The cost of insuring Meta’s and Nvidia’s debt against default has climbed to record highs this month, the Financial Times reported on 20 October 2026: five-year Meta CDS at 108 basis points and Nvidia at 101 basis points.
On 16 October 2026, a Bank of America report found AI capital spending running 50% above forecasts this year while revenue ran only 20% above, according to a summary by TechFlow Research and a post by AI commentator Rohan Paul. We have not read the Bank of America report itself.
Business Pill · FINANCIAL CONDITIONS
A short explainer on financial conditions, how rates and market mood change the price of capital. It teaches the general idea only and says nothing about any company in this story.
The key insight: As we read it, two sets of numbers sit side by side this month: records in what it costs to insure Meta’s and Nvidia’s debt, per the FT, and a Bank of America finding, as summarised, that AI capex ran 50% above forecasts while revenue ran 20% above.
More than $9bn worth of CDS linked to bonds issued by Oracle, SpaceX, Alphabet, Amazon, Meta, Broadcom and Nvidia was traded on average each week in September, up from less than $2bn in July, the FT reported, citing Morgan Stanley Research.
The FT said the price of protection against default at Alphabet and Amazon is also at all-time peaks. Oracle’s CDS was quoted at 322bp on Monday, up from 144bp at the start of the year, per the FT; CNBC put Oracle’s five-year spread at 315 basis points on 19 October, a 311% increase from a year prior.
The five-year CDS tied to CoreWeave’s debt is trading at about 1,090bp, the FT reported. SoftBank Group’s five-year CDS stood at 432 basis points on 20 October, the highest since early November 2022, according to data compiled by Bloomberg.
AI-linked companies have raised an estimated $466bn of debt so far in 2026, up from $101bn in the same period in 2025, the FT reported, citing Morgan Stanley. A basis point is a hundredth of a percentage point, so by our arithmetic 108 basis points is 1.08 percentage points a year.
Five-year credit-default swap levels on AI-linked companies, October 2026: CoreWeave about 1,090 basis points and Meta 108 (a record), with Nvidia at 101, also a record (Financial Times, 20 October); SoftBank 432, the highest since early November 2022 (Bloomberg data, 20 October); Oracle 315 (CNBC, 19 October; the FT quoted 322 on Monday). Circle areas are to scale. The Bank of America figures in the subtitle are from its 16 October report as summarised by TechFlow Research and Rohan Paul.
What Was Posted on X
Bank of America’s new research finds AI capital spending running 50% above forecasts this year while revenue runs only 20% above, and that mismatch drives the rest of its analysis.
Hyperscalers’ weighted average cost of capital has risen from 4% in 2023 to 10%, as risk-free rates sit more than 4% above pandemic lows and debt costs, equity costs and betas have all climbed.
The BoA research still projects capex for the top 5 hyperscalers plus neoclouds at $956B in 2026, $1.32T in 2027 and $1.5T in 2028, which works out to growth of 38% and then 14%.
Its demand math puts 212GW of new data center capacity between 2026 and 2030 at about $9.5T, $2.1T above the current $7.4T capex forecast.
AI commentator Rohan Paul posted this summary of Bank of America’s research at 06:21 UTC on 20 October. He writes that it finds AI capital spending running 50% above forecasts this year while revenue runs only 20% above, and that hyperscalers’ weighted average cost of capital has risen from 4% in 2023 to 10%. This piece uses his post, alongside TechFlow Research’s summary, for the Bank of America figures; we have not read the report itself.
Morgan Stanley puts 2027’s hyperscaler and neocloud AI spending at $1.58T after counting data-center leases as capex.
Hyperscalers hold over $1.1T in leases that haven’t begun, which work like debt-funded purchases but never appear as capex.
About 65% of the 2027 total, roughly $1.03T, lands in S&P 500 revenue, and Nvidia takes about $690B of it, or 44% of the entire budget.
That leaves roughly $340B for the other 499 index members, while about $550B flows outside the index to Korean memory makers, Taiwanese ODMs and private companies.
With 2027 the projected peak cash shortfall, GPU-rental neoclouds funding builds through project debt and convertibles are the most exposed to financing costs.
Rohan Paul posted this summary of Morgan Stanley’s research at 06:17 UTC on 20 October. He writes that Morgan Stanley puts 2027 hyperscaler and neocloud AI spending at $1.58 trillion after counting data-center leases as capex, and that hyperscalers hold over $1.1 trillion in leases that haven’t begun. The section above uses Fortune’s report for the figures and his post for the financing-cost point; we have not read the note itself.
Bloomberg posted Jonathan Levin’s opinion column on X at 10:10 UTC on 20 October. The post says the corporate bond market is waking up to the risks of AI’s huge capex needs, and that the Fed will need to exercise caution to avoid a negative spiral. That is Levin’s view, not reporting.
What Bank of America Found
According to TechFlow Research’s summary, Bank of America noted that AI capex exceeded expectations by 50% this year, while revenue beat forecasts by only 20%. Rohan Paul’s post gives the same two figures and says that mismatch drives the rest of the analysis.
The report argues that hyperscalers’ weighted average cost of capital, the return investors require on their money, has risen from 4% in 2023 to 10%, TechFlow said. Risk-free rates are over 4 percentage points higher than pandemic lows, and debt costs, equity costs and hyperscaler betas are all rising, per the summary. By our arithmetic, 10% is 2.5 times the 2023 level.
Bank of America still projects capex for the top five hyperscalers plus neoclouds at approximately $956 billion in 2026, $1.32 trillion in 2027 and $1.5 trillion in 2028, according to TechFlow. It estimates about 212 gigawatts of new data center capacity will be needed between 2026 and 2030, which at a conservative $40 billion per gigawatt comes to about $9.5 trillion, $2.1 trillion above the current $7.4 trillion forecast.
The report’s scenarios turn on AI revenue per gigawatt, TechFlow said. Its base case shows cumulative excess returns of approximately $200 billion between 2026 and 2030, the bear case cumulative losses of around $500 billion and the bull case near $1 trillion.
A Separate Bank of America Warning
In a separate note reported on 16 October 2026, Bank of America investment strategist Sebastian Raedler wrote that “most major capital cycles of the past have ended in disappointment, even though the underlying technology often lived up to expectations,” according to 24/7 Wall St.
Raedler put forward hyperscaler capex expectations at 3% of US GDP, above the 2.2% of the 1870s railroad boom and the 1.1% of the 1990s tech boom, 24/7 Wall St reported. It said his warning targets investor returns and stops short of a crash forecast.
The Leases That Don’t Show Up as Capex
Morgan Stanley’s North America Insight team, led by Nicholas Lentini, estimates that hyperscalers and AI cloud providers will spend $1.58 trillion in 2027, up roughly 60% from about $980 billion this year, and $1.8 trillion in 2028, Fortune reported on 19 October 2026 from the bank’s 15 October note. About $195 billion of the 2027 figure is data center leases, which never show up as capital spending on a cash flow statement; the report calls leases “synthetic capex,” according to Fortune. Hyperscalers have disclosed more than $1.1 trillion in lease commitments that haven’t started yet, Fortune said.
Rohan Paul’s summary of the same research, posted on X on 20 October, says those leases work like debt-funded purchases, that 2027 is the projected peak cash shortfall, and that GPU-rental neoclouds funding builds through project debt and convertibles are the most exposed to financing costs. We have not read the Morgan Stanley note.
Customer concentration is the other worry raised this week. On his podcast “The Real Eisman Playbook,” investor Steve Eisman said Nvidia’s latest 10-Q showed “that 70% of Nvidia’s accounts receivable is from five accounts” and called it “a tremendous amount of concentration risk,” Benzinga reported on 20 October 2026; he also said he is not telling people to short the whole AI trade. Nvidia’s 10-Q for the quarter ended 26 July 2026 lists five unnamed direct customers at 22%, 14%, 13%, 11% and 10% of its accounts receivable balance, 70% in total by our arithmetic. That is a receivables figure, not a share of revenue. Separately, Rohan Paul’s summary of the Morgan Stanley research says 91.7% of Nvidia’s 2027 revenue is exposed to infrastructure demand; we have not read that figure in the note itself.
An opinion from Bloomberg: in a 20 October column headlined “CoreWeave’s Bond Rout Should Make the Fed Cautious”, Bloomberg Opinion columnist Jonathan Levin wrote that “Corporate credit, monetary policy’s sleeping giant, is finally stirring.” He noted that CoreWeave’s bonds dropped this month, sending yields on its 2032 notes as high as 15%, and that yields for junk and investment-grade debt reached the highest since 2025 and 2023, respectively. Bloomberg’s post on X summarised his argument: the Fed “will need to exercise caution to avoid a negative spiral.” This is commentary, not reporting, and we have read only the column’s opening.
Update, 20 October: TD Cowen Says Nvidia May Be Backing Oracle Leases
TD Cowen analyst Michael Elias said his industry checks suggest Nvidia could be guaranteeing some lease payments or renting facilities itself before subleasing them to Oracle, TipRanks reported on 20 October 2026. In plain terms, Nvidia would be standing behind Oracle’s rent. Under one possible arrangement described in the report, Nvidia would provide financial guarantees that reassure data center owners they will receive payments if Oracle runs into financial trouble; another possibility is that Nvidia would lease facilities itself and then rent the space to Oracle. Either approach could make lenders more willing to finance construction, according to TipRanks.
Per TipRanks, neither company has confirmed such an arrangement, and TD Cowen did not say how large it could be. Elias also said Oracle is discussing additional financing arrangements with suppliers as it works to deliver computing capacity that customers have already ordered. As we read it, if TD Cowen’s checks are right, part of Nvidia’s exposure to the build-out would sit in lease commitments of the kind described above, carried by the chip supplier rather than only by the tenant.
Update, 21 October: Ellison Has 413 Million Oracle Shares Pledged
Oracle founder Larry Ellison added to a multibillion-dollar borrowing spree to help pay for his son David’s takeover of Warner Bros. Discovery, putting up shares in Oracle, Tesla, SpaceX and Paramount as collateral and increasing his borrowing from five banks, Citigroup, HSBC, JPMorgan, UBS and BNP Paribas, The Wall Street Journal reported on 20 October 2026, citing regulatory filings in California. Oracle’s proxy statement, filed on 25 September, says Ellison had pledged 413 million shares of Oracle common stock as collateral to secure certain personal indebtedness. The Journal values those shares at around $58 billion and says the figure was 346 million a year earlier. Oracle also said in September that he had scrapped a plan to sell roughly $7.5 billion of stock, the Journal reported.
Correction, 29 October 2026: An earlier version attributed the roughly $58 billion value of Ellison’s pledged shares to Oracle. Oracle’s 25 September proxy statement gives the 413 million share count but no dollar value; the $58 billion is the Journal’s figure.
The banks that disclosed how much Oracle stock they hold security over reported 87 million shares, up from around 30 million a year ago, according to the filings cited by the Journal; it was not clear exactly how much cash Ellison borrowed. Oracle has said its board governance committee reviews Ellison’s pledging arrangements quarterly and has determined he can repay his personal loans without having to sell pledged shares, per the Journal. As we read it, this is the founder’s personal borrowing, not Oracle’s own debt; it sits in this credit story because it puts more lenders’ collateral on Oracle’s share price while the company’s AI build-out financing, including the lease arrangements described above, is under scrutiny.
Update, 21 October: A Report Likens Data Centre Debt Sales to Subprime
“Speculative data center debt is being repackaged and sold to pension and retirement funds, university endowments, and other institutional investors in a manner eerily reminiscent of the subprime bubble,” says “Brace for Impact: Protecting People and the Economy When the AI Bubble Bursts”, a report released on 16 October 2026 by the Open Markets Institute and the Americans for Financial Reform Education Fund. Its authors are Matthew Scherer, an Open Markets Institute fellow, and Maya Jenkins, a senior policy analyst at Americans for Financial Reform Education Fund, according to the groups’ press release. The line circulated on X on 21 October credited to “AFR”, which here is Americans for Financial Reform, not the Australian Financial Review.
The report says data center loans “are often securitized — that is, sliced up and packaged into bonds that are sold to investors, often institutional investors such as insurers and pension funds,” and that many lenders have issued them on a “booked-but-not-billing” basis, meaning the loan is made on a model developer’s promise to lease the facility rather than on payments that have started. It calls on regulators, Congress and state policymakers to demand financial transparency across the AI ecosystem, and says its purpose “is not to predict a specific sequence of events or portray an economic crisis as inevitable.” As we read it, this is an advocacy report rather than a regulator’s finding, and it adds a policy voice to the credit-market signals above.
Our follow-up sets out the report’s figures, including the Wall Street Journal calculation it cites of at least $3.1 trillion in off-balance-sheet commitments for nine tech giants: Report Says Big Tech’s $3.1T Off-Book Debt Tops Subprime.
The Other Side
The same 16 October report keeps a long-term bullish stance on US equities and views the S&P 500 as offering balanced risk-reward over the next 12 months, according to TechFlow. It says hyperscalers trade at a 24% discount to the S&P 500, per the summary.
Bloomberg Opinion columnist John Authers wrote that CoreWeave’s credit and business model deserve cold-eyed scrutiny, but that this panic looks more like 2023 than 2007, according to Bloomberg’s post on X.
A lender’s view: HSBC Chief Executive Georges Elhedery, in an interview with CNBC’s Dan Murphy published 20 October, urged selectivity over data-centre financing, according to CNBC’s summary of the interview, which ran under the headline “Have to be ‘selective’ with data center investments”. We have read CNBC’s summary, not a full transcript.
The Structural Read
As we read it, the FT’s data show the price of default protection rising even for the largest AI spenders, with Meta at 108 basis points and Nvidia at 101, and with Alphabet and Amazon at all-time peaks.
As we read it, Bank of America’s cost-of-capital figure, as summarised by TechFlow, sets the bar those projects are measured against: a weighted average cost of capital of 10%, up from 4% in 2023.
As we read it, the report’s own scenarios span a wide range, from cumulative losses of around $500 billion in the bear case to near $1 trillion of excess returns in the bull case, and the report keeps a long-term bullish stance on US equities.
Manish Kabra, head of US equity strategy at Société Générale, to the Financial Times
“CDS has become a key bellwether of investor sentiment towards hyperscaler funding and the broader AI trade”
Three Implications
RECORD PRICES Five-year Meta CDS climbed this month to 108 basis points and Nvidia’s to 101, both record highs, the FT reported.
A BUSIER MARKET More than $9bn of CDS tied to seven AI-linked borrowers traded on average each week in September, up from less than $2bn in July, per Morgan Stanley Research via the FT.
A HIGHER BAR Hyperscalers’ weighted average cost of capital rose from 4% in 2023 to 10%, according to Bank of America’s report as summarised by TechFlow.
The framework puts it this way: “the capex race is no longer a corporate spending cycle. It is a sovereign-scale financing problem dressed in tech-company clothing”.
As we read it, credit-default swaps on AI-linked borrowers are one place where the market puts a price on that financing. The framework was published in May 2026 and does not address these October CDS moves.
What Is Not Established
We have not read Bank of America’s 16 October report or Raedler’s note; their figures here are as TechFlow Research, Rohan Paul and 24/7 Wall St report them. The capex and revenue figures are overshoots against forecasts, not a comparison of total capex with total revenue.
CDS levels are market data that move daily, and the FT and CNBC figures for Oracle differ by date and source. Nothing here links any single company’s spread to the Bank of America figures, and nothing here is a forecast.
This Week In The AI Supercycle
Execution Fault Lines, Concentrated Demand, and Geopolitical Firewalls
This piece is part of this week’s issue: the AI build is not short of demand for its parts. It is short of cheap money, and its end revenue rests on a narrow base of large, often venture-funded buyers.
Five-year CDS on Meta and Nvidia reached record highs this month, the FT reported, while a 16 October Bank of America report found AI capex running 50% above forecasts against revenue 20% above, according to TechFlow Research and Rohan Paul.
95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. This piece is built on the Financial Times’ 20 October 2026 report (read in full), TechFlow Research’s 19 October 2026 summary of Bank of America’s 16 October report (read in full), Rohan Paul’s 20 October post on X, the readable part of Bloomberg’s 20 October report on SoftBank, CNBC’s 19 October report (read in full), 24/7 Wall St’s 17 October report on a separate Bank of America note, Bloomberg’s X post for John Authers’ column, and the readable opening of Jonathan Levin’s 20 October Bloomberg Opinion column.
The Morgan Stanley section rests on Fortune’s 19 October 2026 report (read in full) and Rohan Paul’s 20 October post; we have not read the Morgan Stanley note either. We have not read either Bank of America note. The HSBC line rests on CNBC’s 20 October video headline and summary only. Nothing here is a forecast, and nothing here is financial or investment advice.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
