Increasingly, it seems like at least some of the blame for rising US Treasury yields is being laid at the door of the hyperscalers. Kevin Warsh himself listed competition for capital as one of three reasons for surging yields during his recent press conference, specifically calling out “the so-called hyperscalers [ . . . ] out in the market raising funding”.
And why not? 10-year Treasury yields have risen by ca 90bps 100bps over the past year — and just look at the hyperscaler issuance numbers! Lumping Amazon, Alphabet, Meta, Oracle, SpaceX and Nvidia together, we count well north of $200bn of spanky new investment-grade public bonds in 2026. And we haven’t even hit Q4!
If whoever’s buying this stuff wasn’t being tapped for funds, maybe they coulda shoulda woulda been buying US Treasuries instead, keeping bond yields low?
It all feels pretty intuitive. But we wonder whether people are getting carried away.
The case for the prosecution
Way back in February, Hugo De Vere, Srini Ramaswamy and Seth Searls of the Dallas Fed set out how the sheer weight of supply of AI debt could push yields higher. Wall Street estimates at the start of the year were, they tell us, for $300bn of AI-related investment-grade issuance, or “as much as $360bn in 10-year equivalents over the course of 2026, or about an eighth of the duration supply from US Treasury issuance”.
The last part of that — the bit about 10-year equivalents and duration supply from the US Treasury market — is the important bit. But it’s written in bond nerd, so let’s unpick it.
Translating bond volumes into 10-year equivalents is basically a tiny calculation that bond types do to allow them to compare the volume of interest rate risk that bonds of different maturities represent on a semi-apples-to-apples basis.
Because while a two-year bond ties up the same amount of balance sheet as a 30-year bond (eg, $100mn), it will carry a lot less interest rate risk (eg, ca 1.9yrs of modified duration, versus ca 15.5yrs). It’s the combination of balance sheet and interest rate risk that matters to anyone managing a portfolio, and for this reason it matters to bond issuers too. So bondies multiply each instrument’s nominal value by its duration and divide the result by the duration of a 10-year US Treasury to turn it into a 10-year equivalent duration supply volume.
To see how nominal hyperscaler debt issuance translates into 10-year equivalents, we’ve popped a drop-down toggle into our pretty column chart up top. And the jump is impressive. It’s not the $360bn of 10-year equivalents that the Dallas Fed tells us Wall Street was looking for at the start of the year, but we’re only showing hyperscalers and Nvidia rather than whatever expansive definition The Street might’ve had in mind, and 2026 is not yet done.
The authors also outlined two other routes through which AI financing could push up yields: synthetic duration supply, and potential shrinking financial issuance.
The first of these basically flows from private credit investors wanting to receive floating-rate coupons while the data-centre manager wants to pay fixed on the same loan. So a bunch of fixed-payer swaps enter the market, weighing on the supply of duration. Can we quantify this? No. So let’s park this one.
The second route is wildly fascinating, involving complex chains of synthetic transactions across multiple counterparties that sometimes — but not always — happen each time a financial company issues a bond. We were all set to make a massive scrolly with boxes, arrows (and prime our lawyers to expect a legal letter from Matt Groening), but given that issuance coming out of banking has increased rather than fallen this year, we’ll park this one too.
Still, the first route by which AI financing pushes Treasury yields higher is tractable and, as JPM’s Nikos Panigirtziglou put it in last week’s Flows and Liquidity, it has become “a regular feature in our conversations”.
It was one of three factors weighing on the long end of the US Treasury market, according to a Barclays research note in mid-August, along with the budget deficit outlook, and the changing Treasury buyer base.
This was followed by Yardeni Research writing that “the AI revolution is producing a classic crowding-out effect, causing Treasury yields to rise”, and a BNY diagnosis that “price-insensitive borrowers”, issuing debt with longer-duration tenors to fund AI capex, were affecting investor demand for long-end US Treasuries.
Then came this jaw-dropping chart from JPMorgan Asset Management’s Michael Cembalest that has been making the rounds:
The line uses that same bond geek metric — 10-year US Treasury equivalents. But it does a further transformation of the data, showing it as a percentage of US Treasury issuance. And doing so shows the kind of hockey stick that leaves little room for doubt as to the magnitude of the crowding out.
Can this be right? No.
We contacted JPMorgan Asset Management, whose people told us they’d included off-balance-sheet lease commitments in their number. And, importantly, the denominator for their chart wasn’t actually US Treasury bond issuance, but US Treasury long-duration borrowing.
It does strike us as frankly weird to count up all the scraps of duration from two-year, five-year, 10-year, etc., hyperscaler new issues, translate these into 10-year US Treasury equivalents, and then compare the result to the duration supplied into the over-15-year part of the US Treasury curve. Particularly because, as MainFT reported over the weekend, the big shift by the US Treasury is to shorten the average tenor of government debt issuance, and ramp up issuance of T-bills. So the overall supply of duration from long-dated Treasuries is tracking smaller:
But this is what they’ve done, and if we’d bothered to read their text closely rather than gawk at the crazy chart, we probably could’ve worked this out.
However, compared to overall US Treasury duration supply, hyperscaler supply is actually hitting the one-eighth levels set out in that original Dallas Fed note. So the claim that hyperscalers are crowding out the US Treasury maybe looks legit, if not to the degree implied by a misreading of a niche-viral chart.
But shouldn’t we look at the rest of the corporate market too, rather than just the hyperscalers?
The rest of the market
We know that hyperscalers have increased issuance, and now account for around an eighth of US Treasury duration supply at current run rates. So presumably we can pull up a chart to show the total duration-weighted volume of investment-grade corporate bonds is at least mini-hockey-sticking, displacing Treasuries?
Oh.
While the blue line squiggles gently higher, and while investment-grade corporate bonds are tracking around 40 per cent of the combined supply of duration coming from Treasuries and IG corporates to the market this year (rather than the more usual third), this year’s squiggling looks more like the squiggling of yesteryear than we’d expect.
Partly, this looks like it’s because, despite all the gross bond issuance, the growth of the corporate bond market continues to be outpaced by the growth of the US Treasury bond market.
And partly, it’s because, despite all the hyperscaler bond supply, the average duration of the corporate bond market, like the US Treasury market, is actually falling at the margin.
Every time the earth orbits the sun, the average number of years to maturity for all previously issued bonds falls by exactly one, reducing the average duration of the market. And so, for the average maturity of the market as a whole to stay relatively stable, a bunch more longer-dated bonds need to be issued, supplying the market with nice fresh duration.
Comparing the supply of duration year-to-date with where we got to this time last year, it looks like hyperscalers account for about half the change in gross supply:
Though in a complicating wrinkle, about a quarter of the change comes from banks and financials, which the Dallas Fed says should boost demand for Treasuries rather than reduce it.
So yes, the hyperscalers are issuing a lot of bonds. And yes, any bond issuer who issues any single bond could arguably be said to be crowding out the US Treasury — just as the Dallas Fed team and other economists note.
But the pace at which the hyperscalers are issuing bonds appears insufficient either to move the market’s average net duration higher, or to enable the public corporate bond market to keep up with the growth of the US Treasury market on either a nominal or duration-weighted basis.
So it looks to us as though anyone seeking to explain the big move higher in US Treasury yields will need to lean much more heavily on Warsh’s other two reasons — economic strength and geopolitics.
They might also want to consider one he must’ve forgotten in the heat of the moment: the US government’s gaping fiscal incontinence.
Further reading:
— AI debt vs Treasuries (Unhedged)
— US government debt hits $40tn as borrowing rises at historic rate (MainFT)