I spent some time this vacation with a short BIS report called Financing the AI boom: from cash flows to debt. It’s only six pages, but it gave some really useful perspectives on the current build-out.
The part that stuck with me wasn’t whether AI is “a bubble”. The report doesn’t really make that argument. It is more interested in the shift happening underneath the excitement: AI is becoming an infrastructure and financing story. It always has been in a way. Either way, it changed the questions worth asking.
A software story is mainly about capabilities, adoption and margins. An infrastructure story is also about construction, power, utilisation, depreciation and financing. It is about how much has to be built before the demand is fully known, who pays for it, and what happens if the revenues arrive later than expected.
The spending is already large enough to move the economy
By mid-2025, US spending on data centres and IT manufacturing facilities had reached roughly 1% of GDP. Since 2022, those investments have contributed an average of around 0.4 percentage points to GDP growth.
Broader IT investment, including other equipment and software, has reached 5% of GDP. That is higher than its previous peak around the dot-com boom. In some recent quarters, this broader category accounted for almost half of US economic growth.
There is an important distinction here. It would be misleading to say that “AI produced half of US growth”. The larger number includes a much broader category of IT spending. Some of the equipment going into data centres is also imported, meaning the figures may overstate the amount of domestic value being created.
Even with those qualifications, the basic picture is striking. A relatively small group of companies, acting on a shared view of future computing demand, is making investment decisions large enough to influence national growth figures.
But investment spending and productivity are obviously not the same thing.
When a data centre is built, the construction work and equipment purchases enter GDP now. Concrete gets poured, servers are installed, cooling systems are connected and electrical capacity is added. There is economic activity regardless of whether the facility eventually earns an attractive return.
The productivity question gets answered much later than the economic impact.
It depends on whether businesses find valuable uses for all that computing capacity, whether those uses generate enough revenue, and whether the gains spread beyond the companies building and supplying the infrastructure.
This is one reason I find the current comparison with the dot-com period slightly misleading. The question is not simply whether today’s investment is larger or smaller.
The BIS points out that the composition is different. Much of the dot-com investment boom came from companies buying and using IT. The present boom is being driven more heavily by companies producing AI capacity.
In other words, a great deal of supply is being installed before broad-based demand has been demonstrated.
That doesn’t mean the demand won’t appear. But it does mean we can observe the investment boom well before we can observe the productivity boom it is supposed to enable.
The companies building it can still afford it, but the gap is widening
The large technology companies driving the build-out have historically carried much less debt than other US companies. They have been able to finance investment from the enormous cash flows generated by advertising, cloud services, software subscriptions and ecommerce.
That has made the first stage of the AI boom unusually self-financing.
The report’s chart covers Alphabet, Amazon, Meta, Microsoft and Oracle. Their combined quarterly capital expenditure is approaching $120 billion, while free cash flow is closer to $60 billion.
I don’t read this as evidence that these companies are suddenly in financial trouble. They remain highly profitable and entered the cycle with unusually strong balance sheets.
What it shows is that the investment programme is growing faster than the pool of money available to fund it internally.
That creates a gap. And the gap increasingly has to be filled with bonds, leases, loans and other forms of external financing.
There is nothing inherently alarming about that. Debt is a normal way to finance expensive assets that are expected to produce revenue over many years. If a data centre will operate for decades, spreading its cost over time can make good economic sense.
But debt changes the clock. Both in type and in clock speed.
A company investing its own cash can wait for the return, reduce future spending or accept that a particular project did not work out.
Debt arrives with its own timetable. Interest has to be paid. Loans mature. Covenants have to be respected. Assets may need to be refinanced under whatever market conditions exist at the time.
The technology may eventually succeed, but the cash flows still have to arrive on that financing schedule.
That is a different kind of risk.
The collateral may not be as solid as it sounds
The physical nature of the build-out can make the financing sound reassuring. These aren’t speculative apps with no tangible assets. They are data centres, grid connections, servers and power infrastructure. Physical things you can point at.
But the composition of a data centre matters.
The rule of thumb used in the report is that the physical building accounts for only about one quarter of the cost of a new data centre. The other three quarters is IT equipment.
A building may have a long useful life. GPUs, servers and networking equipment are another matter.
They wear out, consume large amounts of energy and risk being overtaken by new generations of hardware. Their economic life may be considerably shorter than the life of the building that contains them.
The value of the facility also depends on things that aren’t captured by the word “property”: access to electricity, cooling, network connections, the requirements of particular tenants and the continued usefulness of its equipment.
So when AI lending is described as infrastructure finance, I think it is worth asking what kind of infrastructure is actually securing the loan.
A railway line does not normally become obsolete because a faster railway line is released two years later. Computing equipment lives in a different world.
Private credit is becoming part of the story
Anyone not living under a rock has probably read or heared something about how private credit is spreading is web into financing and comes with its own set of problems. That is a whole different rabbit hole we wont go into here, so lets focus on the present case.
Private-credit lending to AI, cloud and big-data companies has grown from almost nothing to more than $200 billion. It now represents close to 8% of total outstanding private-credit volumes.
The BIS estimates that it could reach somewhere between $300 billion and $600 billion by 2030. That is a wide scenario range, based on an equally wide range of possible investment growth. The direction is more useful than the precise value here.
Private credit is, in some ways, well suited to this market. Data-centre projects come with construction risk, power constraints, unusual collateral and dependence on a small number of large tenants. A privately negotiated loan can be adapted to those conditions more easily than a standard bond.
The individual loans are already large. The average AI-related private-credit loan in the BIS data is $169 million, compared with $90 million for loans to companies in other sectors.
At the same time, the report doesn’t present this as an immediate systemic crisis.
Around 20% of private-credit funds have some exposure to AI-related companies. For the average fund, AI-related lending still accounts for only about 5% of outstanding volumes. Many private-credit funds also lock up investor money for several years, making them less vulnerable to the kind of sudden withdrawals that can destabilise banks or open-ended funds.
So the risks here may be slower and less visible: refinancing problems, repeated loan extensions, disputes over collateral values or losses that take time to surface. Also the lack of transparency on the structuring of these loans.
There is also the question of where the borrowing ultimately sits. Leasing arrangements and project vehicles can move obligations away from the most visible corporate balance sheets. That may change the appearance of the leverage without changing the underlying economics.
The exposure can look dispersed across funds and projects while still depending on the same small group of tenants, the same expectations about AI demand and the same assumptions about future hardware utilisation.
Debt and equity seem to be telling different stories
The number I found hardest to make sense of concerns the price of the loans.
Private-credit loans to AI-related companies carry an average spread of 6.2 percentage points over the reference rate. For other private-credit borrowers, the figure is 6.1 percentage points.
The average maturity is also nearly identical: 4.7 years for AI-related borrowers and 4.8 years for everyone else. AI loans are roughly as likely to be secured as other loans.
Lenders, in other words, are not obviously charging more for exposure to the AI build-out.
This sits strangely beside the equity market.
AI-related companies have been valued on the assumption that they will capture unusually large future cash flows. The private-credit market appears to be pricing their loans as if the risk is fairly ordinary.
I don’t think this proves that one side is clearly wrong.
Debt and equity are different claims. A lender has limited upside but is paid before shareholders and may benefit from collateral, covenants or guarantees. Equity investors absorb more downside in exchange for potentially much more upside.
The borrower groups in the report are also not identical. The cash-flow analysis looks at five major technology companies, while the private-credit data cover a much broader collection of AI, big-data and cloud businesses. This is not a clean comparison of the bonds and shares of the same companies.
Still, the difference tells some story here.
Perhaps equity investors are too optimistic about the size of the eventual profits. Perhaps lenders are underestimating construction, utilisation and obsolescence risk. Perhaps the loans are structured in ways that genuinely protect lenders from most of the downside.
Or perhaps all three things are partly true. Perhaps none!
What the numbers show is that different parts of the financial system are placing very different bets on how this investment cycle will end.
It doesn’t have to be a bubble to end badly
The BIS is fairly measured in its conclusion.
At roughly 1% of GDP, the narrow AI infrastructure boom is similar in size to the US shale boom and only around half the size of the increase in IT investment during the dot-com period. The commercial-property boom in Japan and the mining boom in Australia were more than five times larger relative to their economies.
By that measure, the current build-out does not look like an economy-wide financial-stability event.
But the size of an investment boom does not tell us everything about the size of the adjustment that follows it.
Across the historical episodes examined in the report, GDP growth slowed by more than one percentage point on average after the boom ended. The sharpest slowdown followed the US dot-com boom, even though it was not the largest investment boom in the comparison.
These examples of course cover different industries, countries and financial systems. But they point to a distinction that is easy to lose when talking about “transformative technologies”.
A technology can matter enormously while the companies financing its expansion earn disappointing returns.
The internet did not disappear after the dot-com crash. The underlying infrastructure and technical advances continued to reshape the economy. That did not prevent investors from financing too much capacity, at the wrong prices, on the wrong timetable.
I can imagine something similar being true of AI.
AI may become a general-purpose technology, or it may not. Demand for computing may continue to rise. The data centres being built today will support products and businesses that do not yet exist.
None of that guarantees that every investment being made now will generate enough revenue, quickly enough, to justify its cost.
That is what I took from this little report. Not a hot take by any measure, but the reading did clarify add nuance.
The AI boom is no longer only a bet on model intelligence (that bet is lost given the catch-up rate of open weight models is like.. three months now?). It is becoming a bet on utilisation rates, electricity prices, hardware lifetimes, refinancing conditions and the speed at which future demand turns into cash.
The next phase may depend less on another jump in benchmark performance and more on something fairly ordinary: whether revenues arrive quickly enough to cover electricity, depreciation and debt service.