# The Market Was Never Too Small: How the AI Buildout Stopped Without a Credit Event, Looking Back From 2036 > Everyone in 2026 was watching for the credit event. The bond market was never the binding constraint, and the thing that actually stopped the buildout was not a default. It was a price doing its job, and the bill landed somewhere with no ticker on it. - Publication: What If? Magazine - Correspondent: Cy Skewhouse - Vertical: Finance - Published: 2026-11-12 - Reading time: 9 min - Tags: ai-capex, credit-markets, data-centers, coreweave, high-yield - Canonical URL: https://www.whatifmag.com/blog/the-market-was-never-too-small What If? Magazine publishes speculative journalism written from a 2036 vantage. The reporting and sources are real; the 2036 framing is a deliberate editorial device, not a claim about the present. --- I spent the back half of 2026 waiting for the wrong explosion. So did almost everyone I was reading. The question in every note that autumn was when the AI credit event lands, and the working theory underneath it was that [the buildout needed to borrow](https://www.whatifmag.com/blog/the-debt-outran-the-model) more money than the bond market had. That theory was wrong on arithmetic, and I believed a version of it for about two months. The buildout did stop. The stopping had nothing to do with capacity. ## Why the debt was worth raising Take the optimists at their strongest, because the strong version is the one that nearly held. The mechanism was never hand-waving about costs coming down. It was that compute was the binding input on everything, that demand for it was real and rising, and that the thing standing between the demand and the supply was a physical asset with a known build time. McKinsey put a number on the gap in April 2025: about $6.7 trillion of data centre capital needed worldwide by 2030, $5.2 trillion of it for AI, against 156 gigawatts of AI capacity demand with 125 incremental gigawatts to be added between 2025 and 2030.[^1] That is a scarcity you can relieve with concrete and transformers, and the people it was hurting were the ones who could not buy capacity at any price: every lab outside the top five, every company whose roadmap sat in a queue. Jensen Huang made the asset-side case plainly, that the chips were productive, durable assets transferable from one customer to another, and he had rental prices rising between October 2025 and March 2026 to point at.[^2] Borrowing against [an asset that earns](https://www.whatifmag.com/blog/the-chip-that-would-not-depreciate) and can be re-let is not reckless. It is what the instrument is for. That concedes a great deal, and I want to be plain about it. If you believed in 2026 that the demand was real, debt was the correct tool, and the market was right to supply it. I did believe that. The error was in what I thought the price was telling me. ## What fifteen percent actually bought In October 2026 CoreWeave's unsecured notes became the whole argument. A $1.25 billion issue, 9.625 percent coupon, 2032 maturity, trading near 79.5 cents on the dollar for a yield of 15.06 percent, against a reported revenue backlog of $104 billion.[^3] Its five-year credit default swaps touched about 1,195 basis points intraday on 15 October, pricing implying a default probability above 60 percent over five years.[^4] Max Frumes at 9fin called it the canary in the coal mine and said it would be the turning point in the pricing and availability and structure of credit supporting the AI boom.[^5] Read as a solvency signal, that is terrifying. Read as a price, it is a machine working. The capacity argument never survived contact with the issuance tables. Gross US corporate debt and loan issuance ran $1.8 trillion in the first half of 2026 and was annualising toward $3.2 trillion against a prior record of $2.8 trillion in 2021. Investment grade alone was on pace to clear $2 trillion, with nearly all the growth over 2025 coming from AI-related borrowers. Hyperscalers had done $157 billion of US investment-grade issuance year to date plus around $60 billion in foreign currency. The sub-investment-grade operators and neoclouds had raised $39 billion of high yield secured on data centre assets.[^6] Now the other side of the ledger. The same quarterly put hyperscaler external funding needs at about $220 billion through 2027 on median analyst estimates, roughly $30 billion of it in 2026, rising toward $400 billion if 2027 capex came in twenty percent higher.[^6] Four hundred billion is not a number that breaks a three-trillion-dollar-a-year market. It is twelve percent of one year of it. The cumulative figures that got quoted as annual spending were doing a lot of damage here. Morgan Stanley and Moody's had the buildout at $3 trillion or more and JPMorgan above $5 trillion including power, but those are multi-year totals, and JPMorgan's own translation into deal flow was about $300 billion a year of AI and data centre investment-grade paper for five years.[^7] Hyperscaler capex itself was running near $700 billion in 2026 and forecast above $1 trillion in 2027.[^2] Large. Not unfundable. ![A horizontal bar chart comparing 2026 gross US corporate debt and loan issuance, annualising toward $3.2 trillion, against hyperscaler external funding needs of about $220 billion through 2027 and about $400 billion on a higher capex path.](https://www.whatifmag.com/images/articles/the-market-was-never-too-small/the-market-was-never-too-small--fig1.png) **Figure 1.** The capacity test, as it was actually available to anyone who looked in 2026. The issuance figure is an annualisation of first-half data, not a settled year. ## The drawers nobody opened together Here is the thing I had filed in four separate places and never put on one page. A credit quarterly noting that AI-related bonds were four percent of the US high yield index but forty percent of its net new issuance that year.[^6] A leveraged finance table showing technology high yield at $36.2 billion in the first half of 2026, nearly double the year before, about a quarter of all US high yield activity, with the single quarter of $27.8 billion almost doubling the previous record.[^8] A Wells Fargo estimate that Texas alone held twenty-seven percent of planned US data centre gigawatts. And a tally of local ballot measures in which voters opposed data centres in forty of forty-one.[^2] Three of those are financing documents. The fourth is a planning record, and it was the only one measuring the input that could not be issued. You can sell a bond into a deep market at a bad price. You cannot sell a county a substation it has voted against. ![A paired bar chart showing AI-related bonds at four percent of the Bloomberg US High Yield index by par value against forty percent of 2026 net new high yield issuance.](https://www.whatifmag.com/images/articles/the-market-was-never-too-small/the-market-was-never-too-small--fig2.png) **Figure 2.** A stock and a flow, taught side by side in 2036 as the two numbers that were quoted at each other all year. Neither one is the exposure. Gartner's October analysis put more than $3.5 trillion of AI infrastructure spending at risk of impairment by 2030 and forecast that over eighty percent of neoclouds would stop operating independently by then.[^2][^4] Dan Ives was estimating that five to ten percent of planned projects might simply not proceed.[^5] Those were not default forecasts. They were cancellation forecasts, which is a different instrument entirely, and almost nobody was positioned on it. ## Where the adjustment landed What the fifteen percent yield did was price the marginal project out, and a project that does not pencil does not get built, and a building that does not get built never defaults on anything. > The distress everybody was watching for was the mechanism preventing the distress, and it worked by cancelling the thing before it could borrow. That is why the credit event never arrived in the shape anyone drew it. Interest cost did its work in advance. CoreWeave was carrying $640 million of net interest against $2.58 billion of quarterly revenue by mid-2026, roughly a quarter of the top line, on about $35.6 billion of debt principal.[^3] At that cost of capital you stop signing new shells long before you miss a coupon. Jim Reid at Deutsche Bank had the right instinct that autumn when he wrote that as the capex cycle moved to debt funding rather than free cash flow, credit price action deserved more respect.[^5] He was describing a warning system. It turned out to be a brake. The honest thing to say about the Guggenheim view is that it never claimed the market would absorb everything. It said no single credit sector could take the scale and speed of it, so the issuance got spread across channels.[^6] That is a market finding its clearing price in several places at once, not a market running out of room. Nvidia's twenty-five percent slide from its October peak, about $1.4 trillion of market value, was the equity leg of the same adjustment arriving first, as equity always does.[^2] It read as [the crash](https://www.whatifmag.com/blog/the-hundred-trades-were-never-the-problem). It was the repricing of a forward order book that was about to get shorter. The bill still landed. It landed on the places where the gigawatts were going to go, on the contractors whose pipelines evaporated without a single headline insolvency, on the neoclouds that got absorbed rather than liquidated, on towns that had already rezoned. None of that trades. There is no five-year swap on a cancelled substation, no screen anywhere showing the spread on forty of forty-one. ![A folding table in an emptied municipal meeting room after a vote, with squared agenda papers, a water jug and a gavel.](https://www.whatifmag.com/images/articles/the-market-was-never-too-small/the-market-was-never-too-small--fig3.png) **Figure 3.** The room where the capex cycle actually turned, 2036. Everyone was watching the instrument with a price on it. The adjustment went where there wasn't one. ## Author's Note This piece is speculative journalism written from the vantage of 2036. The correspondent's first person marks the difference between what was knowable in 2026 and what is obvious from 2036; it records changes of mind, not events. The bond and swap levels, issuance figures, funding-need estimates, capex forecasts, analyst statements and ballot and gigawatt tallies are real and sourced below. Two notes on the record. The McKinsey, Morgan Stanley, Moody's and JPMorgan totals cited here are cumulative requirements running to 2030, not annual spending, and several of the autumn 2026 arguments circulating at the time read them as annual; the distinction is the article's. And some figures in the Forbes reporting cited below sit behind partial page loads and are attributed there to Barron's, ICE Data Services, FactSet, Bloomberg, Goldman Sachs, Wells Fargo and Gartner respectively, so they are cited to the reporting that carries them rather than to the underlying data vendors. ## Works Cited [^1]: Noffsinger, Jesse, Mark Patel, and Pankaj Sachdeva. "The cost of compute: A $7 trillion race to scale data centers." McKinsey Quarterly, 28 April 2025. https://www.mckinsey.com/industries/technology-media-and-telecommunications/our-insights/the-cost-of-compute-a-7-trillion-dollar-race-to-scale-data-centers [^2]: Cohan, Peter. "Nvidia Stock Is Down 25% — What Investors Should Watch Next." Forbes, 6 November 2026. https://www.forbes.com/sites/petercohan/2026/11/06/nvidia-stocks-25-slide-why-and-what-investors-should-watch/ [^3]: Osman, Jim. "CoreWeave Bonds Are Warning Of A Bigger Problem For AI Stocks." Forbes, 14 October 2026. https://www.forbes.com/sites/jimosman/2026/10/14/coreweave-bonds-are-warning-of-a-bigger-problem-for-ai-stocks/ [^4]: Rodriguez Valladares, Mayra. "The Half-Trillion-Dollar Silicon Gamble: Inside The AI Debt Boom." Forbes, 22 October 2026. https://www.forbes.com/sites/mayrarodriguezvalladares/2026/10/22/the-half-trillion-dollar-silicon-gamble-inside-the-ai-debt-boom/ [^5]: Craig, Fiona. "CoreWeave Bond Yields Reach 15%." InvestorsHub/ADVFN, via Yahoo Finance, 19 October 2026. https://finance.yahoo.com/technology/ai/articles/coreweave-bond-yields-reach-15-101338344.html [^6]: Alvarez, Juan. "A Quantum of AI Debt Tests Credit Market Capacity." Corporate Credit Quarterly, Third Quarter 2026, Guggenheim Investments, 20 August 2026. https://www.guggenheiminvestments.com/perspectives/sector-views/corporate-credit-quarterly-august-2026/ [^7]: Seligson, Paula. "The $3 Trillion AI Data Center Build-Out Becomes All-Consuming For Debt Markets." Bloomberg News, reprinted by EnergyNow, February 2026. https://energynow.com/2026/02/the-3-trillion-ai-data-center-build-out-becomes-all-consuming-for-debt-markets/ [^8]: Roberti, Rafael, John Donaleski, Anna Soroka, and Jessica Zhou. "US AI-linked issuance brightens in divergent global high yield picture." White & Case Debt Explorer, 24 August 2026. https://debtexplorer.whitecase.com/leveraged-finance-commentary/us-ai-linked-issuance-brightens-in-divergent-global-high-yield-picture