Everything About AI
NextFin News - Wall Street has a one-line message for the technology companies racing to build the infrastructure of artificial intelligence: for pity's sake, please slow down. The warning comes from the bond market itself, where a historic borrowing spree to fund AI data centers, chips, and servers has collided with investors' limits, pushed long-term interest rates to their highest levels in more than two decades, and turned the cleanest balance sheets in corporate America into a source of market instability.
From next to nothing in 2024, tech-sector borrowing reached roughly $500 billion in the nine months after January, as Google, Meta, Amazon, Microsoft, and others raised debt hand over fist to finance the chips, servers, and data centers that power AI. The scale is now large enough to matter to the Treasury market — the anchor of global finance — and the transmission is running in both directions: rising yields are rattling tech stocks, and tech borrowing is helping keep yields elevated.
The central tension is this: the AI build-out is a structural capital supercycle, but the bond market absorbs new supply in cyclical waves. When the two collide, the price of capital rises for everyone — not just for the borrowers, but for homeowners with mortgages, for the U.S. Treasury funding its deficit, and for every company that needs to refinance. The question investors are now asking is whether the market has been pricing AI as a pure earnings story when it is, increasingly, a rates story.
The Borrowing Wave, By the Numbers
The size of the shift is difficult to overstate. Investment-grade-rated global bond issuance from the hyperscalers totaled $108 billion in 2025, roughly 26% of their capital spending, according to Goldman Sachs. Through the first half of 2026 those companies had already issued $194 billion of investment-grade debt, on track for about $250 billion for the year — roughly a third of capex. In 2027, Goldman expects the debt-funded share to climb further, with $400 billion of expected global investment-grade issuance from the hyperscalers against an estimated $1.14 trillion of capital spending, a roughly 35% share. Total hyperscaler capex, by comparison, is expected to reach $750 billion by the end of 2026 and approach $1.2 trillion in 2027.
Those figures cover only the hyperscalers. AI-linked debt issuance by companies outside that group — software, semiconductor, and data-center financing — has already totaled roughly $412 billion in 2026. Morgan Stanley forecasts global AI-related debt issuance will more than double to nearly $570 billion for the year, after standing at about $236 billion as of May 31, roughly four times the same period a year earlier. Meta, Oracle, and other technology companies raised $250 billion in global debt markets this year alone, at a scale that would have been hard to imagine only a few years ago.
The borrowing is not abstract. Amazon is projecting $200 billion in capital expenditures for the year, Alphabet is targeting $175 billion to $185 billion, Meta has guided to $115 billion to $135 billion, and Microsoft is tracking toward $190 billion for the calendar year. At the low end of that range, the four hyperscalers would spend about $635 billion in 2026 — a roughly 67% spike from their $381 billion in 2025. Meta's total debt, for one concrete example, has climbed from about $36 billion in 2023 to $84 billion at the end of the first quarter. AI-related companies have issued about $140 billion in investment-grade bonds year to date, accounting for 49% of total investment-grade issuance, with another roughly $21 billion in high-yield bonds, or 38% of that market, according to analysis from the Kobeissi Letter.
For most of the early build-out, the market was happy to fund it. That window has closed. Over a several-week stretch, the investment-grade corporate bond market struggled to absorb a combined $75 billion of issuance from Nvidia, SpaceX, and Amazon. Nvidia and SpaceX were able to borrow at reasonably low interest rates, but their newly issued bonds quickly slumped in the secondary market, disappointing investors who often like to flip such deals. Amazon, meanwhile, had to pay unusually steep rates by its own standards to complete its debt sale. Lenders also demanded higher yields on bonds for a new data-center project in El Paso, Texas, leased by Meta, compared with a similar project a year earlier. The price to finance the AI data-center boom is rising, and the borrowers can feel it.
Why a Corporate Borrowing Spree Moves the Treasury Market
The first-order story is simple supply and demand: more bonds for sale means lower prices and higher yields. But the mechanism that makes AI borrowing a market-wide problem is more specific than that. It runs through duration — the bond market's measure of interest-rate exposure — and through the term premium, the extra compensation investors demand for holding long-term risk.
Oracle, once a relatively minor issuer of long-term debt, has become one of the largest suppliers of duration risk in the investment-grade universe, according to Srini Ramaswamy, a senior financial economist at the Federal Reserve Bank of Dallas. Ramaswamy estimates that AI-related issuance now accounts for roughly 15% of the duration supplied by total Treasury issuance. That is not a rounding error. The Treasury, by comparison, sold $540 billion in 10-year notes in the past year.
The effect is amplified by financial engineering that does not show up in headline issuance statistics. Because many institutional investors face internal limits on how much debt they can hold from any one company — and on duration — technology borrowers often issue shorter-dated or floating-rate debt and then use interest-rate swaps to convert that exposure into longer-term fixed-rate funding. Ramaswamy estimates this swap activity alone accounted for about $50 billion in 10-year-equivalent supply in the fourth quarter, and that the figure is probably higher now. A large share of AI financing is therefore long-duration risk in disguise, landing on the same part of the curve that the Treasury, pension funds, and insurers are all fighting over.
The timing matters. The recent surge in AI-related infrastructure investment was partly behind the May rout that pushed 30-year Treasury yields to their highest level since 2007, alongside inflation fears and shifting expectations for Federal Reserve policy. Yields have come down from their May highs but remain above where they began the year, reflecting in part the wave of bond issuance. On October 7, the 10-year Treasury yield climbed above 5.35%, its highest level since 2002, while the 30-year bond touched 5.70%. In a recent week the benchmark 10-year yield soared 17.1 basis points to end at 5.167%, levels not seen since June 2007, while the 30-year note posted its biggest weekly climb since May.
There is a second-order channel that most investors are not pricing cleanly: the feedback loop between higher yields and government borrowing. Higher yields raise the Treasury's debt-service costs; higher debt-service costs require more issuance; more issuance can push yields higher still. The AI build-out is creating a new source of private supply at the exact moment the market is already being asked to absorb unusually heavy government borrowing. A private capital supercycle and a public deficit problem are competing for the same pool of long-duration capital, and the price of that capital is going up.
The Market Is Splitting in Two
The clearest evidence that this is indigestion rather than a credit-quality crisis is how the market has bifurcated. The market for highly rated corporate credit has split in two: bonds issued by AI-related firms are being met with caution, while those sold by so-called traditional issuers such as financial and industrial firms are prompting spirited bidding. The sheer volume and unpredictability of borrowing needed to finance data centers, chips, and AI infrastructure are prompting bond buyers to demand generous concessions and to rethink portfolio concentration limits.
Bonds issued by major AI spenders such as Meta Platforms and Alphabet have consistently traded wider than similarly rated peers, even though the firms generate lots of cash and have strong balance sheets, said Lon Erickson, a portfolio manager at Thornburg Investment Management. BlackRock's Brownback said the widening in AI-related spreads reflects straightforward supply-and-demand dynamics rather than growing concerns about credit quality, and that the trade-off remains attractive for both issuers and investors.
That distinction is crucial. If investors were worried that Meta or Microsoft could not repay, spreads would blow out on credit risk and the equity market would be pricing default. What is actually happening is that investors are being asked to hold too much duration from too few names at once. It is a market-structure problem, not a solvency problem — which is why it can resolve quickly in either direction.
The equity market has already started to reprice the cost of capital. An aggressive spending forecast from Alphabet's Google sparked a selloff in tech stocks and helped push down prices of bonds from companies like Microsoft and Amazon. On another session driven by a worldwide bond selloff, the S&P 500 lost 53.30 points, or 0.69%, to 7,691.76, while the Nasdaq Composite declined 355.20 points, or 1.33%, to 26,289.71, as long-term yields hit their highest levels in years. Fears around the AI trade have kept investors leery of technology stocks even during sessions when other sectors rotated higher, with AI-focused names such as CoreWeave swinging as much as 9.5% intraday on resurfacing investment concerns. "There's nothing that can crack a momentum rally quite like interest rates moving higher," said Tony Welch, chief investment officer at SignatureFD.
"Will AI kill us in 10 years… buddy, the bond market is going to kill us next week," said Clocktower Group market strategist Eric Wallerstein, as the 10-year and 30-year Treasury yields climbed to their highest levels since 2002.
Cyclical Indigestion or Structural Regime Shift?
This is the judgment the market has to get right, because it determines whether the pain is temporary or permanent. The answer is both — and they operate on different horizons.
The cyclical leg is real and already visible. Goldman Sachs' head of credit strategy, Amanda Lynam, said the market is probably past the peak of supply in 2026 for the AI-related theme "because there's been a bit of indigestion and fatigue," and that 2027 will be the accelerator. That is a cyclical pattern: a supply wave outruns near-term absorption, yields gap higher, issuance pauses, and the market digests. History offers comparable episodes — the 2013 taper tantrum, the 2021 meme-credit frenzy, the 2007 leveraged-loan peak — in each of which a crowded trade hit capacity limits and then mean-reverted once the flow normalized. Three features mark this as cyclical: it is driven by the pacing of supply rather than a change in fundamentals; the borrowers are investment-grade with strong cash generation; and the widening is concentrated in AI names while similarly rated peers trade tight.
But beneath the cyclical indigestion sits a structural shift that will not revert on its own. Three pieces of evidence support the structural read. First, the capital intensity of the AI build-out is a regime change in an industry that spent the 2010s and early 2020s as a light-asset, cash-rich sector. The hyperscalers are on track to spend upward of $650 billion in 2026, a roughly 67% jump from 2025, and the debt-funded share is rising from 26% to 35% and beyond. Second, the financing is creating a permanent new source of duration supply — Oracle did not exist as a major long-duration issuer five years ago, and now it is one of the largest in the investment-grade universe. Third, the build-out is large enough to show up in macro aggregates: by most estimates, AI investment will soon account for 3% to 5% of U.S. GDP.
The structural signal is visible in the composition of the yield move. In a typical inflation scare, long-term inflation expectations rise sharply. This time, real yields have risen while inflation expectations have stayed relatively contained — a pattern that Jonathan Hill, head of U.S. inflation market strategy at Barclays, says is consistent with an AI-driven investment boom pushing up capital demand now but potentially improving productivity and easing inflation over time. A cyclical inflation impulse would lift breakevens; a structural capital-demand impulse lifts real yields. The data are pointing to the latter. The 10-year Treasury real yield recently stood above 2%, well over a percentage point higher than its 10-year average near 0.9%.
"The issuance numbers on this are dramatic," Hill said. "It's running in the background and it truly is the big story."
The Counter-Thesis, Taken Seriously
The strongest case against this narrative is the one its own proponents make: none of this means AI borrowing is the dominant force in Treasuries. Federal Reserve policy, inflation data, fiscal deficits, and global demand for safe assets still matter more. Yields could come down as fast as they have climbed if the Fed turns more dovish, if oil retreats, or if a slowdown in growth sends investors back into bonds. On that view, blaming AI borrowing for higher yields is a classic post-hoc story — an attempt to explain a multi-causal move with a single fashionable cause.
That counter-thesis has force. The 10-year yield is a global price set by many margins: foreign official demand, pension rebalancing, inflation expectations, the federal funds path, and term premium all move it. AI-related supply at 15% of Treasury-supplied duration is large, but it is not the largest margin. And the May rout that pushed the 30-year to its highest since 2007 had inflation fears and Fed repricing as co-drivers, not AI alone.
But the counter-thesis proves too much if it is used to dismiss the channel entirely. The question is not whether AI borrowing is the only driver — it clearly is not — but whether it is a new and persistent marginal driver at the exact point in the curve where the market is least elastic. Long-duration supply is the answer to that: it is concentrated at the belly and long end, it is incremental to government supply rather than substituting for it, and it is tied to multi-year capital commitments that cannot be switched off when yields move. Even a secondary factor, applied at the inelastic part of the curve, can move prices meaningfully.
The falsifying signal is specific and observable. If the 10-year Treasury real yield falls back and holds below 2% while AI-related issuance stays above $500 billion a year, the supply-driven thesis is wrong — real yields would be telling us that capital demand, not supply, is the binding constraint, or that another factor has overwhelmed it. A second falsifier: if hyperscaler capital-expenditure guidance flattens or declines while the debt-funded share keeps rising, that would confirm the cyclical read and undercut the structural one. Until one of those prints, the structural interpretation carries more weight than the "just a cyclical flow" dismissal.
What Comes Next: Three Horizons
Short term (weeks to a few months): volatility and indigestion. The market is in the digestion phase Lynam described. Expect episodic concessions on new deals, wider spreads on AI-linked names relative to peers, and equity sensitivity to any 10-basis-point move in the 10-year yield. Issuers will respond by stretching maturities, tapping private credit, or delaying deals — the normal adjustments of a crowded primary market. The beneficiaries here are relative-value credit investors who can pick between AI and non-AI issuers of the same rating; the exposed are momentum-long tech positions that assume a stable discount rate.
Medium term (6 to 18 months): the cost of capital becomes the selection mechanism. As the debt-funded share of capex climbs toward 35%, the interest cost of AI infrastructure moves from immaterial to material on income statements. That will separate the hyperscalers with pricing power — who can pass higher capital costs through to cloud and AI service customers — from the rest. It also raises the hurdle rate for AI projects: a cash flow five to ten years out is worth materially less at a 5.3% discount rate than at 3.5%. The medium-term winners are the companies whose AI revenue arrives soon enough to outrun the rising discount rate; the losers are the promises priced on cash flows that only arrive in the 2030s.
Long term (years): a re-rated term premium, if the build-out persists. If AI investment sustains 3% to 5% of GDP and keeps adding long-duration supply alongside heavy government issuance, the term premium that was compressed to near-zero in the 2010s may not fully return to those lows. That is a regime change for every long-duration asset, not just tech: growth stocks, real estate, and the Treasury curve itself all price off it. The asymmetry is that the market has spent the past decade underpricing term-premium risk; the AI build-out is one of the forces that could force a repricing.
The base case is that yields remain elevated and volatile rather than grinding steadily higher: the cyclical indigestion caps the near-term supply pace, but the structural capital demand puts a floor under long rates. The upside case for bonds is a growth scare that sends investors back into Treasuries and pushes the 10-year back toward 4%. The downside case is a continuation of the May-style rout if capital-expenditure guides rise again while inflation expectations re-accelerate — a scenario in which the private and public supply waves coincide.
What to watch, concretely: the 10-year real yield and the steepness of the 2s10s and 10s30s curves; quarterly hyperscaler capital-expenditure guidance and the debt-funded share; the spread between AI-linked issuers and similarly rated non-AI peers; and the Treasury's quarterly refunding announcements, which will show whether government supply is compounding or offsetting the private wave.
The sharpest takeaway is also the most uncomfortable for AI bulls: the bond market is not arguing about whether artificial intelligence will work. It is arguing about the price of the capital required to build it — and after a decade of cheap money, that price has gone up. The companies that treated the AI build-out as an engineering problem now have to solve it as a financing problem. The ones that get the financing right will own the next cycle; the ones that assumed the bond market would stay asleep will find out how quickly it can wake up.
Explore more exclusive insights at nextfin.ai.
