AI capex is now big enough to move the yield it borrows at

Every deal I underwrite has a cell near the top for the ten-year Treasury, and for most of my career that cell was scenery. You set it, you moved on, the interesting work sat elsewhere. It is scenery no longer.

Look at what the people building AI have done in the bond market. Between 2020 and 2024 the five big hyperscalers issued roughly $35 billion of debt a year on average, per Vanguard (Tier 2)¹; year-to-date 2026 they are at about $132 billion, including a single $53 billion multi-tranche deal and a rare hundred-year bond (Tier 2)¹. cut² A rounding error became a market event in about eighteen months.

The supply is long-dated, and that is the whole problem

The supply this borrowing pushes onto the long end is already large.

Stat $360bn · the 10-year-equivalent duration supply this issuance adds, per the Dallas Fed · Tier 2

Here is how AI capex reaches a government curve, step by step:

1. Capex outran cash flow. Across Alphabet, Amazon, Meta, Microsoft and Oracle, investing cashflow rose from $95 billion in FY2020 to about $490 billion in the twelve months to May 2026, and incremental debt as a share of that capex went from 9% in FY2024 to 32%, per FactSet’s aggregation (Tier 2, secondary aggregation)³.
2. So they borrowed, and long. cut⁴
3. Long-dated corporate supply competes with the long end of the government curve. cut⁵ Reuters has described the buildout as feeding a wider surge in issuance that touches the Treasury market directly (Tier 2)⁶.
4. The auction results show the strain. On 13 August 2026 the US paid its highest 30-year auction cost since 2001, at 5.22% (Tier 2)⁷.

There is a second channel that has nothing to do with supply. If AI genuinely lifts productivity, the equilibrium real rate climbs with it – the Minneapolis Fed calls AI “a force that increases both natural rates and potentially price pressures” (Tier 3)⁸. A higher neutral rate and heavier long-end supply push the same way.

The honest counter-case is strong

I would be selling you something if I left it there. The most careful reading in front of me cuts against my thesis. cut⁹ cut¹⁰

The official narratives point elsewhere too. The IMF’s Fiscal Monitor puts the weight on fiscal deficits and sovereign issuance rather than corporate capex, estimating that a one-percentage-point rise in the US primary deficit lifts term premia by about 11 basis points (Tier 1)¹¹. Its Global Financial Stability Report attributes rising term premia chiefly to inflation-risk premia and global factors (Tier 1)¹². And in the cleanest natural experiment available, MIT Sloan researchers found Treasury, TIPS and corporate yields fell by more than 10 basis points on average in the weeks after fifteen major AI model releases in 2023–24 (Tier 2)¹³ – markets, around those dates, priced AI as if it would lower rates rather than lift them.

Even the funding mix is contested. A Fidelity-based aggregate view has US public companies funding AI largely from free cash flow, while FactSet’s narrower hyperscaler cohort shows the sharp turn to external debt – two conflicting readings of the same question (Tier 2)¹⁴. And no source in what I have isolates AI’s specific share of the yield move; the public evidence does not pin one down (Tier 5, open question).

So take the thesis as directional, not a magnitude. Issuance volume, tenor and the neutral-rate channel all lean one way. How much of the yield rise is AI, exactly, I cannot show, and neither can anyone I have read.

What it costs the person underwriting the deal

Here is why the direction is enough to matter. Whatever share of the base rate AI owns, the spread you pay sits on top of it – and those spreads have held or widened rather than absorbed the move.

The AI-specific paper carries its own premium on top. JPMorgan has technology credit spreads about 30 basis points wider year-to-date against just 2 for the broad investment-grade index, with hyperscaler spreads near 105 basis points duration-adjusted versus roughly 80 for the market (Tier 4)²⁰. Octus found eleven AI and data-centre SPV notes pricing at roughly double the option-adjusted spread their ratings implied (Tier 3)²¹ – the market charging extra for structure it does not yet trust.

The scale of what still needs financing tells you this is no passing squeeze. Brandywine Global sketches the funding gap as $800 billion of private credit, $200 billion of corporate debt, $150 billion of ABS and CMBS, and $350 billion from everywhere else (Tier 3)²²; Morgan Stanley frames roughly $200 billion a year against a $1.5 trillion data-centre financing gap (Tier 2)²³. That supply has to clear at some price, and it competes with everything else you are trying to fund.

The base case needs rebuilding

If AI capex is now helping set a floor under yields rather than riding beneath one, the discount-rate assumption baked into every model still anchored to 2010s cap rates is wrong at the scale of a regime shift. I can say that much with a straight face. What I cannot hand you is a tested way to act on it: there is no verified underwriting playbook in the public record with demonstrated outperformance for any single response (Tier 5, open question). Rebuilding the base case is work each desk has to do on its own numbers, then watch.

This is an observation from my own desk, not investment advice – no recommendation to buy, sell or hold anything, and no outcome guaranteed. The decision stays yours.