# The AI money wobbles: demand held through October, the financing underneath it slipped

URL: https://www.thedeepfeed.ai/posts/2026-10-16-the-ai-money-wobbles/
Category: Business
Published: 2026-10-16
Author: the-deep-feed
Tags: ai-economics, data-centers, private-credit, openai, oracle, nvidia, coreweave, softbank
Kind: deep

> Between late September and mid-October, chip sales and backlogs kept setting records while OpenAI's run-rate shrank by a definition, CoreWeave bonds hit distress prices, Firmus pulled its IPO and SoftBank paid junk rates to fund OpenAI.

## TL;DR

- Measured demand kept setting records: **Nvidia** booked **$96.2B** in quarterly revenue and guided to **$108B**, Micron posted **$54.2B**, and Oracle's contracted backlog reached **$664B**.
- OpenAI's run-rate fell from a widely reported **$68B** to about **$50B** on October 8 because the larger number counted partners' gross sales. Growth was 77% in the quarter; the drop was in the definition, and AI stocks still sold off on it.
- The financing layer is where the strain showed: **CoreWeave**'s 2032 bonds traded below **80 cents**, Firmus pulled a **$5B** IPO, Oracle served force majeure on a Stargate campus, and SoftBank paid up to **9.75%** to fund its OpenAI stake.
- Suppliers have become lenders. Nvidia's non-marketable stakes rose from **$22.3B** to **$51.2B** in six months, and the IMF counted **$1.4T** of future revenue commitments resting on circular deals.

The International Monetary Fund's managing director picked the right verb on October 13. Speaking in Bangkok as the Fund published its autumn outlook, Kristalina Georgieva described a world economy "standing on one leg, the AI leg," and then asked the question every lender in the sector had spent the previous three weeks pricing: "What if it wobbles?" [Reuters recorded the remarks](https://www.kitco.com/news/off-the-wire/2026-10-13/imf-flags-risks-global-economy-ai-market-correction-war-and-debt) at the Fund's annual meetings.

It had already started to. Between September 24 and October 14, Oracle served a force majeure notice on its largest Stargate campus. SoftBank paid junk-bond rates to fund the last tranche of its OpenAI commitment. OpenAI's widely reported $68 billion run-rate came back as roughly $50 billion. An Nvidia-backed data-center operator pulled what would have been one of Australia's largest IPOs. CoreWeave's unsecured bonds fell to prices that usually signal distress. None of these events said that people had stopped buying AI. Several came in the same weeks as record sales from the companies that make the chips and memory.

That gap is the story of the last eight weeks. The demand underneath the AI build-out held. The stack of instruments that converts expected demand into money today wobbled: annualized run-rate claims, supplier-provided financing, special-purpose vehicles and syndicated construction debt. The visible stress showed up in credit spreads, IPO order books and loan syndication desks, not in shipment numbers. The binding constraint on the build-out is moving from chips and power to the price of capital, and that price is set by lenders, who read cash flow, not by equity holders, who read growth.

# Three weeks, read in order

The events in this window came from different companies and reached the market through different reporters, so they rarely appeared together. Placed in date order, they move along one channel: from a single borrower's problem toward the price of money for the whole sector.

| Date (2026) | Event | Layer of the stack | Source |
|---|---|---|---|
| Sep 24 | Oracle serves force majeure notice on Project Jupiter, New Mexico | Construction and tenant risk | [The Next Web, citing Bloomberg](https://thenextweb.com/news/oracle-force-majeure-project-jupiter-new-mexico) |
| Sep 24 | SoftBank prices a record $11.1B high-yield bond to fund OpenAI | Investor leverage | [Bloomberg and Reuters via The Star](https://www.thestar.com.my/business/business-news/2026/09/24/softbank-raises-us111bil-in-world039s-biggest-high-yield-corporate-bond-sale) |
| Sep 30 | Micron reports $54.23B quarter | Measured demand | [Micron 8-K](https://www.sec.gov/Archives/edgar/data/0000723125/000072312526000018/a2026q4ex991-pressrelease.htm) |
| Oct 7 | Fed minutes cite "competition for capital" from AI private debt | Cost of capital | [Federal Reserve](https://www.federalreserve.gov/monetarypolicy/fomcminutes20260916.htm) |
| Oct 8 | OpenAI's run-rate restated to about $50B from $68B | Revenue claims | [CNBC](https://www.cnbc.com/2026/10/08/open-ai-revenue-nvidia-oracle-coreweave.html) |
| Oct 9 | Firmus shelves $5B IPO | Equity exit | [Reuters via NST](https://www.nst.com.my/amp/business/corporate/2026/10/1551088/nvidia-backed-firmus-scraps-us5bil-australian-ipo-poor-demand) |
| Oct 13 | IMF models an AI investment slump; flags circular financing | System risk | [Reuters via Kitco](https://www.kitco.com/news/off-the-wire/2026-10-13/imf-flags-risks-global-economy-ai-market-correction-war-and-debt) |
| Oct 14 | CoreWeave 2032 bonds trade below 80 cents; shares close down about 15% | Credit pricing | [Reuters via The Star](https://www.thestar.com.my/tech/tech-news/2026/10/15/coreweave-ceo-notes-strong-backlog-in-memo-to-reassure-staff-over-market-scrutiny) |

One correction belongs at the top. The working hypothesis behind this piece was that the strain first became visible in late September. The record does not support "first." The bank-syndication problem was reported by the Financial Times in May, S&P cut Oracle to BBB- on July 9, [per Startup Fortune](https://startupfortune.com/oracle-sues-wisconsin-over-a-7-billion-data-center-collateral-bill/), and Oracle's credit default swaps closed at a record on July 17, [per the same outlet](https://startupfortune.com/bond-investors-are-starting-to-say-no-to-the-ai-debt-binge/). What changed after September 24 is breadth. Strain that had been read as an Oracle story started appearing at SoftBank, at CoreWeave, at an Australian IPO and in the Federal Reserve's minutes, which is a different and more serious condition.

# The $18 billion that was a definition

Start with the number that moved the most stocks. On October 8 the FT reported, and CNBC confirmed, that OpenAI told investors it reached roughly $50 billion in annualized revenue at the end of September, against the $68 billion figure "widely reported late last month." A person familiar with the matter told [CNBC](https://www.cnbc.com/2026/10/08/open-ai-revenue-nvidia-oracle-coreweave.html) that the larger figure "included gross revenue from OpenAI's partners," built so investors could compare OpenAI directly with Anthropic.

The comparison problem is mechanical. [Reuters reported](https://www.thestar.com.my/tech/tech-news/2026/10/09/openai039s-annualized-revenue-20-billion-less-than-previously-signaled-ft-reports) that Anthropic counts sales made through cloud partners such as AWS and Google Cloud, pays those partners about 16% of each dollar, and that such sales made up half of its revenue last year, while OpenAI does not count partner sales. Anthropic's $65 billion July run-rate and OpenAI's $68 billion September figure were both gross numbers. OpenAI's $50 billion is the net one. On quarterly reported revenue, which removes the run-rate multiplier, Reuters put the second quarter at $6.7 billion for OpenAI and $11.5 billion for Anthropic.

The underlying business did not shrink. The same investor presentation, per CNBC's source, showed 77% total run-rate growth in the third quarter and 107% growth in OpenAI's enterprise business. OpenAI's chief financial officer said as much in public a week later:

> "So what we said is in Q3, we hit close to $50 billion. It was an incredible, it was a blockbuster quarter, right? Could make a CFO's heart sing. We grew over 70 percent, just quarter over quarter. Our enterprise business doubled."
>
> — [@firesidealpha](https://x.com/firesidealpha/status/2110410713558905044), Oct 14, 2026, quoting Sarah Friar

The market sold anyway. CNBC recorded Nvidia down 3%, Oracle down nearly 6% and CoreWeave down nearly 8% that Thursday, with AMD, Broadcom and Intel each down 4% to 5%. The [Guardian](https://www.theguardian.com/technology/2026/oct/09/openai-forecast-revenue-gap) had the Nasdaq down 1.4% and Micron down 4.8%. No demand fell by $18 billion on October 8. What fell was confidence in a metric that had been doing the work of an audited figure, at a company whose $852 billion valuation, planned 2027 listing and reported $30 billion raise at about $1.4 trillion all lean on that figure.

That is the first layer of the stack, and its weakness is structural. A run-rate multiplies one month by twelve. It does not say whether the revenue is gross or net, recurring or one-off, or prepaid. In a private market where the two largest labs are both preparing IPOs, the number is chosen by the company, repeated by investors and treated by suppliers as a demand forecast. Oracle's stock rose 5.3% on reports of the larger figure in late September, [per Crypto Briefing](https://cryptobriefing.com/openai-70b-annualized-revenue-growth/), and fell on the $50 billion figure ten days later; the business in between was the same.

![Schematic reconciling OpenAI's run-rate from $68B gross to about $50B net, with Anthropic's gross basis alongside](/post-images/2026-10-16-the-ai-money-wobbles/run-rate-reconciliation.jpg)

A second, quieter signal about revenue quality arrived in August. The FT reported, in a story summarized by [Simon Willison](https://simonwillison.net/2026/Aug/23/anthropics-best-ai-model-struggles-to-attract-users-as-cheaper-t/), that Anthropic's top-priced Fable model was struggling to win spend while cheaper models gained, using Ramp's index of billing data from 70,000 companies. Anthropic's run-rate still rose from $47 billion in May to $65 billion in July. Buyers were not leaving; they were moving down the price list. The same pattern appeared in [July's pricing cuts](/posts/2026-07-28-the-margin-is-the-message/). For a lender, the distinction matters: volume can keep rising while revenue per unit of compute falls, and the debt is sized to the compute.

# Demand, measured in filings

The bull case deserves to be stated with its numbers, because its numbers are the strongest in this story and most of them come from audited filings.

| Measure | Figure | Period | Source |
|---|---|---|---|
| Nvidia revenue | $96.2B, up 106% year on year | Quarter to Jul 26 | [Nvidia 8-K](https://www.sec.gov/Archives/edgar/data/0001045810/000104581026000073/q2fy27pr.htm) |
| Nvidia data center revenue | $89.0B, up 117% | Same | Same |
| Nvidia next-quarter guidance | $108.0B, plus or minus 2% | Quarter to late Oct | Same |
| Micron revenue | $54.23B vs $11.32B a year earlier | Quarter to Sep 3 | [Micron 8-K](https://www.sec.gov/Archives/edgar/data/0000723125/000072312526000018/a2026q4ex991-pressrelease.htm) |
| Micron operating cash flow | $43.97B | Same | Same |
| Oracle remaining performance obligations | $664B, up $209B year on year | Quarter to Aug 31 | [ERP Today](https://erp.today/oracle-q1-fy27-results-664b-backlog-ai-contracts/) |
| Oracle GPU utilization | 97.9% | Same | Same |
| OpenAI run-rate growth | 77% in the quarter | Q3 | [CNBC](https://www.cnbc.com/2026/10/08/open-ai-revenue-nvidia-oracle-coreweave.html) |

Nvidia's chief executive framed the quarter as a shift in the customer base. "This time last year, one lab alone was driving the buildout," Jensen Huang said in the release; now "multiple frontier labs" are "scaling in parallel." Micron, which sells the memory that sits beside every accelerator, reported a single quarter worth 1.45 times its entire prior fiscal year. Oracle said capacity coming up for renewal was resold at a 20% premium, per ERP Today's account of the call. At the time of writing, traders on [Polymarket](https://polymarket.com/event/will-nvidia-nvda-q3-data-center-revenue-be-above-20261008120135000) priced a 96% chance that Nvidia's next data-center print clears $95 billion and 73% that it clears $100 billion, after the October 8 sell-off.

Even the sharpest critic of AI investors in this window conceded the demand side. Broadcom chief executive Hock Tan, the same speaker who compared AI investors to casino gamblers, told a Singapore conference on October 12 that demand for computing still outstrips supply and could do so through 2029, possibly into 2030, [per The Next Web](https://thenextweb.com/news/broadcom-hock-tan-ai-investors-casino-capital). He called OpenAI's $50 billion "by itself, not that bad" for a startup.

The skeptics' track record argues the same way. Dan Luu's [audit of Ed Zitron's predictions](https://danluu.com/zitron/), which drew 1,056 comments on [Hacker News](https://news.ycombinator.com/item?id=49526069) in early September, scored most of the dated calls as wrong. Those calls included a bubble that would pop "no later than Q2 2026" and a CoreWeave that could not survive six months without new money. "Zitron is wrong on the predictions and also wrong on the reasoning," Luu concluded. Anyone arguing that October proved the bears right has to explain why the demand forecasts the bears mocked kept clearing.

# The supplier became the lender

The second layer is where demand and financing start to blur. In 2026 a growing share of AI hardware is paid for with money that the hardware's sellers supplied, guaranteed or arranged.

Nvidia's balance sheet shows the pattern without commentary. Between January 25 and July 26, 2026, the [company's own filing](https://www.sec.gov/Archives/edgar/data/0001045810/000104581026000073/q2fy27pr.htm) shows non-marketable securities rising from $22.3 billion to $51.2 billion and marketable equity securities from $12.9 billion to $42.8 billion. Purchases of equity securities totaled $42.4 billion in the first six months of the fiscal year, against $1.2 billion in the same period a year earlier. Accounts receivable rose from $38.5 billion to $63.1 billion, and long-term debt from $7.5 billion to $32.4 billion. In the second quarter, net income of $59.7 billion included $7.8 billion of other income, and operating cash flow was $24.1 billion. The filing does not name the companies Nvidia bought into, and none of these figures implies impropriety. They do show a supplier whose customers' fortunes now sit on its balance sheet as both receivables and equity. Nvidia's [deal for Hugging Face](/posts/2026-09-30-nvidia-buys-the-hub/) is the most visible use of that balance sheet this autumn; the equity line above is the less visible one.

![Labeled balance-sheet schematic of Nvidia's equity stakes, receivables and debt growing between January and July 2026](/post-images/2026-10-16-the-ai-money-wobbles/supplier-as-lender.jpg)

Broadcom has moved further. In August, Bloomberg reported that it was in talks to raise more than $60 billion through a special-purpose vehicle to finance AI chips for Anthropic and others, with a junior tranche of roughly $30 billion and a senior-secured tranche of $60 billion to $70 billion that Broadcom would partly guarantee, [per The Next Web](https://thenextweb.com/news/broadcom-60bn-ai-chip-debt-anthropic). Under that structure Anthropic does not buy the chips. Investors finance the purchase and lease the hardware to the lab, and the debt sits on the vehicle rather than on either company. The first deal under the same Broadcom, Apollo and Blackstone partnership raised $35 billion in June.

SoftBank is the clearest case of an investor borrowing to keep its commitment. On September 24 it priced $11.1 billion of bonds, the largest high-yield corporate sale on record, with dollar coupons of 8.625%, 9.25% and 9.75%, [per Bloomberg and Reuters](https://www.thestar.com.my/business/business-news/2026/09/24/softbank-raises-us111bil-in-world039s-biggest-high-yield-corporate-bond-sale). Its 2021 issue had yielded 2.125% to 5.25%. The deal's own counsel, [Latham & Watkins](https://www.lw.com/en/news/2026/09/latham-advises-on-softbank-group-record-breaking-bond-offering), said the proceeds fund "the final US$10 billion tranche of its US$30 billion follow-on investment in OpenAI." SoftBank's five-year credit default swap exceeded 400 basis points that week, against about 280 in June, and both OpenAI and SoftBank's SB Energy unit had delayed IPO plans that were meant to return cash. "Now AI-driven debt issuance has reached the high-yield market at scale," Fitch's Satoru Aoyama said.

The IMF put a frame around these loops. Its Global Financial Stability Report counted about $300 billion of deals between hyperscalers and "weaker AI firms" over three years, $650 billion of associated debt and $1.4 trillion of future commitments, according to [diginomica's reading of the report](https://diginomica.com/node/31298). Jason Wu of the Fund's markets division described the risk as a triple hit: "If there are questions about the monetization of AI, the equity investments will deplete in value, the debt may be difficult to service, and the revenue commitments in the future may not materialize." A run-rate restatement is exactly that kind of question. One company's revenue line now appears in three places: as OpenAI's run-rate, as the customer behind roughly half of Oracle's backlog, and as the asset story behind SoftBank's bonds.

# The loan that would not clear

The third layer is construction debt. Banks underwrite data-center projects, hold the loans briefly, then sell pieces to other lenders. That is ordinary practice, and the strongest objection to the October headlines rests on it.

The headline that put this on Hacker News, Startup Fortune's October 11 ["Banks are now trying to offload the AI data center debt they just underwrote"](https://startupfortune.com/banks-are-now-trying-to-offload-the-ai-data-center-debt-they-just-underwrote/), mostly repackages Financial Times reporting from May. At that time [Silicon UK, citing the FT](https://www.silicon.co.uk/cloud/datacenter/bank-data-centre-debt-629721), reported that JPMorgan, Morgan Stanley and Sumitomo Mitsui were looking for ways to distribute data-center debt, including variants of significant risk transfers that move the riskiest slice of a loan to outside investors. Lenders including JPMorgan and Mitsubishi UFJ had already spent more than six months distributing $38 billion of construction debt for an Oracle-leased project in Texas and Wisconsin, some at a discount to non-bank lenders. Matthew Moniot of Man Group told the FT the sizes were "out of scale to anything we've thought about, ever." The verified news in October is that the theme was back on the front page. No new loan sale was confirmed.

The Hacker News thread argued the point in both directions. One commenter defended the banks:

> Nothing out of hte ordinary here, but to be fair the original article isnt' a financial site so they would have no real idea about how this work behind the scenes.
>
> This doesn't seem like a story at all given how robust this debt is trading.
>
> — chollida1, [Hacker News](https://news.ycombinator.com/item?id=50094321), Oct 14, 2026

Another accepted the premise and pointed at the duration instead:

> What is unusual about this is that syndications usually clear in a few weeks but this has been stuck in distribution for six months and still not closed.
>
> — lewdwig, [Hacker News](https://news.ycombinator.com/item?id=50094006), Oct 14, 2026

The second reading has more support in the record. In August the FT reported that a Morgan Stanley-led group was preparing a $15 billion bond to move debt for a Google-backed campus in Hubbard, Texas, leased to Anthropic, off bank books, [per EnterpriseAM](https://enterpriseam.com/egypt/2026/08/06/morgan-stanley-led-bank-consortium-prepares-usd-15-bn-bond-sale-to-dump-pre-built-ai-construction-risk/). Google's guarantee covers the finished facility, not the build, so bondholders would carry construction risk and the paper was expected to price at speculative grade. The same report said banks had spent months looking for buyers for more than $50 billion of debt tied to separate Oracle-leased projects.

![Schematic of a data-center construction loan stuck at the syndication valve between underwriting banks and outside buyers](/post-images/2026-10-16-the-ai-money-wobbles/the-loan-that-would-not-clear.jpg)

Oracle sits at the center of this layer because it is both tenant and borrower. Its first fiscal quarter, [as reported by ERP Today](https://erp.today/oracle-q1-fy27-results-664b-backlog-ai-contracts/), combined the $664 billion backlog with $28.5 billion of capital spending, negative free cash flow of $5.4 billion, a $20 billion at-the-market equity sale, total debt of $125 billion and interest expense up 55%. Roughly half the backlog is tied to OpenAI. Management said most of the quarter's new AI contracts were structured as customer prepayments or bring-your-own-hardware, so they require no added Oracle capital. That is a sensible response to a funding problem, and it confirms the funding problem exists.

Then the physical schedule started to slip. On September 24, Oracle sent a force majeure notice on Project Jupiter, the New Mexico Stargate campus that could cost up to $165 billion, seeking the right to delay payments if the site misses its 2028 opening, [per The Next Web's account of Bloomberg's report](https://thenextweb.com/news/oracle-force-majeure-project-jupiter-new-mexico). The state land office had twice refused rights of way for the gas pipeline, and its start date moved from August 15 to February 1. Oracle says the project remains on schedule. A week later, [The Register](https://www.theregister.com/on-prem/2026/10/02/power-approval-set-to-delay-oracles-wisconsin-ai-datacenter/5300832) reported research firm Aterio's base case for Oracle's Wisconsin campus: partial grid power in December 2027 and full supply in October 2028, against Oracle's target of delivering customer capacity in the second half of 2027. A three-senator report published October 9 found that data-center operators dispute paying for shared grid upgrades, [per Senator Warren's office](https://www.warren.senate.gov/newsroom/press-releases/ai-data-center-companies-reveal-to-warren-blumenthal-van-hollen-they-are-not-paying-their-full-costs-will-continue-using-ndas-and-seeking-tax-breaks/), which leaves the question of who carries the power-side risk to state regulators.

A delay of this kind does not reduce demand. It moves revenue later and leaves interest costs where they are, which is the specific shape of risk a construction lender cannot syndicate away.

# The price of money moved first

The fourth layer is the market that prices all of the above, and it moved before equity analysts changed their demand forecasts.

The Federal Reserve noticed first, in official language. The minutes of its September 15-16 meeting, released October 7, listed "competition for capital from heavy private debt issuance to finance the development of artificial intelligence" among the factors lifting term premiums and Treasury yields. Nominal yields rose about 35 basis points across the 2- to 10-year segment between meetings, and the committee raised its target range to 3.75% to 4%. The open-market desk reported that "yield spreads on hyperscaler debt used to finance AI infrastructure remained wide," [per the minutes](https://www.federalreserve.gov/monetarypolicy/fomcminutes20260916.htm). When a central bank's minutes name a single industry's borrowing as a source of upward pressure on the whole yield curve, that industry's cost of capital has stopped being only its own business.

Equity buyers then declined a price. Firmus, an Nvidia-backed operator of AI data centers, shelved a $5 billion Australian IPO on October 9, [per Reuters](https://www.nst.com.my/amp/business/corporate/2026/10/1551088/nvidia-backed-firmus-scraps-us5bil-australian-ipo-poor-demand). The offer valued the company at $30.6 billion, nearly triple the $10.5 billion it reached in early August, with two operating sites and analyst-estimated debt of about $30 billion. "They were asking for a very big price tag for what would likely be expected to happen in the future assuming near flawless execution," said Joseph Koh of Blackwattle Investment Partners, who looked at the deal and did not bid. Ten Cap's Jun Bei Liu called it "an important reality check for the AI investment boom," adding that she would not read it as "the beginning of the end."

Five days later the bond market repriced the most debt-dependent name in the sector. CoreWeave's unsecured notes due 2032 traded below 80 cents on the dollar on October 14, and its shares closed down about 15%, [Reuters reported](https://www.thestar.com.my/tech/tech-news/2026/10/15/coreweave-ceo-notes-strong-backlog-in-memo-to-reassure-staff-over-market-scrutiny), citing Bloomberg. Chief executive Mike Intrator told employees that the market can "misunderstand and act irrationally" and that the backlog, $104.2 billion at the end of the second quarter plus more than $25 billion of new third-quarter commitments, more than covers the debt. He also wrote that problems "in the attempted IPOs and debt financings of some of CoreWeave's rivals" were "bleeding into the narrative around CoreWeave." That sentence describes contagion inside a financing layer, written by the company affected by it.

> The big AI reversal is finally here?
>
> 1) Kospi down another 7%
>
> 2) SK Hynix and Samsung down 7-10%
>
> 3) CoreWeave Inc. bonds fall to levels that usually indicate "distress"
>
> 4) Most chip and AI related stocks are down 4-10%
>
> — [@_anujsinghal](https://x.com/_anujsinghal/status/2110550486944448656), Oct 15, 2026

![Panel of five credit gauges: SoftBank's credit default swap and bond coupon, CoreWeave's 2032 bonds, the Firmus IPO and Treasury yields](/post-images/2026-10-16-the-ai-money-wobbles/credit-gauges.jpg)

The IMF's downside case puts a size on where this channel leads. A steep fall in AI investment could cut global output by 1.1% and US output by 1.8% by 2028, the Fund said, [per Reuters](https://www.kitco.com/news/off-the-wire/2026-10-13/imf-flags-risks-global-economy-ai-market-correction-war-and-debt). Deputy research director Petya Koeva Brooks added that even that scenario was not likely to produce a global recession. That is a correction, not a crash, and the Fund called it a risk, not a forecast.

# The strongest case that nothing broke

A thesis about financing strain owes the opposite case its best form. It has five parts, and each holds up against the evidence above.

First, the demand numbers are audited and accelerating. Nvidia guided up, Micron guided up and Oracle's backlog grew $26 billion in a quarter. Second, the OpenAI restatement was a definitional correction, not a miss, and its chief financial officer said the fourth quarter was "continuing at pace." Third, syndication is routine. The loans in question have not defaulted, and chollida1's point that the debt is trading well is not contradicted by any verified price on the Oracle-leased loans. Fourth, the Fed minutes also recorded that spreads on hyperscaler corporate bonds "decreased slightly over the period," so the investment-grade borrowers at the top of the stack are still financing cheaply. Fifth, the skeptic record is poor, and Dan Luu's audit documents it call by call.

The counter to the counter is narrow. Every one of those points concerns the strongest borrowers or the demand side. The strain this window recorded sits in the weaker names and the newer structures. On the borrower side that means neoclouds and a junk-rated investor; on the instrument side, special-purpose vehicles, construction-stage paper and IPOs priced ahead of revenue. Those are the parts of the stack built to carry demand forecasts forward in time, and they were the parts that repriced. A sector can have rising demand and a financing problem at once if the money arrives before the revenue does. That was the condition of fiber and telecom in 2000, though the gap in balance-sheet quality between those borrowers and today's hyperscalers is large, and the analogy weakens as it moves up the stack.

| Layer | What it converts | October reading | Direction |
|---|---|---|---|
| Measured demand | Chips and memory into revenue | Nvidia $96.2B, Micron $54.2B, Oracle backlog $664B | Holding |
| Revenue claims | One month into a year | OpenAI $68B gross to $50B net | Repriced |
| Supplier finance | Seller's balance sheet into buyer's capacity | Nvidia stakes $51.2B; Broadcom SPV over $60B | Growing |
| Investor leverage | Borrowed money into lab equity | SoftBank 9.75% coupon; swaps above 400bp | Costlier |
| Construction debt | Bank loans into buildings | Syndication slow; Jupiter force majeure | Strained |
| Credit pricing | All of the above into a rate | CoreWeave below 80 cents; Firmus pulled | Stressed |

![Two-ledger schematic separating the demand layer from the financing layer, with dated checkpoints from October 2026 to 2028](/post-images/2026-10-16-the-ai-money-wobbles/two-ledgers.jpg)

# What settles it before year-end

The next checkpoints arrive quickly and most produce verifiable numbers rather than narratives. Oracle's investor day, which ERP Today placed in October, is expected to give the first detailed account of how the $664 billion converts, with management currently expecting around half to become revenue within 36 months, per ERP Today. Nvidia's third-quarter report, which the Polymarket contract expects by November 18, will land against the $108 billion guidance and the odds above. Anthropic is expected to pursue its IPO as soon as November, per the Guardian, which would put the first audited frontier-lab accounts in front of public investors. OpenAI's reported $30 billion raise at about $1.4 trillion would test whether private equity buyers accept the $50 billion figure at the valuation built on the $68 billion one. None of the reports reviewed here shows the Hubbard campus debt fully placed, and the Jupiter pipeline's new date is February 1.

Each of those events will be read in two ledgers. The first records whether customers keep paying for compute, and through mid-October it shows yes, at record scale. The second records whether lenders keep accepting the instruments that pay for compute before customers do, and in this window that ledger recorded its first broad markdown. The difference between the two is the time between construction and revenue, and money borrowed against that time now carries a coupon of 9.75% at SoftBank and trades below 80 cents on the dollar at CoreWeave.

# A one-legged economy, priced by lenders

Georgieva's image works because it locates the risk correctly. The AI leg is not hollow: the chips, memory and backlogs in this window are measured, audited and growing. The wobble is at the joint between that leg and the rest of the body, where a run-rate becomes a valuation, a valuation becomes collateral, collateral becomes a construction loan and the construction loan waits for a gas pipeline. Every one of those conversions held in a rising market. In October each was asked to show its terms, and several did so at a higher price than the build-out was planned around. The demand case was not disproved by any of this. The build-out will be paid for at whatever rate lenders set, and in October that rate started to rise.

## Sources

- [CNBC: Nvidia, Oracle, CoreWeave and other AI stocks sink on OpenAI revenue report (Oct 8, 2026)](https://www.cnbc.com/2026/10/08/open-ai-revenue-nvidia-oracle-coreweave.html)
- [Reuters via The Star: OpenAI's annualized revenue $20 billion less than previously signaled (Oct 9, 2026)](https://www.thestar.com.my/tech/tech-news/2026/10/09/openai039s-annualized-revenue-20-billion-less-than-previously-signaled-ft-reports)
- [The Guardian: OpenAI projected to bring in $20bn less in revenue than expected (Oct 9, 2026)](https://www.theguardian.com/technology/2026/oct/09/openai-forecast-revenue-gap)
- [NVIDIA: Financial results for second quarter fiscal 2027, Form 8-K exhibit (Aug 26, 2026)](https://www.sec.gov/Archives/edgar/data/0001045810/000104581026000073/q2fy27pr.htm)
- [Micron: Record fiscal fourth-quarter and full-year 2026 results, Form 8-K exhibit (Sep 30, 2026)](https://www.sec.gov/Archives/edgar/data/0000723125/000072312526000018/a2026q4ex991-pressrelease.htm)
- [ERP Today: Oracle's $664 billion backlog comes with a new financing model (Sep 11, 2026)](https://erp.today/oracle-q1-fy27-results-664b-backlog-ai-contracts/)
- [The Next Web: Oracle cites force majeure over New Mexico data centre, Bloomberg reports (Sep 24, 2026)](https://thenextweb.com/news/oracle-force-majeure-project-jupiter-new-mexico)
- [The Register: Power approval set to delay Oracle's Wisconsin AI datacenter (Oct 2, 2026)](https://www.theregister.com/on-prem/2026/10/02/power-approval-set-to-delay-oracles-wisconsin-ai-datacenter/5300832)
- [Silicon UK: Banks look to offload AI data centre debt, citing the Financial Times (May 5, 2026)](https://www.silicon.co.uk/cloud/datacenter/bank-data-centre-debt-629721)
- [EnterpriseAM: Morgan Stanley-led consortium prepares $15bn bond sale, citing the Financial Times (Aug 6, 2026)](https://enterpriseam.com/egypt/2026/08/06/morgan-stanley-led-bank-consortium-prepares-usd-15-bn-bond-sale-to-dump-pre-built-ai-construction-risk/)
- [Startup Fortune: Banks are now trying to offload the AI data center debt they just underwrote (Oct 11, 2026)](https://startupfortune.com/banks-are-now-trying-to-offload-the-ai-data-center-debt-they-just-underwrote/)
- [The Next Web: Broadcom seeks more than $60bn in debt to fund AI chips for Anthropic (Aug 21, 2026)](https://thenextweb.com/news/broadcom-60bn-ai-chip-debt-anthropic)
- [Bloomberg and Reuters via The Star: SoftBank raises $11.1bn in world's biggest high-yield corporate bond sale (Sep 24, 2026)](https://www.thestar.com.my/business/business-news/2026/09/24/softbank-raises-us111bil-in-world039s-biggest-high-yield-corporate-bond-sale)
- [Latham & Watkins: SoftBank Group's record-breaking $11.1 billion bond offering (Sep 2026)](https://www.lw.com/en/news/2026/09/latham-advises-on-softbank-group-record-breaking-bond-offering)
- [Reuters via New Straits Times: Nvidia-backed Firmus scraps $5bn Australian IPO on poor demand (Oct 9, 2026)](https://www.nst.com.my/amp/business/corporate/2026/10/1551088/nvidia-backed-firmus-scraps-us5bil-australian-ipo-poor-demand)
- [Reuters via The Star: CoreWeave CEO notes strong backlog in memo to reassure staff (Oct 14, 2026)](https://www.thestar.com.my/tech/tech-news/2026/10/15/coreweave-ceo-notes-strong-backlog-in-memo-to-reassure-staff-over-market-scrutiny)
- [Federal Reserve: Minutes of the Federal Open Market Committee, September 15-16, 2026](https://www.federalreserve.gov/monetarypolicy/fomcminutes20260916.htm)
- [Reuters via Kitco: IMF flags risks to global economy from AI market correction, war and debt (Oct 13, 2026)](https://www.kitco.com/news/off-the-wire/2026-10-13/imf-flags-risks-global-economy-ai-market-correction-war-and-debt)
- [diginomica: What happens to the world economy if the US AI bubble bursts? (Oct 14, 2026)](https://diginomica.com/node/31298)
- [The Next Web: Broadcom CEO Hock Tan likens many AI investors to casino gamblers (Oct 13, 2026)](https://thenextweb.com/news/broadcom-hock-tan-ai-investors-casino-capital)
- [Dan Luu: How accurate have Ed Zitron's AI skeptic predictions been?](https://danluu.com/zitron/)
- [Simon Willison on the FT: Anthropic's best AI model struggles to attract users (Aug 23, 2026)](https://simonwillison.net/2026/Aug/23/anthropics-best-ai-model-struggles-to-attract-users-as-cheaper-t/)
- [Polymarket: Will NVIDIA Q3 Data Center revenue be above __?](https://polymarket.com/event/will-nvidia-nvda-q3-data-center-revenue-be-above-20261008120135000)
- [Fireside Alpha (@firesidealpha): OpenAI CFO Sarah Friar on $50bn versus $70bn (Oct 14, 2026)](https://x.com/firesidealpha/status/2110410713558905044)
- [Anuj Singhal (@_anujsinghal): The big AI reversal is finally here? (Oct 15, 2026)](https://x.com/_anujsinghal/status/2110550486944448656)

---

Canonical: https://www.thedeepfeed.ai/posts/2026-10-16-the-ai-money-wobbles/
Site: https://www.thedeepfeed.ai
Full corpus: https://www.thedeepfeed.ai/llms-full.txt