The Sustainability of AI Investment amidst High Interest Rates

One of the mysteries of recent macro history is why tightening monetary policy failed to significantly slow economic activity, particularly capital investment. For instance, even as the Fed funds rate rose, nonnresidential fixed investment rose throughout 2022-23. FIgure 1: GDP growth, q/q AR (black line, left scale), contribution from nonresidential fixed investment (blue bar), from imports of computer, computer accessories, semiconductors (green bar), rest-of-GDP (red bar), Fed funds rate (blue line, right scale). 2026Q2 growth rates from Atlanta Fed GDPNow of 7/10; computer imports for 2026Q2 estimated off of 2 months of trade data. Source: BEA, Federal Reserve, Atlanta Fed, and author’s calculations. This outcome — attributed to the boom in AI-related capex — has spurred the commonplace description of the US economy as being remarkably resilient. The question is whether this capex boom will continue, and by extension, the growth in the US economy. For recent quarters, the contribution to growth emanating from nonresidential investment has been offset by imports of computer and semiconductor imports — in an accounting sense. In 2026Q2, it looks like a sizable chunk of GDP growth is accounted for by nonresidential fixed investment, which is not offset by imports, although this is contingent on 2 months of import data accurately representing Q2 overall imports. That means continued US growth is highly dependent on what happens with AI related capex. What has the composition of nonresidential fixed investment in AI looked like over the past few years. We don’t have a direct apportionment of investment to AI, but we can look at the change in investment since 2022Q3 (taking the November 2022 release of ChatGPT as a kind of marker for AI capex). Figure 2: Change in investment in software (green bar), in information equipment (red bar), and power and communication structures (blue bar) since 2022Q3, in bn.Ch.2017$ SAAR. Source: BEA and author’s calculations. BIS estimates 2026 investment by US hyperscalers and other AI firms at $800 bn, up from about $750 bn in 2025. Is there any reason to believe that this level of investment won’t be reached? One leading indicator of nonresidential investment in the 2000 dot-com peak was imports of telecom and computing equipment (peak in latter was one quarter before peak in former). We only have data on information equipment investment through 2026Q1, while we have monthly data on imports of computers, computer accessories and semiconductors through May. Splicing series together, one obtains the following picture. Figure 3: Change in investment in information equipment (blue line), imports of computer equipment, parts and semiconductors (red line) since 2022Q3, in bn.Ch.2017$ SAAR. 2026Q2 observation is based on April and May import data. Source: BEA and author’s calculations. The 2026Q2 observation is just a guess, based on two months of data, which will be revised over time. It may very well be the case that imports continue to rise through Q2. However, there is reason to believe that there may be a downside surprise to AI-related capex. First, Magnificent 7 stock prices have come down, and are below peak. As the equity market cools, there are (at least) two effects. First, the cost of capital will rise for AI related firms (as well as others). Second, wealth effects supporting consumption will dissipate. For AI capex should slow relative to what would have otherwise occurred. Source: Bloomberg. Second, investment is now outstripping cash flow, so that firms now have to rely upon external financing, i.e., through the bond markets. Source: Economist. The hyperscalers are hence moving up the finance pecking order. Internal funds (cash flow) was relatively cheap, and less influenced by market rates. Notice that as the corporate bond rate goes up, now investment will decline (as shown in Figure 1 below). Source: Fazzari et al. (1988) as modified by Chinn. Accessing bond markets mean that investment will now be subject to a higher hurdle rate, more closely tied to bond yields — and at the same time some firms will have to pay a noticeable risk premium. (Below, Microsoft bond yields serve as risk free rate). Source: Economist. Hence, while AI-related capex seemed relatively immune to higher interest rates in the past, this may be less true going forward. And indeed uncertainty regarding future expected cash flow may weigh more heavily on investment than in the past. All the more reason to reduce policy uncertainty, and reduce upward pressures on interest rates emanating from cost-push inflation. Time for a little “on the one hand…on the other hand” chatter about tech sector debt. Understand, these generalizations have to do with tech firms overall, not specifically with data center or AI developers. First, the average maturity of tech-sector debt is fairly long, which helps insulate borrowers from interest rate fluctuations and liquidity problems, and free cash flow pays for the majority of tech-sector capital spending: https://www.ubs.com/global/en/wealthmanagement/insights/chief-investment-office/house-view/daily/2025/latest-18112025.html However, because of the high level of debt in the sector, there can be a lot of debt maturing any old time. Through 2028, there is something around $330 billion in high-yield bonds and leveraged loans are due: https://fenado.ai/articles/tech-sector-faces-330-billion-debt-wall-through-2028-amid-rising-rates-and-ai-scrutiny That may not include private credit. And while NVIDEA and a few others generate freighter-loads of cash, that’s not true for sellers of AI as an end product, and NVIDEA gets it’s cash flow from them. As a practical matter, as long as investors are hungry for AI-sector debt, the party can continue, and demand for big-name AI issuance has, till now, been quite strong. To state the obvious, if demand remains strong, the roll-over of maturing debt and the strong pace of AI investment can be financed without a hitch. Thing is, current plans are for even more AI debt issuance in coming quarters than we’ve seen so far, and there’s a big lump of maturing debt on top of that. In a rising rate environment, with returns on investment still quite thin? Tricky. Off topic – y’all noticed that the felon-in-chief fired the Democratic members of the Election Assistance Commission? One-time White House lawyer for the felon administration, Ty Cobb, states the obvious: “You can’t look at the termination of the Election Assistance Commission as anything other than another effort by one side to try to take over part of the referee role for the elections.” https://thehill.com/homenews/campaign/5963898-ty-cobb-donald-trump-midterm-influence/ The guy who tried to steal the 2020 election is now going to try to steal whatever he can of the 2026 elections. The Republicans on the Supreme Court gave him permission to steal. Free cash flow; I did not think it that much borrowing for mag 7 More borrowing suggests decline in relative free cash. Consider also that feee cash pays for depreciation expenses, less free cash means more borrowing for obsolescence. I think worry should look at EBITDA, tracking EBIT is too narrow. What if 75% EBIT should be sinking fund. Worse I think NVDA is projecting into compute demand that won’t operate. I suspect the grid needs more capital, to adjust for data center demands and grid variance. thanks to Prof. Chinn for putting into better words and pictures what I have been alluding to in the world of AI investment recently. until now, an enormous amount of AI investment was funding by the AI giants towards other AI giants. these were promises of investment in power plants, data centers, chips, etc. these were basically investment decisions made internally. but that source of capital is drying up-if it ever really existed at all? now you need to go to outside investors and defend your plan. just not sure how that is going to play out. there are enormous upfront costs for the physical infrastructure itself-and big time lags. now I am hearing some companies (see the karp rant recently) arguing that the big AI labs are charging too much for their tokens. I think there is going to be a cool off period, and investors are going to argue more money is available once you show me exactly how this is going to be profitable. the AI labs are now either going to show that this works, or poof, the bubble will burst. plenty of money will flow into AI again, if they can show this is a profitable path. SpaceX was an AI play, and they were wise to get out in front of Anthropic ad OpenAI. as the year rolls on, being an IPO in the AI market may be a tough position. it is not the same bubble as the dot com bust, but it may not yet be as profitable as some folks dream.