The U.S. economic data for the previous few months is wanting decidedly meh. The latest employment numbers have been so dangerous that Trump truly fired the head of the Bureau of Labor Statistics, accusing her of manipulating the numbers to make him look dangerous. However there’s one enormous brilliant spot amid the gloom: an unimaginable AI information heart constructing growth.
AI know-how is advancing quick, threatening (promising?) to upend many sectors of the financial system. No one is aware of but precisely who will revenue from this growth, however one factor that’s sure is that it’s going to take quite a lot of computing energy (or “compute”, as they are saying). AI fashions take quite a lot of compute to coach, however these days additionally they take quite a lot of compute to do inference — i.e., to “take into consideration” and reply every query you ask. Inference compute now represents most of the cost of working superior AI fashions, and will increase in inference compute are accountable for many of the ongoing performance gains. So compute wants are in all probability solely going to develop as AI retains getting higher.
Whoever offers this compute goes to make an enormous quantity of income. Whether or not which means they’ll make quite a lot of revenue is one other query, however let’s desk that for proper now; you possibly can’t make revenue in the event you don’t make income. So proper now, tech corporations have the selection to both sit out of the growth fully, or spend large and hope they’ll work out how one can make a revenue.
Roughly talking, Apple is choosing the former, whereas the massive software program corporations — Google, Meta, Microsoft, and Amazon — are selecting the latter. These spending numbers are fairly unimaginable:
For Microsoft and Meta, this capital expenditure is now greater than a 3rd of their whole gross sales.
Right here’s Chris Mims of the WSJ:
The Magnificent 7 tech corporations have collectively spent a report $102.5 billion on capex of their most up-to-date quarters, almost all from Meta, Alphabet (Google), Microsoft and Amazon. (Apple, Nvidia and Tesla collectively contributed a mere $6.7 billion.)…
Investor and tech pundit Paul Kedrosky says that, as a proportion of gross home product, spending on AI infrastructure has already exceeded spending on telecom and web infrastructure from the dot-com growth—and it’s nonetheless rising. He additionally argues that one rationalization for the U.S. financial system’s ongoing power, regardless of tariffs, is that spending on IT infrastructure is so large that it’s performing as a kind of private-sector stimulus program…
Capex spending for AI contributed more to growth within the U.S. financial system prior to now two quarters than all of client spending, says Neil Dutta, head of financial analysis at Renaissance Macro Analysis, citing information from the Bureau of Financial Evaluation. [emphasis mine]
Right here’s that chart from Kedrosky, who has been doing an excellent job following this story because it unfolds:
Facet be aware: I’m truly unsure Kedrosky is true in regards to the peak of the telecom growth. His chart exhibits the extent of capex for 2020, when 5G and fiber infrastructure was being constructed out. However Doms (2004) exhibits U.S. telecom capex reaching $120 billion in 2000:
That may have been round 1.2% of U.S. GDP on the time — about the place the information heart growth is now. However the information heart growth continues to be ramping up, and there’s no apparent motive to assume 2025 is the height, so Kedrosky’s level nonetheless stands.
I believe it’s necessary to have a look at the telecom growth of the Nineties fairly than the one within the 2010s, as a result of the previous led to a gigantic crash. The railroad growth led to a gigantic crash too, in 1873 (earlier than the funding peak on Kedrosky’s chart). In each instances, corporations constructed an excessive amount of infrastructure, outrunning development in demand for that infrastructure, and suffered a devastating bust as expectations reset and loans couldn’t be paid again.
In each instances, although, the massive capex spenders weren’t improper, they have been simply early. Finally, we ended up utilizing all of these railroads and all of these telecom fibers, and way more. This has led lots of people to speculate that large funding bubbles may truly be beneficial to the economy, since manias go away behind a surplus of low-cost infrastructure that can be utilized to energy future technological advances and new enterprise fashions.
However for anybody who will get caught up within the crash, the long run advantages to society are of chilly consolation. So a lot of people are worrying that there’s going to be a crash within the AI information heart business, and thus in Huge Tech on the whole, if AI business income doesn’t grow fast enough to maintain up with the capex growth over the subsequent few years.
A knowledge heart bust would imply that Huge Tech shareholders would lose some huge cash, like dot-com shareholders in 2000. It will additionally sluggish the financial system immediately, as a result of Huge Tech corporations would cease investing. However the scariest chance is that it will trigger a monetary disaster.
Monetary crises are likely to contain financial institution debt. When a monetary bubble and crash is generally a fall within the worth of shares and bonds, everybody takes losses after which simply kind of walks away, a bit poorer — like in 2000. Jorda, Schularick, and Taylor (2015) survey the historical past of bubbles and crashes, they usually discover that debt (additionally referred to as “credit score” and “leverage”) is a key predictor of whether or not a bubble finally ends up hurting the actual financial system. They write:
Utilizing a complete dataset, overlaying a variety of macroeconomic and monetary variables, we show that it’s the interplay of asset value bubbles and credit score development that poses the gravest danger to monetary stability. These outcomes, based mostly on long-rung historic information, supply the primary sound statistical help based mostly on giant samples for the broadly held view that the monetary stability dangers stemming from of an unleveraged fairness market growth gone bust (such because the U.S. dotcom bubble) can differ considerably from a credit-financed housing growth gone bust (such because the U.S. 2000s housing market).
They use the phrases “unleveraged” and “credit score”, however what they actually imply right here — and the best way they define their variable for credit score development — is particularly financial institution loans, not bonds. Within the 2008 crash, many of the debt was held by banks, in a single kind or one other. Thus, a wave of defaults threatened the solvency of the banking system, inflicting all the financial system to freeze up.
However the banking system wasn’t in any actual hazard of collapse from the dot-com/telecom crash, as a result of banks hadn’t lent some huge cash to individuals concerned with the tech business, so there weren’t quite a lot of loans to go dangerous. The telecom corporations had borrowed loads, however primarily by way of the bond markets fairly than from banks, which on the time have been extra targeted on housing and small enterprise. Households additionally didn’t turn into considerably extra indebted within the 90s (relative to their revenue), which might be why there wasn’t a protracted and painful interval of family deleveraging following the dotcom growth, the best way there was after 2008.
So if we consider this primary story of when to be afraid of capex busts, it signifies that now we have to care about who’s lending cash to those Huge Tech corporations to construct all these information facilities. That method, we will work out whether or not we’re frightened about what occurs to these lenders if Huge Tech can’t pay the cash again.
Paul Kedrosky has an inventory:
The place is all this capital coming from?
For probably the most half, six sources:
Inside Money Flows (Main for Microsoft, Google, Amazon, Meta, and many others. )
Debt Issuance (Rising function)
Fairness & Comply with-on Choices
Enterprise Capital / Personal Fairness (CoreWeave, Lambda, and many others.)
SPVs, Leasing, and Asset-Backed Autos (like Meta’s latest)
Cloud Consumption Commitments (largely hyperscalers)
And The Economist writes:
[C]apex is rising quicker than [Big Tech’s] cashflows…The new centre of the AI growth is shifting from stockmarkets to debt markets…Through the first half of the yr investment-grade borrowing by tech corporations was 70% larger than within the first six months of 2024. In April Alphabet issued bonds for the primary time since 2020. Microsoft has diminished its money pile however its finance leases—a kind of debt largely associated to information centres—almost tripled since 2023, to $46bn (an extra $93bn of such liabilities should not but on its balance-sheet). Meta is in talks to borrow round $30bn from private-credit lenders together with Apollo, Brookfield and Carlyle. The marketplace for debt securities backed by borrowing associated to information centres, the place liabilities are pooled and sliced up in a method much like mortgage bonds, has grown from nearly nothing in 2018 to round $50bn at this time…
CoreWeave, an ai cloud agency, has borrowed liberally from private-credit funds and bond traders to purchase chips from Nvidia. Fluidstack, one other cloud-computing startup, can be borrowing closely, utilizing its chips as collateral. SoftBank, a Japanese agency, is financing its share of an enormous partnership with Openai, the maker of ChatGPT, with debt. “They don’t even have the cash,” wrote Elon Musk when the partnership was introduced in January. After elevating $5bn of debt earlier this yr xAI, Mr Musk’s personal startup, is reportedly borrowing $12bn to purchase chips.
A few of these funding sources don’t appear that harmful, within the macroeconomic sense. When Huge Tech or different corporations spend their very own money, points inventory, or subject bonds, it seems to be extra just like the telecom growth of the Nineties than the railroad growth of the 1800s.
However what about all this “non-public credit score”? These are the possibly scary half. Personal credit score funds are principally corporations that take funding, borrow cash, after which lend that cash out in non-public (i.e. opaque) markets. They’re the debt model of personal fairness, and lately they’ve grown quickly to turn into one of many U.S.’ financial system’s main classes of debt:
Personal credit score funds take a few of their financing as fairness, however additionally they borrow cash. A few of this cash is borrowed from banks. In 2013, only one% of U.S. banks’ whole loans to non-bank monetary establishments was to personal fairness and personal credit score corporations; today, it’s 14%. A latest be aware by Berrospide et al. of the Federal Reserve exhibits the rise in financial institution lending to personal credit score:
BDCs are “Business Development Companies”, that are a kind of personal credit score fund. If there’s a bust in non-public credit score, that’s an acronym you’ll be listening to loads.
And I consider the graph above does not embody bank purchases of bonds (CLOs) issued by non-public credit score corporations. If non-public credit score goes bust, these financial institution property will go bust too, making banks’ steadiness sheets weaker.
A recent article by Fillat et al. of the Boston Fed argues that financial institution lending to personal credit score funds may pose a systemic danger to the banking system:
The meteoric rise of personal credit score presents necessary questions in regards to the function of banks going ahead and the implications for stability within the US monetary system…Our evaluation of Federal Reserve and proprietary loan-level information signifies that the expansion of personal credit score has been funded largely by financial institution loans and that banks have turn into a key supply of liquidity, within the type of credit score traces, for PC lenders. Banks’ in depth hyperlinks to the PC market might be a priority as a result of these hyperlinks not directly expose banks to the historically larger dangers related to PC loans.
The authors level out that the majority of those are very short-term loans — therefore safer than long-term loans like those that sunk banks within the monetary disaster of 2008. And most of them are senior loans — if non-public credit score goes bust, banks receives a commission out first. However they warning:
[B]anks would endure losses [on their private credit lending] solely in severely opposed financial circumstances, resembling a deep and protracted recession. However losses might additionally happen in a much less opposed situation if the default correlation among the many loans in PC portfolios turned out to be larger than anticipated—that’s, if a larger-than-expected variety of PC debtors defaulted on the identical time. Such tail danger could also be underappreciated.
If all of the non-public credit score funds are lending to information facilities, then their correlations are in all probability fairly excessive — if there’s a bust in AI, quite a lot of them will go bust without delay. This might be the “tail danger” that the Boston Fed people are frightened about, and it looks as if there’s an opportunity it might harm the U.S. banking system.
A personal credit score bust additionally might harm insurance coverage corporations, who’re the principle LPs within the funds, however who additionally lend cash to the funds. For instance, Carlino et al. (2025) present how life insurers have principally turn into a type of financial institution that borrows cash from institutional traders and lends it to personal credit score:
They be aware, ominously, that “life insurers’ publicity to below-investment-grade agency debt has boomed and now exceeds the business’s publicity to subprime residential mortgage-backed securities in late 2007.”
I’m not precisely positive how systemically necessary life insurance coverage corporations are; they’re intertwined with the remainder of the monetary system by various channels, and it’s laborious to inform how vital these challenges are. Nevertheless it looks as if the likelihood is there for some systemic results. In the meantime, different forms of insurers are beginning to lend to personal credit score as properly. Do not forget that AIG, an insurance coverage firm, was one of the most important bailouts within the monetary disaster of 2008.
So after I take a look at this complete panorama, it appears to me that a number of the primary circumstances of a monetary disaster are at the very least beginning to fall into place:
-
Now we have an enormous story about why “this time is completely different” — the concept that AI will change the whole lot, and that information facilities will thus earn enormous returns.
-
Now we have a big and growing quantity of debt getting used to fund one single sector of the financial system (information facilities), which means that the loans’ default likelihood might be extremely correlated.
-
Now we have an opaque nook of the monetary system (non-public credit score) that has just lately grown from a tiny piece of the system to a really vital piece.
-
Now we have systemically necessary lenders (banks, and probably insurance coverage corporations) enmeshed within the new sector in a mess of the way.
Up to now, the hazard doesn’t scream “2008”. However in the event you wait till 2008 to start out worrying, you’re going to get 2008. It’s good to start out worrying early. Jamie Dimon, CEO of JP Morgan Chase, is not waiting to start sounding the alarm, warning that non-public credit score might set off the subsequent monetary disaster, at the same time as his personal firm expands into the non-public credit score market. As a former Citibank CEO said after 2008, “So long as the music is enjoying, you’ve obtained to stand up and dance.”
