Earlier this week, in “Food For Thought,” I touched on the increasingly topical interplay between hundreds of billions in high-grade debt sales to fund AI expenditures and Treasury’s issuance roadmap.
Scott Bessent would probably like to cut coupon auction sizes next year if possible in an effort to alleviate persistent oversupply concerns. AI-related corporate bond issuance may give him an excuse, or at least an incentive.
Indeed, last week’s subtle language tweak in the QRA, which introduced two-way “risk” around auction sizes come mid-2027, might’ve represented an early effort to socialize the narrative.
Although supply overhang’s not the hot button issue for the Treasury market it was in late 2023, the term premium remains ~45bps wide versus where it was when 10-year US yields were 5%, as illustrated below.
Whether cutting coupon auction sizes would make any real difference considering the constellation of factors which “explain” the term premium is debatable, but less supply certainly wouldn’t hurt.
Put differently, there are all sorts of arguments — from Fed independence concerns to a five-years-and-counting inflation overshoot to ~seven-handle nominal growth to what budget hawks deride as America’s parlous fiscal trajectory — for why longer-end Treasury yields should remain elevated. It’d be nice if there were at least a few arguments for why they should be lower. Less supply would be one such argument.
But Treasury’s now competing for buyers with a duration deluge from the most creditworthy companies in the history of capitalism, which is doubly bad: It means taxpayers will get a worse deal as Treasury has to offer higher yields to compete and adding supply to an already saturated high-grade market could push up borrowing costs for the companies working to keep America ahead in the “existential” AI race.
This may end up being one of the more important macro-market discussions of 2027, so I want to stay up on it. On Tuesday I mentioned a note from Nomura’s Jon Cohn, who wrote that although he still doesn’t expect cuts to longer-end auctions, “the scenario has gone from a wingy tail to a real prospect [as] AI-related supply ultimately gives Treasury enough motivation (or cover)” to reduce coupon sizes.
The figure above uses figures from Cohn’s note. As you can see, nearly two-thirds of AI-related high-grade supply since 2025 is 10 years and out.
“To put [this] in context, mega-cap tech issuance this year is about 25% the size of net Treasury issuance (ex-bills) to private investors (i.e., excluding the Fed), up from 5% in 2025,” Cohn wrote, adding that some of the most commonly-cited estimates of AI-related corporate supply in 2026 are actually “quite restrained.” (If you “widen the scope of what qualifies as ‘AI-related’ issuance,” you can peg the supply tsunami at nearly half a trillion YTD.)
The bottom line, Cohn went on, is that AI and data center debt “leans quite long in maturity [and] the sheer amount of duration supply forced onto the market at the long-end” should be a concern for Bessent.
I wonder if the customers are the same for both long duration treasuries and hyperscaler bonds.
I will bet they aren’t. Some spread action coming up
I agree with this article and its logic almost completely. However, I think there is another logic too: CYCLES. A.I. generally agrees with this. The U.S. dominated. equity and rates cycle last peaked on JAN 1 2022. The next peak is in 3 years and 9 moths. It sounds pretty damn close.
Curious: when hyperscaler bonds have durations longer than the lifespan of the chips they are buying with the proceeds, is this really a good idea? Not the way I think of structuring loans for anything in my personal life, but I’m not C-level material in my professional life either.
This is a good comment. I don’t know if it’s quite a hall-of-famer, but it’s high-quality dry humor.
I’ve been wondering this myself for a while so I asked the beneficiary of the long bonds itself, AI, and it told me that while the chips will be outdated in 5 years the investments are primarily in the land, buildings, air conditioning systems, and electrical infrastructure. They plan to just swap out the chips in the servers every 5 years, but are looking at 30 year lifespan building infrastructure to justify the expense. Someday all the homeless people may rise up against the data centers, so like Flock cameras they need to install as many as they can now while ‘the getting is good’.
If that’s the case, it sure seems odd that we haven’t seen any air conditioning or electrical system manufacturers join the ranks of the trillion dollar market cap club.
Snark aside, I don’t doubt that some of these construction firms (and maybe even some of the workers) are seeing more money thrown at them than they’ve ever seen in their life, just not anywhere in the same ballpark as our beloved tech titans.
It’s been happening for a while. I have a relative who graduated high school in 2021 and went straight to work with an air conditioning company making $70K a year installing units in Texas data centers. He’s now a “Cooling Customer Engineer” with his high school degree making six figures.
The trouble with the logic about most of the money being spent on land, buildings, air conditioning etc, is that it is essentially worthless. Who could afford to by all this space if there are problems with AI growth in a few years.
Agree, of course. They’re ripping apart a 500-acre, century old farm in my neck of the midwest for a huge Oracle data center… not just who can afford it, but who’s going to want that real estate when the infrastructure is obsolete in a few years? It’s miles from nowhere, not the sort of “location, location, location” ideal.
I’ll hazard a guess: DHS, ICE, GEO Group, CoreCivic, and whatever the Trump and Kushner bros got their dirty fingers in.
The hardware is the majority of the cost.
Consumers will borrow over 7-8 years for a new car, and we know where that goes.
Should include state and municipal bond issuance as well in the analysis. Over half a trillion each year past 2 years.
i would argue that the next 20 years most important macro story for the US is ,
“34% of US adults (OECD average: 25%) scored at or below Level 1 proficiency in numeracy. At Level 1, they can do basic maths with whole numbers and find single pieces of information in tables or charts, but struggle with tasks needing multiple steps. Those below Level 1 can add and subtract small numbers.” OECD 2023 study
Roughly a third of Americans has only very basic reading and counting skills, a sharp deterioration from 10 years ago and the trend is down further.. but does explain Trump and Kegsbreath