AI Pays Its Spread. Everyone Else Pays the Yield.

JunkBondInvestor and Luke Kawa look at the same 10-year yield and name different culprits. In Credit Weekly: Tightening by Proxy on October 12, 2026, JunkBondInvestor argued that persistent AI spending could itself help keep rates elevated, and that the long end probably gets relief only when AI investment slows and one of the largest sources of new corporate debt leaves the market. Eleven days earlier, on Joe Weisenthal's Odd Lots, Kawa put the rise in Treasury yields down mainly to a global growth impulse on the industrial side, lifting inflation and the chance that central banks respond. DeepStack judged the two accounts in direct contradiction on what moves the yield.

The yield does not choose between them. On monthly averages, the 10-year Treasury went from 4.12% in September 2025 to 4.99% in September 2026, by FRED's monthly series, while the high-quality corporate curve's 10-year spread over it went from 0.80 to 0.87 percentage point, by DeepStack's arithmetic. Freddie Mac's 30-year mortgage rate averaged 6.86% in September 2026, 51 basis points above a year earlier, by its weekly survey. The daily 10-year closed at 4.98% on October 29, 2026, by FRED's daily series, after a closing high of 5.31% on October 5 and up from a 2026 low of 3.97% on February 27.

Part twenty-four asks whether AI's borrowing is making credit dearer for the rest of the economy, and who outside AI pays for it. Part seven concluded on September 27 that supply sets the long end. Part ten set the hyperscalers' bonds against the savings that price stocks, and Part fifteen told readers to count AI borrowing as a rates input. DeepStack read 462 sources by 130 voices and verified 269 agreements and 174 disagreements, 58 of them direct contradictions and 3 refutations.

AI borrowing lifts the yield

Torsten Slok, Apollo's chief economist, made the supply case in his Daily Spark of August 26, listing the fiscal outlook, a Fed weighing a hike and hyperscaler issuance crowding out demand for Treasuries as forces pointing the same way. Orbis Investments wrote in its third-quarter 2026 letter that yields rose on a higher expected path for short rates, driven by AI's capital demands and the government's deficits. Harris Associates Oakmark wrote in its letter for the same quarter that heavy hyperscaler bond sales are pushing yields up.

The case rests on a borrower that does not flinch. Slok wrote on August 20 that AI spending had become largely indifferent to higher rates. Dylan Patel, on Dwarkesh Patel's podcast on August 25, said he saw no reason Meta, having recently borrowed at 5% to 6%, would not pay 8% for compute, and offered it as an intuition rather than a calculation.

JunkBondInvestor's version names who pays. Its central argument on October 12 is that restraint does not fall evenly: AI-infrastructure borrowers keep issuing almost regardless of rates, so the burden lands on ordinary, price-sensitive borrowers while AI takes a larger share of credit. On October 18 it added that an AI slowdown would remove the one large flexible borrower and leave the inflexible one, the government, so real yields would not fall back as they would in an ordinary growth scare.

Growth, not AI, sets the rate

Michael Howell of Capital Wars is the counterweight, and he does not deny AI a part. On August 23 he wrote that long yields are rising because the economy is very strong, its nominal growth lifted by a fiscal deficit of about 6% of GDP and a large wave of AI capital spending. In the same post he called term premia and the curve's slope primarily liquidity phenomena. On his reading AI reaches the yield as growth rather than as supply crowding the auction, and DeepStack judged his liquidity account and Slok's crowding-out note of three days later in direct contradiction.

Others on this side say the economy has stopped minding rates. Matthew C. Klein of The Overshoot wrote on September 1 that consumer spending and business investment have become far less sensitive to credit than in the past. Andrew Sheets of Morgan Stanley said on Thoughts on the Market on October 9 that growth in 2026 may be less rate-sensitive than expected, with deficits and AI spending two possible reasons, and that today's rates do not look obviously wrong.

Slok supplied this side's best fact himself, writing on September 2 that the New York Fed's 10-year term premium had moved sideways for twelve months. The Fed Board's Kim-Wright estimate, a different model, disagrees: 0.49 point on September 2, 2025, 0.90 a year later and 1.07 on October 9, 2026, by FRED's THREEFYTP10.

The strongest evidence for this side is research. Jens Christensen of the San Francisco Fed and Glenn Rudebusch of Brookings found, in a Hutchins Center working paper of August 2026, that AI news does not explain the rise in the natural rate of interest since 2020: four estimates of it fell 23 to 35 basis points in the windows around AI news, while fiscal news accounted for about 20% of the increase. Isaiah Andrews and Maryam Farboodi of MIT found, in a National Bureau of Economic Research paper issued in September 2025 and revised in July 2026, that long Treasury and inflation-protected yields fall around major AI model releases and stay lower for weeks. Lotfi Karoui of PIMCO wrote on September 28 that the evidence of AI bonds directly crowding out Treasuries had looked weak across yields, term premia and swap spreads. The two papers measure what news about AI does to expected real rates, not what AI's bond sales do to the supply of duration, and Christensen and Rudebusch cite work finding that more debt supply moves term premiums more than the natural rate. DeepStack's Call rests on that supply channel and on the September minutes' mention of AI borrowing, and these studies are why its conviction stops at three of five.

Four doors

DeepStack's reading of public series on October 30, 2026 separates what each kind of borrower pays. The differences are its own arithmetic, on the same monthly window unless a date says otherwise. The Treasury door is open. The 10-year averaged 4.99% in September 2026, 87 basis points above September 2025, and closed at 4.98% on October 29, 33 basis points under its October 5 high; the Kim-Wright premium averaged 0.46 point in September 2025, 0.55 in February 2026 and 0.96 in September 2026, and 1.03 over October's readings through October 23, the last posted, which ended at 0.94. The premium fell through those weeks while the hyperscalers kept selling bonds, and the 10-year fell with it; a month does not settle the supply half, which the March 2027 reading tests.

The corporate door moved 7 basis points at ten years. The Treasury's high-quality corporate spot rate at 10 years (HQMCB10YR) sat 0.80 point above the 10-year Treasury (GS10) in September 2025 and 0.87 in September 2026. The curve includes the hyperscalers' own bonds, so it cannot show on its own what other issuers pay, and the Fed's excess bond premium, the part of spreads that default risk does not explain, fell from -0.29 to -0.39 over the same months, a sign that risk appetite for corporate borrowers as a group improved rather than worsened. The long end widened. At 30 years the same curve (HQMCB30YR) over the 30-year Treasury (GS30) went from 1.09 points in September 2025 to 1.43 in August 2026 and 1.31 in September. Karoui reported on September 28 that the hyperscalers' long-dated spread curves had steepened while those of other nonfinancial issuers flattened, so part of that widening may be AI's own long bonds.

The mortgage door widened again in October, from a 2026 trough. Over the year to September 2026 Freddie Mac's mortgage rate rose 51 basis points on monthly averages, less than the 10-year's 87, and its spread over the same day's 10-year narrowed from 2.24 points to 1.90. The weekly spread of Freddie Mac's survey rate over the same day's 10-year then averaged 2.23 over October's five Thursdays, from 2.04 on October 1 to 2.36 on October 22 and 2.26 on October 29, against monthly averages between 1.89 and 2.05 from January to September; it had averaged 2.32 in August 2025. The bank door all but stopped tightening in July and began tightening again for small firms in October. In the Fed's October 2026 loan officer survey, published November 2, a net 5.1% of banks tightened business-loan standards for small firms (DRTSCIS) and 1.7% for large and middle-market firms (DRTSCILM), after 1.8% and none in the July survey, 6.6% and 8.1% in April 2026 and 8.9% and 5.3% in January. The Board said standards tightened for small firms and were "basically unchanged" for large and middle-market firms.

The minutes of the Fed's September 15-16 meeting, published October 7, describe the same doors. The manager said part of the rise in Treasury yields reflected a higher expected path of policy and strong data; market commentary also pointed to competition for capital from heavy AI debt issuance, alongside geopolitics and uncertainty over the Treasury's buybacks, as contributing to higher term premiums and Treasury yields. A few participants listed expectations for AI-related borrowing among potential factors. Hyperscaler spreads "remained wide" given the issuance's volume and long duration. In the staff's review, those spreads "decreased slightly over the period after widening notably," corporate bond issuance "remained robust, even beyond the recent notable activity of AI hyperscalers," and conditions were "generally accommodative for larger businesses and municipalities but were somewhat restrictive for residential mortgage borrowers and small businesses." Corporate yields rose "mostly in line with Treasury yields," and mortgage rates "a bit more than 10-year Treasury yields."

On October 28 the Fed held its target range at 3.75% to 4%, with Lisa Cook and Beth Hammack voting for a quarter-point rise, and its statement called capital investment robust; the minutes of that meeting are due November 18. The next day Microsoft priced $25.0 billion of notes in ten tranches maturing from 2028 to 2066, its first sale of notes under a 424B2 prospectus supplement since January 2017, by its pricing term sheet on EDGAR. For an issuer both agencies rate AAA, the ten-year notes priced 65 basis points over the Treasury and the 2066 notes 105.

How large a competitor AI is depends on the denominator. The Dallas Fed estimated on February 10, 2026 that AI-related high-grade issuance could add as much as $360 billion of 10-year equivalents in 2026, about an eighth of the Treasury's total duration supply. Michael Cembalest of J.P. Morgan Asset & Wealth Management counted only the long end in his Eye on the Market of September 1, 2026: the long-duration part of 2026 issuance by the five hyperscalers and Nvidia, about $160 billion in 10-year equivalents, is 35% of the Treasury's new long-duration borrowing. On DeepStack's reading, either ratio makes AI a large competitor for savings, though the government still supplies about eight times as much duration on the Dallas Fed's measure, and about three times as much at the long end on Cembalest's count.

That is what paying its own spread means. Oakmark told clients in that third-quarter letter that it holds hyperscaler bonds for their wider spreads and lower leverage than the rest of investment grade. Harley Bassman, on Erik Townsend's MacroVoices on September 17, said the hyperscalers have the cash flow for their coupons, so a lab's failure would be an equity problem rather than a bond problem. Michael Burry of Cassandra Unchained cut against that comfort on September 24 with Amazon's numbers: by its own filings, Amazon raised $81.9 billion of long-term debt in the twelve months to June 2026, against $746 million a year earlier, while its free cash flow over those months was an outflow of $7.6 billion. DeepStack judged his figures a refutation of Bassman's. Ayesha Tariq, a guest on Michael Spencer's AI Supremacy on October 12, added that higher rates can push off-balance-sheet guarantees back on, citing S&P's plan to count about $17 billion of leases Alphabet backed as its debt once they begin.

DeepStack's reading: the AI issuer pays a premium for its size and duration, and the high-grade market absorbs it while charging other issuers little more at ten years, 7 basis points in a year on DeepStack's approximation of an aggregate that includes the hyperscalers. Karoui's count agrees: AI-related borrowers supplied nearly a quarter of nonfinancial dollar high-grade issuance in 2026 through September 23, and spreads for other issuers had not widened meaningfully while the hyperscalers' had. The cost leaks out where AI and everyone else draw on one pool, the Treasury yield, and it lands hardest on borrowers who cannot refinance around it.

The weaker borrowers pay

Housing is where the hike works the old way, JunkBondInvestor wrote in Credit Weekly on September 27: mortgages through 7% for the first time since early 2025, and refinance applications 62% below a year earlier. Slok named the pattern on October 9: a "Dutch disease" in which a rate-insensitive AI boom pulls capital, power and labor away from housing and autos.

Sheets, on the other side of the yield argument, agreed on October 9. The rate that balances a spending boom is not the rate that keeps a weaker, more indebted company whole, he said, and in 2005, when a large US deficit, Chinese capital spending and US housing wealth pushed rates up, the economy held for years while weaker credit suffered first. Slok had written on September 25 that transmission is running mainly through the bottom of the credit stack, and Quanta 72 wrote in its brief of October 18 that the stress sits in CCC bonds rather than in BBB. The ICE BofA indexes on FRED (BAMLH0A3HYC) show it: from September 18 to October 16, 2026, the CCC option-adjusted spread widened 214 basis points, to 12.97 points, while the BBB spread widened 12 and the AAA spread 2.

Michael W. Green of Yes, I Give a Fig gave the household version on August 30. Where the problem is physical scarcity, he wrote, restraint mostly leaves households less able to pay for what is scarce without making it any less scarce. Klein reads the same redistribution as a feature: amid rearmament and new energy building, higher rates can discourage lower-priority consumer and housing activity.

What would settle it

DeepStack will read five public pages on their dates. The first is the minutes of the Fed's December 8-9 meeting, due on December 30 if the Board keeps its three-week lag: do they attribute any part of the level or rise of long Treasury yields or term premiums to AI-related borrowing, issuance or competition for capital? The second is the mortgage door, read on January 4, 2027: does Freddie Mac's 30-year rate average 2.10 points or more above the same day's 10-year Treasury over December 2026? The third is the Fed's January 2027 loan officer survey, due in early February: does a net 5% or more of banks tighten business-loan standards for small firms? The fourth is the supply premium, read on April 20, 2027: does the Kim-Wright 10-year term premium average 0.80 point or more over March 2027? If the rise in yields came from the expected path of policy rates, as Kawa's account has it, the premium can fall back while the yield holds; if it came from the supply of duration, it stays up. The fifth is the long end: is March 2027's 30-year high-quality corporate spot rate 1.30 points or more above the 30-year Treasury on monthly averages, read by April 30, 2027?

DeepStack's Call is that AI pays its spread and the rest of the economy pays the yield. The first half rests on hyperscaler spreads that remained wide in the September minutes, though they narrowed slightly over the period, beside robust issuance by others, on Oakmark holding hyperscaler bonds for their wider spreads and on Karoui's count of other issuers' spreads that did not widen; the aggregate high-grade 10-year spread, up 7 basis points in the year to September 2026, fits them, though it includes the hyperscalers. The second half rests on the Treasury yield, up 87 basis points over the same months, a Kim-Wright term premium that rose from 0.55 point in February to 0.96 in September, and the September minutes' mention of AI borrowing among the factors behind long yields. The minutes and the term premium are the Call's tests; the mortgage, bank and long-end readings are information. Attribution is the weak joint: the higher expected policy path the minutes also cite, large deficits and Howell's growth reading fit the same yields, Christensen and Rudebusch find AI news lowering expected real rates rather than raising them, and no public series isolates AI's share.

AI will pay the rate. The question is who cannot.

Three readings would change the Call: a Kim-Wright term premium averaging under 0.80 point in March 2027 while the hyperscalers keep selling bonds, meaning AI's supply no longer lifts the yield; a month through March 2027 in which the 10-year corporate spread tops 1.00 point while the Fed's excess bond premium rises above zero, meaning other issuers' spreads widen too; or December minutes that explain long yields without AI-related borrowing. The spread counts only with the premium, because the curve carries the hyperscalers' own bonds. The premium, measured across a broad panel of nonfinancial firms' bonds, last read above zero in June 2024, by the Board's data. AI will pay the rate. The question is who cannot.