AI capex and the US economy
We have written a few times recently about the common if somewhat shapeless worry that if the AI boom were to turn into a bust, it would take the US economy into recession. A recent X post from the Harvard economist Jason Furman gives the worry more of a shape. He wrote:
Investment in information processing equipment & software is four per cent of GDP. But it was responsible for 92 per cent of GDP growth in the first half of this year. GDP excluding these categories grew at a 0.1 per cent annual rate in the first half.
Here are his charts:
We walked through Furman’s numbers and they look right. For a sense of scale, here is tech spending and AI superscaler capex as a percentage of nominal GDP (note Furman’s numbers are in real terms):
Unhedged has made two optimistic counterpoints to the dreary AI bust/recession hypothesis. First, the tax law that comes into effect next year allows investment projects to be fully depreciated in the first year. If this tax break could spark a wider capital spending boom, that could pick up where AI leaves off. Next, the build-up in tech investment as a share of the economy has been more gradual than the upwards bursts in spending during the housing or telecom bubbles. It might be part boom and part permanent secular change.
Dario Perkins of TS Lombard, arguing in a recent note that “AI is NOT the thing that is keeping the US economy out of recession,” makes similar points, and adds several other sharp ones. A lot of the equipment going into data centres is imported, he notes, so there will be offsetting negative contributions to GDP elsewhere in the national accounts. And recessions, Perkins argues, are not defined by GDP; they “are a very distinct process that takes place in the labour market. We have not seen that dynamic in 2025, and that has nothing whatsoever to do with AI capital spending. Big tech capex is NOT the reason the US labour has stayed afloat.”
Unhedged is quite sympathetic to these views — but also with another, more bearish point Perkins makes. If the returns on massive AI investments turn out to be low, that could cause a serious stock market correction. In other words, the mechanism by which an AI bust could cause a recession is not through suddenly lower growth (the national income statement, as it were) but through asset writedowns (the national balance sheet).
Perkins thinks we don’t need to worry about the AI bubble in asset values quite yet, however, because AI assets are not highly leveraged:
Leveraged bubbles (which typically involve property markets) are far deadlier than unleveraged bubbles (in stock markets, tulips, Bitcoin etc). The reason is simple. Falling asset prices will damage balance sheets, and if debt (which is fixed in nominal terms) has increased a lot, that can trigger a painful dynamic of asset fire sales and forced deleveraging . . . there is not (yet) a lot of leverage in the current AI capex bubble
We hope this is right, but we are not totally confident. It is true that the big tech companies are mostly financing their AI investments out of free cash flow. At the same time, we hear a lot about private debt investment in data centres, and we have deep respect for the financial system’s ability to conceal leverage in unexpected places, especially when it is in the grip of a “next big thing” narrative. “The AI bubble could become more dangerous if it continued to inflate,” Perkins warns. We think there might be quite a lot of air in it already.