The AI build-out is now financed by credit

Everyone in markets has an opinion on whether artificial intelligence is a bubble. It is the wrong question to spend time on, because the answer only tells you how far prices fall. The more useful question is a plumbing question: if the AI build-out disappoints, whose balance sheet takes the loss?

The question people are asking, and the question that matters

For the first two years of this cycle the answer was simple and comfortable. Four American companies with large profits were paying for data centres out of pocket. If the spending turned out to be premature, the loss showed up as a lower share price. Shareholders own shares knowing they can lose money. The damage stopped there.

That is no longer how the build-out is paid for. Over the past eighteen months the money has increasingly come from borrowing, from loans secured against the chips themselves, and from private credit funds whose ultimate backers are life insurance companies. Each of those channels moves the risk from someone who chose to own an asset to someone who is owed a fixed sum on a fixed date. When the second kind of investor is disappointed, the consequences do not stay inside the technology sector.

I have been writing about the AI trade for most of this year, and the binding question keeps moving. In February it was capital, capex as a macro variable. In April it was labour. In May it was the price of compute itself. This month it is the financing, and I think it is the most important version of the question so far.

The size of the thing, in two numbers

Two figures give the scale without drowning anyone.

The first: in the first quarter of 2026 the four largest American cloud companies spent USD 129.8 billions on capital projects between them, roughly 80% more than the same quarter a year earlier (house capex tracker, rebuilt 28 August 2026). That is one quarter, four companies. The second quarter pushed harder still. Microsoft, Alphabet and Meta together spent USD 110.8 billions, up 98% on the same quarter of 2025, with Meta alone going from USD 19.0 billions to USD 30.1 billions in three months. Amazon, which spends more than any of the three, had not published its capital-spending line for the quarter in either feed the house tracker reads, so the four-company total is not yet complete.

The second: artificial intelligence now accounts for 15.77% of all physical business investment in the United States, structures and equipment combined, up 5.79 percentage points in a year (house capex-share tracker, 26 August 2026). On a trailing twelve-month basis that is USD 433.9 billions, growing at 72.7%. AI has supplied roughly 78% of the growth in American physical capital investment.

Put plainly: a single technology theme is now most of what American business is building. At that share of national investment it becomes a macro variable, and any macro variable of that size deserves a question about how it is funded.

The chart above tracks only the three largest spenders whose quarter we can see in full, and even that partial view climbs without a pause. What it cannot show is the rest of the market, and that is the part I would ask readers to sit with. Specialist data-centre developers, so-called neoclouds, sovereign projects and private operators are building alongside the giants. Almost none of them have thirty years of retained profits to draw on. They build with borrowed money, because borrowed money is the only money they have. The house tracker holds no verified series for that group, so I put no number on it here.

Three pipes, and what runs through each

The financing arrives through three channels, and they carry different risks.

Corporate bonds. The large technology companies have gone to the bond market at a pace they never used to need. In 2025 they issued more debt than in the previous seven years combined, while their cash relative to assets fell. This is the cleanest of the three channels. The borrowers are profitable, the debt is publicly rated and publicly priced, and if it goes wrong, you can see it going wrong in real time. I do not lose sleep over this pipe on its own.

Lending against the chips. This is newer and stranger. A data-centre operator buys graphics processors, then borrows against them, pledging the chips and the rental contracts they generate as collateral. In many cases the loans are then bundled together and sold on to investors as a security, which is what “securitization” means: a lender turns a pile of individual loans into a tradeable bond so someone else carries the risk.

Anyone who lived through 2008 recognises the shape of that sentence. I would still urge some care with the comparison. Bundling loans is a normal, useful technique, and it broke in 2007 for a specific reason: the underlying collateral, American houses, was mispriced, and the bundling hid that fact from the people buying it. The question to ask about chips is therefore the same question, applied honestly: what is this collateral worth if the borrower stops paying?

A building can be re-let for decades. A processor generation turns over in three to five years, and its rental value is set by a market that swings hard. In May, the on-demand rate for an Nvidia H100 sat near USD 1.50 an hour for four months and then roughly doubled to USD 3.00. That was a demand signal at the time and I read it as one. It doubles as a warning about collateral. An asset whose income can double in a month can halve in one, and the recovery value of a used chip five years out is an open question that no lender has yet had to answer in a downturn.

Private credit funded by insurance. This is the channel that concerns me most, and the least visible of the three.

Since 2008, regulation has pushed risky corporate lending out of banks and into private credit funds. More recently, private equity firms have paired those credit funds with life insurance companies they own. The logic is that a life insurer collects premiums today against payouts decades away, which is exactly the kind of patient money that suits a loan you cannot sell quickly. The insurer earns a higher yield, the fund gets stable capital that cannot demand its money back, and the group collects fees at both ends.

Roughly USD 750 billions of life insurance assets now sit within private equity’s reach, and insurers hold something like 10% to 15% of their assets in private credit. Much of that credit carries what the industry calls a private letter rating, a credit grade visible only to the state regulators and not to the public.

Two details are worth pulling out. The first is that the safety net is weaker than most people assume. Bank deposits in the United States are insured by a fund that collects money in advance and charges riskier banks more. Insurance is protected state by state, by a system that only collects money after a company has already failed, charges surviving insurers by size rather than by risk, and in about 34 states lets them claim the whole assessment back as a tax credit over five years. The bill, in the end, is public, and no legislature ever votes on it.

The second is that the opacity is not theoretical. Guggenheim’s Delaware Life restated its holdings of assets affiliated with its own owner from a reported 3% to 5% up to 40%. That is close to a tenfold revision: the same balance sheet, described two ways, and it shows how fast this can surface.

I would not claim that AI infrastructure is the dominant use of insurance-linked private credit. I have not seen data that would let anyone say so. What I would claim is narrower and still uncomfortable: the fastest-growing borrower class in the economy is being served by the least transparent lending channel in the economy, and the backstop behind that channel is thinner than the one behind banks. I hold the size of the overlap as a research question, not as a settled number.

What a Z-score is, and what the live model is saying

The house macro model does something simple with a large amount of data. For each economic indicator it asks how unusual today’s reading is against that indicator’s own history, and expresses the answer in standard deviations, written as sigma. Zero means perfectly normal. Plus one sigma means hotter than roughly five readings in six. The model then groups indicators into categories and takes an aggregate.

The point of doing it this way is that it strips out narrative. The model does not know there is an AI story. It only knows whether the banking data looks like itself.

Three readings carry this article, all from the live model computed on 27 August 2026. Each value is the category’s distance from its own mean, in standard deviations, the one the chart above plots.

Banking System sits at +1.40σ against its own mean. Among the categories with a reliable number of live inputs, this is the hottest thing in the model. Banking activity, by its own historical standard, is running warm.

Liquidity sits at +0.34σ. This is the most completely observed category in the entire model, every input live, and it says the monetary environment is ordinary. There is no flood of new money.

Credit and Consumer sits at +0.03σ. Household credit conditions are sitting almost exactly on their own long-run average.

Read those three together and the picture is specific. Banking is hot while liquidity is normal and household credit is unremarkable. A broad, everybody-is-borrowing credit boom would light all three. Only one is lit. Whatever is running through bank balance sheets is concentrated, and it is not coming from the consumer and not from an abundance of money.

Why neutral liquidity makes this worse rather than better

There is a comforting way to read a neutral liquidity score, which is that nothing is going wrong. I read it the other way round.

Michael Howell’s framework, which I lean on for this, treats the world as a giant refinancing machine rather than a savings machine. Most borrowing gets rolled over into a new loan when it comes due. What matters is whether the pool of available money is growing fast enough to absorb the debt that comes due each year.

On the house measure, American non-financial debt now stands at 361% of broad money as of the first quarter of 2026. Debt grew 5.7% over the year while that money pool grew 4.6%, a gap of 1.1 percentage points (house debt-liquidity tracker, 24 August 2026). Debt is pulling ahead of the money available to refinance it, slowly and steadily.

The timing is the uncomfortable part. The cheap borrowing done during the pandemic was mostly termed out to about five years, which brings it back for refinancing in bulk, and the peak of that corporate and sovereign wall falls between 2026 and 2028. Into that window walks the largest private capital-spending programme in modern American history, looking to borrow.

The chart above shows the long end of the curve that borrowing runs into. Three-month money is at 3.86% and thirty-year money is at 5.28%, as of 18 August 2026. The curve slopes up, so the longer you borrow, the more you pay, and a data centre is a fifteen-to-twenty-year asset. Every operator financing one is buying money at the expensive end of that line.

Meanwhile the Federal Reserve’s policy rate sits at 3.63%, which after inflation leaves a real rate of about +0.33%: mildly restrictive. The central bank is not going to inflate this problem away in the next few quarters unless something forces it to.

So the setup is a record borrowing requirement, arriving during the heaviest refinancing years of the decade, at the most expensive point on the curve, with the money supply growing more slowly than the debt. None of those four facts is dramatic on its own. They compound.

How a credit drawdown actually differs

The reason to care about the funding mix is that equity losses and credit losses behave differently in an economy, and the difference is mechanical.

When a share price falls, the loss lands on whoever owns the share. They feel poorer and may spend less, which is real but slow and diffuse. The company keeps operating. Nobody is legally required to do anything.

When a loan goes bad, three things happen quickly. The lender’s capital falls, which reduces how much it can lend to everyone else, including businesses with nothing to do with AI. The borrower must find cash on a fixed date, which usually means selling assets into a market that has just been told those assets are impaired. And every other holder of similar loans has to re-mark their own book, whether or not their particular borrower is in trouble.

That third step is the transmission channel, and it is why the identity of the lender matters more than the size of the loss. A USD 50 billions equity loss is a bad year for a group of shareholders. A USD 50 billions credit loss spread across banks and insurers tightens lending conditions for the whole economy, because those institutions are regulated on their capital and must respond by shrinking.

Which brings the argument back to the model. If the AI build-out were still equity-funded, a disappointment would show up first as a fall in technology valuations. Funded the way it now is, a disappointment shows up first in bank credit quality, in the marks on private credit portfolios, and in the reserves of insurance companies that hold them. That sequence is worth holding on to, because it also tells you the early-warning indicators are the boring ones: lending surveys, credit spreads, insurer disclosures, and the banking category of a macro model.

For a European or Swiss reader, the exposure is less remote than it looks. European banks and insurers have been buyers of American private credit for years, in search of yield they could not find at home. A credit event that starts in American data-centre lending arrives here through the asset side of institutions that most people would describe as conservative. The reflex response, which I would expect rather than predict, is the familiar one: a bid for the Swiss franc, a bid for gold, and a widening of the gap between companies that fund themselves from profits and companies that fund themselves from markets. That last gap is what a quality preference means in practice. In a regime where credit is the transmission channel, quality describes which companies can decline to go and ask for money.

What would tell me I am wrong

Three developments would weaken the argument, and I would want to say so publicly if they appeared.

If AI revenue starts covering AI capital spending, the debt is simply early rather than excessive, and early debt against a real cash flow is ordinary corporate finance. The tell would be operators funding new builds out of contracted revenue instead of new facilities.

If the banking category cools back toward zero over the next few monthly readings while the build-out continues, then the warmth I am pointing at was something else, and my reading of it was wrong.

And if central banks re-expand their balance sheets meaningfully, the refinancing wall stops being a wall. Abundant money forgives a great deal of leverage, for a while.

What to watch next

The build-out will continue regardless of what any of this analysis concludes, because the companies doing it believe compute is the binding constraint on the decade and they are probably right about that. The machines get built. The open question is who is holding the paper when the first cohort of chips reaches the end of its useful life and someone has to write down what that collateral is worth.

This is markets research and commentary written for a general readership. It is not personalised investment advice and does not take account of any individual’s circumstances.