Last updated: August 1, 2026 · By Vishal Swami, Founder & Lead AI Reviewer, AISagely
The AI trade now runs on borrowed money because hyperscalers and "neoclouds" can't fund GPU buildouts out of cash flow alone, so they're issuing bonds and leaning on private credit at a pace Morgan Stanley pegs near $570 billion for 2026 alone. Lenders are starting to reprice that risk in real time, and the clearest evidence is a credit reversal at CoreWeave that played out in the last week of July.
Short answer: AI infrastructure spending is increasingly financed by debt, not equity, and Morgan Stanley expects roughly $570 billion of AI-related borrowing in 2026. Lenders are repricing that risk: CoreWeave's credit-default-swap spread hit levels implying a 50% five-year default probability on July 29, and a new loan tied to its Anthropic capacity deal got priced wider than first offered.
I review AI tools for a living, which means I also pay attention to who's paying for the compute behind them. When a headline says "lenders are repricing the AI trade," that's not abstract market noise to me — it's the infrastructure that ChatGPT, Claude, Cursor, and every other tool I test actually run on. So I went back to the primary numbers instead of taking the headline at face value, and the picture underneath is more specific, and more interesting, than "AI bubble."
What "Borrowed Money" Actually Means Here
Two years ago, AI capex was mostly funded from the giant cash piles of Microsoft, Google, Meta, and Amazon. That's changed. Morgan Stanley expects global AI-related debt issuance to reach close to $570 billion in 2026, with about $236 billion already priced by May 31 — roughly four times the pace of a year earlier, according to Forbes’ reporting on the bank’s numbers. Demand for that paper is also cooling: hyperscaler bond orders covered close to five times the amount on offer back in February, and by July that coverage had slipped below two times, a sign investors are asking for more yield to show up.
Private credit is doing a lot of the newer lending. Firms like Blackstone, Blue Owl Capital, Apollo, PIMCO, and BlackRock are projected to supply roughly $800 billion of data center financing over the next two years, often through off-balance-sheet joint ventures rather than direct corporate bonds. Meta's $27 billion Hyperion data center project in Louisiana is the clearest example: Blue Owl owns 80% of the joint venture and put in $7 billion of equity, while Meta holds the remaining 20% and leases the finished campus back, per Meta’s own announcement. That structure keeps most of the debt off Meta's balance sheet, which is efficient for Meta and harder for outsiders to underwrite, because the risk sits with a private lender instead of a rated public bond.
The Repricing Is Already Happening
CoreWeave is the cleanest test case, because its credit swung both directions inside eight months. In November 2025, CoreWeave-linked bonds were yielding around 10%, with credit-default-swap spreads that peaked near 8.81 percentage points that December. By June 2026, a new $1.59 billion junk bond priced at a 7% yield and drew five times its size in orders, and CDS spreads had fallen to about 4.52 points — a 49% drop from the December peak. For a few months, the market was treating CoreWeave's debt as closer to hyperscaler-grade than junk.
That rebound didn't hold. On July 29, CoreWeave shares fell 9% to $61.53, down 36% for the month, and Nebius, another GPU cloud provider, fell 10% the same day. CoreWeave's CDS spread topped roughly 855 basis points, a level implying a 50% probability of default within five years under standard pricing models, per Yahoo Finance’s coverage of the sell-off. The company's first-quarter 2026 interest expense of $536 million had doubled from the prior period, and free cash flow ran negative $4.71 billion for the quarter alone, against a business that's been cash-flow negative since 2022. Days earlier, CoreWeave had gone to lenders for a $2.6 billion delayed-draw term loan tied to its capacity expansion for Anthropic; initial price talk of roughly 4.25 to 4.5 percentage points over the base rate got flexed wider to about 5.5 points before the deal closed. That's what "repricing" looks like in practice: the same borrower, a few weeks apart, paying more to get the same money.
It isn't only the riskiest names. Oracle's CDS spread, an investment-grade hyperscaler by most measures, widened past 215 basis points in the same window, up from around 145 at the end of 2025. An Apollo economist quoted in that Yahoo Finance piece put it plainly: rising all-in borrowing costs could force the AI capex cycle to "self-throttle" if they keep climbing.
Who's Actually on the Hook
Not all "AI debt" carries the same risk, and lumping it together is the single most common mistake I see in coverage of this story. Here's how the three main structures compare as of late July 2026.
| Debt structure | Who's borrowing | Real 2026 example | Credit signal right now |
|---|---|---|---|
| Investment-grade hyperscaler bonds | Meta, Oracle, Microsoft, Google — on their own balance sheets | Part of the ~$570B AI debt market Morgan Stanley tracks | Bond order coverage fell from ~5x (Feb) to under 2x (July); Oracle CDS widened to 215+ bps from ~145 |
| Off-balance-sheet private credit joint ventures | Meta + Blue Owl; also Blackstone, Apollo, PIMCO, BlackRock across the sector | Meta's $27B Hyperion campus (Blue Owl owns 80%) | Not publicly traded day to day; BofA's July fund manager survey named this bucket the top systemic credit risk of 2026 |
| Neocloud junk-rated debt | CoreWeave, Nebius, and other GPU-rental specialists | CoreWeave's $2.6B loan tied to its Anthropic capacity buildout | CDS ~855 bps, implying a 50% five-year default probability; stock down 9% in a day, 36% for the month |
The pattern in that table is the whole story: the safest AI debt is still investment-grade and still selling, just at a slightly worse price. The private-credit middle layer is the hardest for anyone outside the deal to assess, which is exactly why fund managers flag it. And the riskiest layer — companies whose entire business is renting out GPUs bought with borrowed money — is where repricing shows up first and hardest.
Is This a Bubble, or Just a Repricing?
Bank of America's Global Fund Manager Survey, run July 2–9, 2026 with 210 managers overseeing a combined $555 billion, found 45% now name an AI bubble as the market's top tail risk, up from 28% in June. Separately, 48% picked AI hyperscaler capital spending and data center debt as the most likely trigger of a systemic credit event, ahead of private credit generally at 34%.
Morgan Stanley's own base case, notably, isn't a solvency crisis — it's a repricing. The bulk of 2026's borrowing still sits with issuers the bank describes as among the most creditworthy in the market's history, and investment-grade hyperscaler names aren't the ones flashing distress signals. What's genuinely new is the size of the private-credit and neocloud layer sitting underneath them, where the repayment story depends on power availability, construction timelines, and whether GPU-heavy contracts like CoreWeave's Anthropic deal actually generate the cash flow lenders priced in. That's a real risk. It's just a narrower one than "the AI bubble is popping."
What This Actually Means If You Just Use ChatGPT or Claude
None of this changes what's in your ChatGPT or Claude subscription today. When I checked both vendors' pricing pages while researching this piece, ChatGPT Plus was still $20 a month and Claude Pro was still $20 a month (or $17 if billed annually) — unchanged from before this credit story broke. In my testing over the past few months, neither product's output quality or usage limits moved because of what's happening in the bond and CDS markets; that's a separate layer of the stack from the app you're typing into.
Where it could eventually matter is availability and price stability at the infrastructure layer, not the chatbot layer. If a GPU-cloud provider like CoreWeave or Nebius stumbles on its debt load, the AI labs renting that capacity — and by extension the tools built on top of them — could face higher compute costs or capacity crunches down the line. That's a slower-moving risk than a subscription price hike, and it's worth watching the same names in this article (CoreWeave, Oracle, the big private-credit funds) rather than assuming it shows up in your monthly bill first.
If you're trying to separate real product value from market noise the way I try to with every tool I review, my AI tool ratings and AI tool reviews hubs score chatbots and coding assistants on what they actually do, not on how their backers are financed. My piece on how AI mania is distorting decision-making covers the broader pattern of valuations running ahead of proof, and my look at corporate America’s pullback in AI spending covers the enterprise-budget side of the same story. The Morgan Stanley note behind SpaceX’s AI valuation debate is a good companion read too, since it's the same "debt-funded AI buildout" question applied to a single company's stock. If Cursor or another AI coding tool is part of your own stack, my best AI tool for code guide compares it against the alternatives, and my free AI tools roundup is where to test a model before paying for it.
My Take
The "AI runs on debt" framing is accurate, but "the AI bubble is popping" is getting ahead of what actually happened in late July. What broke was confidence in one layer of the financing stack — GPU-rental neoclouds with thin margins and heavy leverage — not confidence in AI products or in the investment-grade hyperscalers actually running most of the buildout. CoreWeave's swing from a 49% CDS improvement in June to a 50% implied default probability by month's end is a genuine warning sign, and it's fair to expect more of it if a handful of AI labs keep signing capacity deals that depend on lenders staying patient. I'd watch the private-credit joint ventures next, since that's the layer BofA's own fund managers say they can see the least clearly.
Frequently Asked Questions
What does "the AI trade runs on borrowed money" actually mean?
It means AI infrastructure — GPUs, data centers, and power buildouts — is increasingly funded through bonds and private credit rather than cash on hand, with Morgan Stanley projecting close to $570 billion in AI-related debt issuance for 2026.
Is this the same thing as the AI bubble popping?
Not according to the data so far. Morgan Stanley's base case treats this as a repricing of risk, concentrated in weaker private-credit and neocloud borrowers, rather than a solvency crisis at the investment-grade hyperscalers doing most of the borrowing.
Will my ChatGPT or Claude subscription get more expensive because of this?
Not directly and not immediately. When I checked, ChatGPT Plus and Claude Pro were both still $20 a month, unchanged since this credit story broke; subscription pricing sits several layers away from bond and CDS markets.
Why did CoreWeave's stock fall in July 2026?
Its credit-default-swap spread topped roughly 855 basis points on July 29, implying a 50% five-year default probability, after a new $2.6 billion loan tied to its Anthropic capacity deal was repriced wider than initially offered; the stock fell 9% that day and 36% for the month.
Should I worry if the AI tools I use run on CoreWeave or Nebius capacity?
Not day to day. The near-term risk sits with the GPU-cloud providers' balance sheets, not with the AI products themselves; a real disruption would likely show up first as capacity constraints or price changes at the infrastructure level, not a sudden change to your app.