On September 24th, Oracle sent a force majeure notice to the developer of Project Jupiter, a Stargate data centre campus in New Mexico. Oracle was reserving the right to delay payments if the facility did not come on line in 2028 as planned.
Project Jupiter has faced a string of setbacks, most of which stem from energy supply to the site. A pipeline to deliver gas to the site has been delayed to 2027 after regulators repeatedly denied the requisite permits.
Oracle's caution points to three pressures building on the AI boom at once: borrowing is more expensive, revenue is lagging well behind forecasts, and the build-out is running into physical and political limits. These would matter less if the boom were being judged on its cash flows alone. It has been judged instead on a story, and behavioural finance calls our habit of letting a vivid story stand in for evidence the narrative fallacy. But when that story gets revised, the losses travel well beyond the companies telling it, into the index funds, pension plans and private credit funds that investors hold.
Oracle’s quarterly interest expense rose 55 per cent to US$1.43 billion. CoreWeave, an AI-focused cloud provider, saw its quarterly interest expense more than double to US$640 million, with Q3 interest expense guided to US$860–940 million. In late September, the yield on the US 10-year treasury reached about 5.17%; the highest it has been since 2007. Much of that has to do with the current US fiscal challenges and volatile oil prices but it impacts other borrowers. JPMorgan expects US$4.1 trillion in AI-related debt to be issued through 2030. Apollo Global Management’s Chief Economist Torsten Slok warned that pricier debt could slow the spending cycle itself.
The market's verdict arrived on October 13th. The yield on CoreWeave's US$1.25 billion of bonds due in 2032 closed at a record 14.5 per cent. Its shares fell as much as 15 per cent on Wednesday and Oracle’s fell about 7 per cent. CoreWeave's debt has grown to roughly US$52 billion from US$15-billion a year earlier, and it lost US$626 million last quarter on revenue of US$2.58 billion. D.A. Davidson’s Gil Luria estimates CoreWeave borrows at about 9 per cent and warns the equity could be wiped out if the AI boom fades.
Much of AI’s debt is funded from another source. Ares Management’s credit team estimates that U.S. insurers hold roughly US$9 trillion of invested assets and warns that AI-related financing risk is increasingly reaching insurance portfolios, including bonds bought to back retirees’ annuities; the Financial Times reports that CoreWeave is now courting that same pool of capital.
Bain & Co. has done the math. To justify the money going into chips and data centres, the industry would need annual revenue of US$6 trillion by 2031. Existing AI products, on Bain's most generous assumptions, would bring in US$1.8 trillion. That leaves US$4.2 trillion to find, which Bain says would take "nothing short of a technological miracle".
The reporting numbers are not reassuring. OpenAI put its annualized revenue at about US$50 billion at the end of September, some US$20 billion below the US$70 billion previously communicated to investors. Chris Wood of Jefferies goes further, expecting "massive capital destruction" if cheaper Chinese models keep taking market share. In the week to July 19, Chinese models handled 36.39 trillion tokens on OpenRouter, against 7.39 trillion for the leading US ones. Now, tokens are not dollars, and OpenRouter leans toward cost-conscious developers. Still, if the models become commodities while the chips stay expensive, the industry will not run short of AI, it will run short of returns.
The AI buildout is racing towards the walls of physical and political limits. In September, Virginia's governor signed an executive order for more stringent requirements for new data centres and denying fast-track approval to any project that needs more than 25 megawatts. In Olds, Alberta, a proposed C$10 billion data center drew hundreds of rural Albertans to intervene. Since the start of 2025, more than 340 anti-data-centre protests have taken place, and a coordinated July 2026 day of action alone included 142 protests in 42 US states.
After the US and Israel attacked Iran, drone strikes hit three AWS data facilities in the UAE and Bahrain on March 1, 2026. Iran later threatened Stargate UAE specifically and, as a result, the UAE is redesigning its campus as smaller, partly underground sites. The point for investors, both with the protests and physical strikes, is that the costs and timing of build out are increasing, at a time when the hyperscalers can ill-afford it.
What does it mean for the investor if the AI story does come apart? Morningstar puts the five technology giants at about 30 per cent of the S&P 500, and in the United States more than 80 per cent of 401(k) plans default into mostly cap-weighted target-date funds. Scotiabank estimates that a 20-per-cent U.S. equity decline would shave 0.8 per cent from the level of American GDP, and 0.5 per cent from Canadian GDP by 2027, as falling household wealth restrains consumption and higher financing costs deter investment.
In the third quarter, redemption requests reached 39 per cent at one Blue Owl technology-lending fund and nearly 17 per cent at its flagship credit fund; both paid out only about 5 per cent. The immediate strain stems mainly from loans to software firms that AI may disrupt, rather than data centres. But private credit is also a major financier of the AI build-out, so markdowns on data-centre loans would surface through the same semi-liquid funds. Gating may prevent an immediate fire sale, but it creates redemption queues and forces investors seeking cash to sell more liquid assets, including index funds and other public-market holdings.
What makes this fragile is the narrative that fuelled it. Now the account is being revised. Sam Altman of OpenAI said late last year that investors were overexcited and someone would get burned. Last month Dario Amodei of Anthropic urged the industry to ‘pace the frontier’, Altman agreed, and OpenAI ruled out an IPO this year. Each move is defensible, and prudence is the stated reason. But a revised story also lowers the bar a company is judged against, and investors have watched founder narratives change before: WeWork, Theranos, FTX. Those cases differ in material respects. But from the outside a candid reset and a protective one look the same. So investors cannot rely on anyone's story, the bullish original or the cautious sequel. Check what your funds actually own, read the redemption terms of anything sold as semi-liquid, and test every claim against the revenue.
As Herb Stein put it: "If something cannot go on forever, it will stop," Investors just need to make sure that they are ready for that stop
