex task sent to a central system. This matters commercially because another cloud provider merely rearranges the paid computing market, whereas local processing may eliminate the paid cloud transaction itself for certain tasks. AI can therefore be technologically successful and widely used even if the return on part of the central infrastructure being built today falls short of expectations.

The financial constraint may become a social and government constraint

Data centres require not only capital and hardware, but also energy, grid connections, water, land and local political acceptance. In the United States, several states and municipalities are restricting the construction of large data centres or attaching new conditions because of energy prices, pressure on the grid, water use and environmental effects. This is not a general AI ban, but every delay, obligation to generate power independently or lost tax incentive increases the cost of the investment or postpones the start of revenue.

The broader question in the financing debate is not whether the money will suddenly run out either. Quantitative easing after 2008 did not create an unchanged tank of money from which capital is now simply flowing into AI. The point is that a large stock of nominal financial wealth is seeking real returns, while the risk and balance-sheet capacity of insurers, pension funds, banks and bond funds is finite. The constraint therefore appears first not in the disappearance of money, but in a higher price for the next dollar, a demand for stronger collateral and stricter conditions.

At the same time, AI infrastructure is seeking finance in the same long-term capital market in which governments finance growing debt as well as defence, energy and social spending. This is not mechanical crowding out: a dollar allocated to a corporate bond is not necessarily missing from a government bond. But the price relationship is real. Higher government bond yields make corporate financing more expensive, while record corporate bond supply may require higher spreads if investors’ balance sheets become saturated.

AI’s return also has a social precondition. For annual investment of several hundred billion dollars to become sustainably profitable, AI must be embedded deeply in production, services and everyday work. This may create new tasks and occupations, but it may also reduce demand for other jobs. If the transition increases unemployment or income inequality, the state may face higher social spending and weaker revenue based on labour income. The same state whose capital market finances AI infrastructure may therefore also have to bear the social costs of the transformation.

The bubble is not the ultimate question

The significance of Nvidia’s programme is not that it proves the imminent collapse of the AI investment wave. It shows that the next bottleneck may already be capital itself, available at the right price and for the right term. The GPU is transformed from a technological asset into loan collateral, while the manufacturer acquires an interest not only in selling the hardware but also in maintaining the financeability of its customers.

The system’s return cannot be inferred separately from Nvidia’s profit, the growth of the large cloud providers or the revenue of the AI labs. The entire chain is sustainable only if the ultimate benefit to businesses and consumers is sufficient to produce all the intermediate operating costs and capital returns. Meanwhile, it is uncertain not only which accelerator will be competitive in five years, but also where the computation from which today’s data centre must earn its return will take place.

The central question, then, is not whether AI is a bubble. It is whether the real economic cash flow created by AI can grow quickly enough to generate the return required by an increasingly large and complex financing system before technological change reprices the assets behind it. If it cannot, the problem may appear not only in share prices, but also in collateral values, credit spreads, project refinancing and the financeability of the next technological generation.

The purpose of Unus Multorum is not to tell readers what to think about a particular event. Its purpose is to reveal the connections that allow every reader to form their own conclusions.