The AI buildout is bidding up the interest rate on itself

When the Federal Reserve raised its benchmark rate last month to a range of 3.75 to 4 percent, Governor Lisa Cook named a culprit that was not in the inflation story two years ago. Data centers, she explained, compete for the same inputs everyone else uses — construction labor, electricity, water — each up roughly 5 percent over the past year, and she said the inflation pressure may be broadening. ^[1] The boom that readers have followed as a technology story has crossed into the central bank's reaction function. It is now big enough to move the price of everyone's money — including its own. The mechanism is not subtle. Ayako Yoshioka, a portfolio manager, put it plainly: this cycle's buildout is larger than the 1990s internet expansion, and it is physical in a way software never was. ^[2]

this time around, it's a very physical, build out that is creating inflationary impulses in the overall economy. — Ayako Yoshioka

The scale shows in the grids it touches. In Ireland, data centers consumed nearly a quarter of the country's electricity last year. ^[3] Here is the hinge where the thread turns back on itself. The industry pushing up the price level is also the economy's biggest new borrower. More than 40 percent of new long-term investment-grade corporate debt is AI-related, by Apollo chief economist Torsten Slok's count, and he projects a funding gap of about a trillion dollars that the traditional bond market cannot absorb. ^[4] BNP Paribas expects hyperscalers to sell roughly $400 billion of bonds next year, inside a record $3.7 trillion of net fixed-income supply. ^[5] The rise the buildout provoked lands on the buildout's own paper. The market is already pricing that. High-yield risk premiums hit a five-month high after nearly $600 billion of AI-related debt supply this year. ^[6] The repricing has redrawn the hierarchy: TSMC bonds now trade at lower yields than Meta's. ^[6][7] PIMCO's Christian Stracke attributes the widening spreads to something other than credit quality.

It certainly looks like at least one more to come from the Fed. — Christian Stracke

The arithmetic that follows is unforgiving. Bain estimates the hyperscalers need $4.2 trillion in new revenue within five years for the bet to pay. ^[8] Meanwhile Amazon, Nvidia, Meta and Broadcom have moved chunks of their chip financing off the balance sheet — special vehicles, convertible notes, a half-trillion-dollar funding platform — and analysts warn the debt may outlive the useful life of the depreciating chips it buys. ^[9] Central banks have begun counting it formally: roughly $450 billion of AI-related debt by early September, on a path toward $4.1 trillion by 2030. ^[10]

With the industry rapidly issuing large volumes of debt, its importance in public and private credit markets is expected to grow, mirroring the trend observed in equity markets and expanding the range of investors exposed to the AI investment boom. — Reserve Bank of Australia

The revenue that must pay for all this is real and compounding. OpenAI's annualized revenue is approaching $70 billion; Google's rose 24 percent year over year, with a backlog now nearly four times trailing revenue. ^[11][12] The problem is not that nothing is being sold. It is what the selling is gated on. ServiceNow's enterprise AI maturity index found spending up 110 percent while the maturity score rose to just 51 out of 100 — and the adoption curve, the report says, stalls not because the technology failed.

The adoption curve stalls not because the technology failed but the data underneath it did. — Holly Briedis

The tell is where the new money goes. AWS is putting $1 billion into forward-deployed engineers, and roughly $9 billion has been pledged across Amazon, Microsoft, OpenAI and Anthropic for deployment work since May. ^[13] That is not more compute. It is people, hired to carry the integration the models cannot do on their own — the one bottleneck capital cannot buy off. Nobody planned any of this. The Fed is fighting inflation, the bond desks are demanding yield, the hyperscalers are building. Each move is locally reasonable; the loop is what their sum produces, and that is why it holds. That is the context for the strangest detail in the record. In the same rates analysis that lists an AI-buildout slowdown among the drivers of a ten-year Treasury yield possibly heading toward 6.6 percent, OpenAI and Anthropic have themselves argued for a slower pace of infrastructure activity — even as the four largest spenders lift combined capital spending toward $738 billion this year. ^[14] The biggest spenders are asking the buildout to slow down, and raising their own capex, in the same breath. Both things are happening at once.