Dallas Fed: AI bond issuance this year could add $360bn of ten-year duration over 2026

The maturity choice follows from the asset. A data center's underlying economics do not vary with interest rates over a business cycle, the Dallas Fed note argues, so the borrower prefers fixed-rate liabilities to hold income volatility down ^[5]. Issuance from the hyperscalers and others in the buildout has been both large and long in duration ^[3].

Divide the two published figures and the shape of that book appears. $300 billion of face value converting into as much as $360 billion of ten-year equivalents puts about 20 percent more interest-rate risk in every dollar than a ten-year note carries ^[1]. The other comparison runs backwards. If $360 billion is about an eighth of what Treasury will supply, the note is implying roughly $2.9 trillion of ten-year equivalents from Treasury issuance across 2026 ^[2].

Direct issuance is one of three channels the note counts, alongside duration transferred through swap activity and displacement effects ^[12]. Private credit lends floating, because that is what investors in those markets want, so a borrower who needs fixed for the life of the asset pairs the loan with a pay-fixed swap ^[6]. Four quarters at the fourth quarter's $50 billion pace would be $200 billion of ten-year equivalents a year, 56 percent on top of the direct figure. The combined total lands near $560 billion, about a fifth of the implied Treasury number ^[3][4]. The $50 billion is anecdotal evidence from market participants, for one quarter ^[7]. I would not annualise it before the swap-spread evidence turns up.

That evidence has a usable sign. Treasury duration supply, all else equal, cheapens Treasuries against swaps; the pay-fixed swaps pressure swap yields higher relative to Treasury yields and leave Treasuries richer ^[9]. Private markets do not make those transactions observable on their own, and the note says the difference suggests a potential empirical approach to inferring their existence and scale ^[10][11].

Two ways this comes in smaller. The $500 billion to $600 billion spent since 2023 was mostly funded from retained earnings ^[2], and a hyperscaler paying out of cash flow is not issuing bonds. The other is the capex number itself. $3 trillion to $5 trillion over three to five years works out at $600 billion a year at the low end and about $1.7 trillion at the high end ^[5]. Against that, $300 billion of investment-grade issuance covers half the low case and under a fifth of the high one ^[6]. The rest has to come from cash flow, private credit or equity.