Crowding Out

In Downgrades and Competition For Capital I discussed how AI's insatiable appetite for funds was raising interest rates not just for AI-related debt, but also for everyone else. The impact of these rate increases will take time to become evident, as borrowers roll over existing debt or raise new debt.

Below the fold I look at impacts that we don't have to wait for, because they are already happening on a large scale.

Krugman Chart 1AI Versus Everything Else presents a different aspect of the competiton for capital — it is "crowding out" investment in everything else:

Evidence of that crowding out is revealed by a recent article in the New York Times, which noted that the AI Boom is not being slowed by these soaring rates. Yet the spike in interest rates is being felt throughout the non-AI economy. In other words, all that money going into construction of datacenters and purchasing foreign produced semiconductors comes at the expense of investment in everything else.

The chart at the top of this post, which I borrowed from Steve Rattner, clearly makes this point: construction of datacenters appears to be diverting funds and resources from the construction of housing, office buildings, and factories — more or less everything that isn’t AI-related.

That said, AI-related construction — building the structures that house datacenters — is only a fraction of the AI spending boom. Most of the money is going for equipment and software. In fact, that’s one main reason AI spending is so resistant to high interest rates: Borrowing costs matter much less for assets like GPUs, which depreciate rapidly in any case, than they do for long-lived assets like buildings.

Krugman Chart 2other 60% of the spending:

I look at business investment in “tech” — information processing and software — measured as a percentage of GDP. The blue line in the graph below shows how much that measure of tech spending has changed since the beginning of 2023. I compare that with the corresponding change in all other investment. This chart clearly reveals dramatic crowding out by tech spending of all other investment:

Krugman Chart 3different:

Chart 3 shows the same calculation as Chart 2, but applied to the tech boom of the 1990s, and it doesn’t show any crowding out at all — in fact, non-tech investment rose:

SourceCrowding out pits sovereigns against AI, Ryan Avent agrees with Krugman that the crowding out is happening:

Low interest rates imply that people will swap their claims on today’s productive potential for a song. In such cases, a debt-financed investment plan need not yield very much to pencil out. As rates rise, however, the expected return on an investment plan needed to justify the borrowing costs also goes up. Decent investments get squeezed out, then good investments, and then maybe really excellent investments that stand to deliver big economic benefits. So it goes.

We are experiencing this today, and you’ll never guess which part of the economy is responsible for the pronounced and sustained rise in desired borrowing.

But Avent argues that the "part of the economy" isn't AI but governments:

Anyway, gross government debt across advanced economies rose from 67 per cent of GDP in 2001 to 108 per cent today. So one could say that AI borrowing is crowding out other private investment, but one could also, and perhaps more reasonably, say that government borrowing is stymieing what might otherwise be a very broad-based, AI-driven, private sector investment boom.

If this turns out to be the better way to think about it, we’ll know soon enough. Firms in non-AI sectors may abandon efforts to claim more of the economy’s productive potential in the face of rising interest rates, but governments almost certainly won’t. For governments to give way, budget deficits need to fall. Accomplishing this means disappointing people who are expecting to receive claims from the government. It also means people have to give up more of their own claims to the government through higher taxes. In this fractious age, governments that do such things tend to be tossed out in favour of crazier governments which are even less able to do the needful.

Unfortunately, this likely means that we are on one of two paths. Either the AI complex ultimately smashes itself against the irresistible borrowing needs of the public sector and is forced to retrench, with potentially unpleasant consequences for growth, or rising interest rates cause one or another sovereign to slip into a market-destabilising crisis first.

The re-direction of investments toward AI would not be a problem if two conditions were satisfied. First, that the AI investments were more productive than those they crowded out. Skepticism on this question is warranted. There is a distinct lack of empirical evidence that AI is increasing productivity in the economy as a whole. Krugman expresses this skepticism:

Moreover, we shouldn’t assume that the private sector knows what it’s doing. There’s often a double standard when thinking about bad investments, in which public investments that go bad are treated as evidence that the government can’t do anything right, while private malinvestment is brushed off as no big deal.

Think about the $80 billion that Meta, formerly known as Facebook, spent on the “metaverse,” only to basically abandon the concept. A government program that spent $80 billion on a failed project would be the subject of endless Congressional hearings. Yet the waste from a failed private investment is equally real.

Second, that the enhanced productivity from the AI investments was at least as durable as that from those they crowded out. But the benefits of the 60% of the iAI investment that goes into IT hardware decay very rapidly, as I discussed in Mind The GAAP Again and its predecessors.

Krugman concludes:

So it’s clear that crowding out by the AI boom is happening on a very large scale. And unlike the risks of technological unemployment or a burst bubble, this isn’t a hypothetical risk. It’s happening right now.

Subsequently, Krugman also asked Will AI Crowd Out Jobs?, pointing out that a data center that might cost around $8B might only employ about 100 workers, mostly in low-skill jobs like security:

deploying $8 billion in capital, the cost of a data center, in a non-AI sector will, on average, support more than 10,000 jobs, while investing that money in a data center will support only around 100 jobs. Or to put it another way, money spent building data centers will, on average, create only around one percent as many jobs as investing the same sum in other industries would have created.

Citing this paper, Krugman explains the effect of the crowding out:

Van Nieuwerbergh estimates that through 2032 hyperscalers will spend more than $10 trillion on data centers — capital expenditure that, all the evidence suggests, will come at the expense of investment elsewhere in the economy. And because data centers employ so few people, a back-of-the-envelope estimate yields a number that made even my eyes bug out: the capital invested in data centers will employ roughly ten million fewer workers than that capital would have employed if invested elsewhere.

This doesn’t mean that unemployment will rise by 10 million. It means, instead, that reduced demand for labor will put downward pressure on wages and shift the balance of economic power, which has already been moving against labor, even more in capital’s favor. Even before the rise of AI, we were experiencing a rapid decline in the share of national income going to wages relative to profits, along with rising oligarchy. The anti-labor effects of AI will accelerate that trend.

For the AI companies' managements, this seems like a feature not a bug.

SourceBig Tech will fund more than a third of its AI investments with debt in 2027, Goldman Sachs predicts by Jake Conley, reports Goldman Sachs' estimates of the scale of the investments by the hyperscalers, and how much they expect them to borrow:

hyperscalers have turned increasingly toward debt markets. In 2025, investment-grade-rated global bond issuance from the hyperscalers totaled $108 billion, per Goldman Sachs, or roughly 26% of capex. Through the first half of 2026, those companies have already issued $194 billion of IG-rated debt, with that figure expected to reach roughly $250 billion in debt issuance, or roughly 33% of capex.

In 2027, Goldman Sachs said, the share of debt-funded capex is likely to rise even further, with $400 billion in expected global IG bond issuance from the hyperscalers against an estimated $1.14 trillion in capex, or a roughly 35% share.

If these investment plans materialize, the crowding out effect will greatly accelerate. In the 18 months from January 2025 it reduced non-AI construction spending by $132B, or about six week's worth. The chart suggests that spendig will double from around $700B to around $1.4T in 2028. That might equate to crowding out 12 weeks worth of non-AI construction in 2028, or about a quarter. That would be a significant hit to the broader economy

Fortunately, Ed Zitron's Credit Crunch provides a basis for predicting that:

the AI complex ultimately smashes itself against the irresistible borrowing needs of the public sector and is forced to retrench,

Zitron is skeptical that these investment plans will materialize:

In any case, AI data center debt is way more expensive now by virtue of the current state of treasuries, with costs compounded by the anxiety around them in general.

This means that any “virgin” projects — those that aren’t directly backstopped or co-signed by hyperscalers — are guaranteed to hit egregiously-high, 9% to 14% rates, which makes (as I discussed a few weeks ago in my two-part debt series) it near-impossible to make the already-questionable economics of running a data center work.

This means that any AI data center debt being raised right now is doing so under stricter credit conditions and doing so at unrealistic, unsustainable prices, if they’re going to be able to raise at all. This makes the $174 billion in debt that SB Energy needs to raise to fund its theoretical data center project with OpenAI — even when backstopped by NVIDIA — either horrendously expensive or impossible to complete, as does it mean that any future large, multi-gigawatt projects will add billions of dollars in interest payments to already-expensive debt.

This all makes pontifications that data center spending will increase to $32 trillion by 2050 equal parts stupid and wasteful, on top of the overall problem that there isn’t even enough demand right now for hyperscalers to break even on their 2026 and 2027 capex plans.

OpenAI and Anthropic have made huge, non-cancelable "take-or-pay" commitments to the hyperscalers. These form the basis of a vast amount of the hyperscalers' backlog. The compute these commitments cover is for both training, a cost, and inference, which generates revenue. In Thoughts on The Curve and the Future of AI Thomas Wright reported an estimate:

that the labs spend 58% of their compute on R & D and 32% on servicing customers.

Zitron estimates that OpenAI and Anthropic represent 80% of the total demand for compute. Thus about half of the total demand for compute is OpenAI and Anthropic's training and attacking other companies. Halving their internal compute budget would allow for about a 50% increase in compute for customers with no need for additional investment. This is why they are using the claim that their products are too dangerous to release to call for a pause and regulation.

SourceChinese catching up, halving their training budget wouldn't work because the demand doesn't exist:

As a result of their $1.3 trillion in compute commitments, neither company can become “capital efficient” by cutting their training costs, because said training costs are the only means of further growth outside of massive price increases that are unlikely to grow their businesses.

To make matters worse, both are cutting prices to compete with each other, and per Ramp, said price cuts aren’t increasing usage:

Price cuts, flat usage, cutting R&D and revenue disappointments aren't a good look when you're planning a $2T IPO or a $30B funding round at a $1.4T valuation. Skepticism about the projected capex spending, and the increased crowding out is reasonable. But the effects on the broader market of the bubble deflating will be severe.

In AI Bubble Risks Worst S&P 500 Crash Since 2008, Strategist Says, Levin Stamm and Sagarika Jaisinghani report that:

Panmure Liberum’s Joachim Klement said his base case is for that trade to disintegrate as soon as 2027, sending stocks sharply lower.

“My core conviction is that the AI bubble will either burst in 2027 or in 2028, so sometime in the next two years,” Klement said in an interview. Hyperscalers’ free cash flows are largely depleted, while the cost of debt is rising quickly and becoming prohibitive for these firms, he said.

Klement’s 2027 year-end target of 5,000 points for the S&P 500 implies 36% downside from current levels. ... Klement sees Europe’s Stoxx 600 falling to 430 points, more than 30% below current levels.

There are many more pessimistic forecasts. Hang on to your hats.