Who Ultimately Supplied the Money?

Simply Seeking Balance · The Distance Problem · Part 2 of 4

A companion to The Distance Problem. Developed with Quinn; editorial revision by Enzo.

The AI investment boom produced numbers so enormous that eventually I stopped being impressed by them.

Chips and data centers, long-term cloud commitments, special-purpose vehicles, private credit, supplier financing, power contracts, institutional investors, and public support. The announcements seemed to arrive faster than I could sort out their relationships.

After a while, every announcement began to sound like proof of the announcement before it.

Enormous spending seemed to prove enormous demand, while expectations of enormous demand justified still more spending.

That may be perfectly rational. It may also be worth examining.

The question that finally interested me was embarrassingly simple:

Who ultimately supplied the money?

Not who announced the project. Not which company booked the revenue. Not which subsidiary signed the contract.

Whose capital actually became exposed?

Follow the money backward far enough and eventually someone must be taking a genuine risk. Sometimes the answer is easy.

A profitable company uses cash generated by its existing businesses to buy computing capacity from an independent supplier. The supplier earns revenue. The customer receives something it believes will produce more value than it cost.

That is ordinary investment.

But the structure becomes harder to interpret when the participants begin occupying several positions at once.

Imagine a chipmaker investing in a model developer that commits to buying computing capacity using those chips. A cloud company invests in the developer while selling it computing services. A special-purpose vehicle borrows to buy servers and leases them onward; an insurer or pension fund holds some of the debt. A utility builds the connection, and surrounding households may bear part of its cost, depending on the rate arrangement.

Every transaction can be legitimate. But they are not necessarily independent evidence of demand. The same economic relationship can appear several times in the financial story.

That is shared exposure. If the customer, supplier, lender and investor all depend on the same underlying enterprise succeeding, what appears to be diversification may be several claims resting on the same foundation. A related, narrower problem is wrong-way risk: the exposure becomes more dangerous just as the counterparty becomes less able to pay. Source: Basel Framework definitions.

One arrangement makes the overlap concrete. Reuters reported on October 1, 2026, citing Anthropic’s IPO prospectus, that Broadcom had agreed to lend Anthropic up to $42 billion for infrastructure spending. Broadcom was also a hardware supplier and equipment lessor. That does not make demand fictitious. It does mean that financing and future supplier revenue cannot simply be counted as independent confirmation of it.

This does not mean the foundation is weak. It means we should know that it is the foundation.

There is another problem: timing.

I learned something about that when I bought a Tesla Model 3 in early 2020.

Before buying, I asked about rumors of major battery changes. I was willing to wait rather than purchase immediately before a significant upgrade. I was reassured that nothing of the sort was imminent.

Then, in September, Tesla announced its new 4680 battery cell and structural battery pack.

I was annoyed. Naturally.

But the interesting part is what happened afterward. The technology took years to reach production vehicles. The announcement did not establish when—or whether—that particular system would reach the Model 3.

The announcement was real. The proposed improvement was real engineering work. But the timetable, manufacturing difficulties, and eventual products were different questions.

In hindsight, waiting for that announcement would not necessarily have given me a better Model 3.

Now enlarge the problem from one car to a data center containing billions of dollars of processors.

A company buys chips early because capacity is scarce. The electrical connection is delayed, so the processors sit in a warehouse. Meanwhile, a new generation becomes faster or more efficient, software improves, and customers reconsider what they need.

Interest continues accumulating while the equipment produces no revenue.

What began as prudent advance purchasing can become expensive inventory before it ever goes to work.

And the opposite mistake is possible too: waiting for the next technological improvement can mean missing years of useful production from equipment available today.

That is why technological progress and investment timing cannot be treated as the same question.

That question matters because hype does not need to create fake demand to create a boom. It can pull real future demand into the present.

Companies afraid of future shortages reserve capacity now. Customers sign long-term contracts, developers borrow against them, and suppliers expand factories to meet the orders.

Every participant has evidence that demand is enormous because everybody else is behaving as though demand will be enormous.

Again, nobody needs to be lying. Expectations can manufacture their own temporary confirmation. Eventually reality must settle the account.

Were the data centers completed?

Did electricity arrive?

Did customers remain obligated to buy?

Did the hardware remain useful?

Did revenue exceed the financing costs?

Did the underlying AI produce enough economic value to justify everything built around it?

That is why I am less interested in asking whether AI is “a bubble.” AI could transform civilization and still bankrupt investors.

Railroad and internet buildouts both offer historical reasons to distinguish technological importance from returns earned by particular investors.

A technology's importance and an investment's price are separate propositions.

The same financial abstraction that can hide overlapping exposure also makes it harder to see what the capital eventually builds. That becomes particularly interesting for pensions.

Pension portfolios may hold AI-related stocks and debt directly or through broad funds. Concentration can grow without any beneficiary making a deliberate decision about AI. How much exposure exists depends on the particular portfolio; the point is that diversification does not itself tell a beneficiary what the money is doing.

This revives an investment question I used to discuss with clients decades ago:

Are you investing for profit, or are you investing according to principle?

I am vegan. I would not knowingly choose to invest directly in a cattle company.

But what about an index fund containing one?

What about a pension fund containing that index?

What about a bank whose asset-management division owns the cattle company?

At some sufficient degree of removal, nearly everyone is economically connected to nearly everything. So perhaps the useful question is not whether one's money touches something objectionable.

It is:

At what degree of removal does participation cease to be meaningfully my decision?

Voting with one's wallet is different from voting at the ballot box. If I decline to buy beef, I withdraw my own money. If I leave a company, I withdraw my participation.

If I vote to prohibit someone else from buying beef, I am exercising collective authority over that person's choice.

Those acts are related. They are not interchangeable. And investment itself has similar degrees of agency.

There is a difference between deliberately buying shares in a company because I want its particular work funded and owning the company incidentally through a broad diversified pension portfolio.

Financial abstraction makes diversification possible. It also creates distance from what investments eventually become: factories, servers, mines, power plants, warehouses, jobs, pollution, products, animals, and loans. Someone eventually does something with the money.

So my conclusion is not that complicated finance is inherently corrupt or that AI investment is fraudulent.

It is more modest.

Before treating spending as evidence of success, follow the transaction until you can distinguish new outside demand from commitments financed within the same network. Money need not literally move in a circle for the exposure to be shared.

Ask who supplied it. Ask what they were promised. Ask what must happen before they are repaid.

Ask what happens if the technology succeeds more slowly than the financing requires. Ask whether the asset can still earn enough if the next generation arrives sooner than expected. Ask who absorbs the loss.

And perhaps ask one ethical question alongside the financial ones:

Do I understand well enough what my capital enables to regard the resulting return as entirely detached from the thing being funded?

The farther money travels from judgment, the easier it becomes to mistake exposure for understanding.