CoreWeave's credit scare explained: the AI bubble is really three bubbles, and only one is cracking
CoreWeave lost a fifth of its value in hours while its default insurance passed 10 percent a year. Separate the technology, the valuations and the debt, and the risk sits in one specific place: one floor below the hyperscalers.
On 14 October CoreWeave shares fell 15.5 percent, then another 6 percent after hours, with no company news to explain it. The next day the cost of insuring its debt for five years passed 1,000 basis points, a level usually reserved for distressed borrowers. Yet CoreWeave's revenue grew 112 percent last quarter and its backlog stood at $104 billion, with Microsoft, Meta and OpenAI as its main clients.
A video from the French channel Grand Angle uses that paradox to argue that "is AI a bubble?" is the wrong question. There are three possible bubbles stacked on top of each other: the technology, the valuations and the debt. Each has a different diagnosis, and only by separating them can you see where the risk actually sits.
We checked the video's figures against filings, rating agencies, the IMF, the BIS and market data. The framework holds up well. Several numbers did not: Oracle is still investment grade, CoreWeave's flagship loan was rated A3 rather than Ba3, and OpenAI's target valuation is $1.4 trillion, not $140 billion. Here is the corrected picture, layer by layer.
Technology, valuations, debt: what the evidence says about each layer
1. What actually happened to CoreWeave on 14 October
There was no earnings release or company announcement. Reuters tied the drop to a sell-off in CoreWeave's own bonds, whose 2032 notes had touched a 15 percent yield the day before, and to a hotter-than-expected September inflation print that revived expectations of another Fed hike. Oracle, Nebius and Iren fell sharply the same day. CoreWeave's five-year CDS crossed 1,000 basis points on 15 October and was reported at an intraday record of about 1,195, up from roughly 450 in June.
2. The capacity test: no sign of an overbuild yet
The classic tech-bubble symptom is capacity built for demand that never comes, which ends in collapsing rental prices, as happened to fibre after 2000. Today's test points the other way. CBRE puts vacancy in primary North American data-centre markets at a record-tying 1.4 percent in H1 2026, with 80.4 percent of capacity under construction already preleased. The Silicon Data H100 rental index rose from $2.00 an hour in December 2025 to a reported $2.88 on 24 October. The video's "$2.20 in 2022" is a dating error: H100s rented for several times that in 2023.
3. Cloud backlogs really did more than double
At 30 June, Microsoft reported $678 billion of commercial remaining performance obligations (up 84 percent), Google Cloud $514 billion and AWS $496 billion. That is about $1.69 trillion combined, roughly two and a half times the level a year earlier, so the video's "nearly $1.7 trillion" is accurate. Third-quarter figures arrive in the last days of October and are the next test of whether those commitments keep growing.
4. Adoption is broad, but measured payoff is narrow
Apollo's Torsten Slok found that 69 percent of S&P 500 companies said on Q2 calls that they had deployed AI, 29 percent cited a quantified result and only about 2 percent disclose a metric tracked over time. McKinsey's State of AI survey found 88 percent of firms use AI somewhere, while 39 percent report any EBIT impact, most of it under 5 percent. Usage is also concentrated: OpenAI's August Enterprise Signals data shows its top 10 percent of business customers use 8.3 times more output tokens per active user than a typical firm, up from 2.6 times in January.
5. Listed valuations are tight, not delusional, on one thermometer
FactSet puts the S&P 500 at 18.7 times forward earnings on 23 October, close to its 10-year average. Morgan Stanley says the premium of the seven largest tech stocks over the other 493 shrank to about 10 percent in July, the lowest in more than a decade, and Nvidia traded near 18.7 times forward earnings. Cisco peaked above 100 times in 2000. The video also cites tech-sector price, earnings and multiple changes for January to August that we could not source, so we have left them out.
6. On the other thermometer, prices are near dot-com levels
Robert Shiller's CAPE, which divides price by ten years of inflation-adjusted earnings, reached 41.2 in August 2026, the highest since 2000 and well over double its long-run average; the record is 44.2 in December 1999. The speaker argues CAPE misreads an economy whose cycles have accelerated. That is his opinion, and a contested one: the same argument was made in 1999. Both thermometers are accurate; they answer different questions about how long today's earnings will last.
7. The private market is where prices run ahead of profits
OpenAI is reported to be seeking new money at a $1.4 trillion pre-money valuation, after raising $122 billion at $852 billion in March. The Financial Times put its annualised revenue near $50 billion at the end of September, which makes the video's 28-times-revenue multiple correct once the garbled valuation is fixed. The FT also reported an internal projection of about $278 billion of negative free cash flow from 2026 to 2030. There is no earnings multiple to quote, because there are no earnings yet.
8. Earnings are linked in a chain, and the accounting flatters it
Nvidia, Broadcom and SK Hynix's record profits are largely Microsoft, Meta and Amazon's capital spending: one company's revenue is another's cheque. Server useful lives make reported profits look stronger too. Microsoft and Alphabet depreciate servers over six years and Meta over five and a half. Amazon is the exception: it extended to six years by 2024, then cut a subset back to five in 2025, which the video's "extended to 5 or 6 years" glosses over. Longer lives are defensible while older GPUs still rent at high prices.
9. The debt staircase: from bonds to off-balance-sheet vehicles
Alphabet, Amazon, Meta, Microsoft and Oracle are expected to spend about $800 billion on capex this year, roughly double 2025, and Bank of America projects their combined free cash flow at minus $64 billion in 2026 and minus $144 billion in 2027. Funding moves down a staircase. Investment-grade issuance reached roughly $225 billion in H1 by one S&P count; Amazon set a euro corporate record and Alphabet sold the largest yen bond by a foreign issuer (a bond, not a loan). Meta's $27 billion Hyperion campus sits in a joint venture 80 percent owned by Blue Owl funds, which its auditor flagged as a critical audit matter, and the FT tallied up to $300 billion of big-tech AI guarantees.
10. Oracle was downgraded, but not to junk
The video says S&P cut Oracle to BB+ in July, "the last notch before speculative grade". Both halves are wrong. On 9 July S&P cut Oracle from BBB to BBB-, which is the last investment-grade notch; BB+ would already be junk. S&P cited a projected free-cash-flow deficit of about $42 billion in fiscal 2027 and estimated that about half of Oracle's roughly $638 billion of signed contracts is linked to OpenAI. Moody's rates Oracle Baa2 with a negative outlook.
11. The CoreWeave paradox: lenders price the contract, not the company
In March CoreWeave closed an $8.5 billion delayed-draw loan that Moody's rated A3, investment grade, not Ba3 as the video says. Bloomberg reported it is backed by CoreWeave's $14.2 billion contract with Meta. By 15 October CoreWeave's unsecured 2032 notes traded near 78 cents with a yield above 16 percent. That is the key point, and the video gets it right: where Meta's signature stands behind the debt, risk is priced as low; where only CoreWeave stands behind it, risk is priced as severe. The video's "9 percent yield this summer" starting point we could not confirm; it matches the bonds' coupons.
12. Concentration and circular financing
Spreading risk only works if borrowers do not all depend on the same bet. On 17 August Nvidia agreed to guarantee up to $105 billion of lease and power payments, plus a floor on site value, so OpenAI could lease SB Energy's Ohio campus for 20 years, with Nvidia as exclusive chip supplier. Nvidia also invested about $30 billion in OpenAI's 2026 round, as reported. The BIS found that 16.1 percent of AI-to-AI investments since 2021 also involve a supply relationship, but they account for 46.4 percent of the value. The FT reported in May that JPMorgan, Morgan Stanley, SMBC and MUFG were exploring ways to offload data-centre loan exposure.
13. Timing is the risk nobody controls
The Fed raised rates by 25 basis points in September to 3.75 to 4 percent, and the 10-year Treasury yield hit an intraday 24-year high of about 5.36 percent on 7 October before easing to about 5.00 percent by 23 October. Markets priced roughly an 80 percent chance of a December hike, not the "9 in 10" in the video. Fixed-rate borrowers like Meta and Alphabet barely notice. Floating-rate leveraged loans do, though that market is still small: $20 billion is a Morgan Stanley forecast for 2026, and CLO exposure was about $4.5 billion.
14. Why it matters beyond tech investors
ING estimates that AI investment accounts for about a third of US GDP growth this year; the figure is ING's, not the IMF's. The IMF's October downside scenario, a sharp drop in AI investment and asset prices, would leave US output 1.8 percent and global output 1.1 percent below baseline by 2028. The exposure is broad: this debt sits in bond indices, insurers' portfolios and target-date retirement funds, which held about $5.3 trillion at the end of June according to Sway Research.
The corrected verdict matches the video's own. The technology layer shows no overbuild: data centres are full, GPU rents are rising and backlogs have more than doubled. Listed valuations are defensible if earnings keep coming, a big if when one company's profit is another's capex. The private market and the debt below the hyperscalers are where promises outrun cash, and where the risk has been sliced, rated and distributed.
Railways and the internet both won while many of the people who financed them lost money. AI does not need to fail for the same thing to happen; its revenue only needs to arrive more slowly than interest payments and refinancing dates. Signals to watch: the spread between CoreWeave's secured and unsecured debt, the late-October cloud backlogs, the December Fed meeting and big-tech CDS. This article is analysis, not investment advice.
Where we track this
- Is the AI boom a bubble? →
Our earlier look at what is real, what is financed and where the buildout becomes fragile.
- AI's debt boom is competing for capital →
How 5 percent Treasuries and hyperscaler bond issuance change the competition for investors' money.
- Why cheaper AI can still raise the total bill →
The demand mechanism behind full data centres and rising GPU rental prices.
- OpenAI's 722 AI math manuscripts →
The research release the video points to as evidence of how fast the technology is moving.
Questions people ask
- Why did CoreWeave stock drop in October 2026?
- CoreWeave fell 15.5 percent on 14 October 2026 and about 6 percent more after hours without company news. Reporting tied the move to a sell-off in its own bonds, whose yields passed 15 percent, a hotter-than-expected US inflation print that raised the odds of another Fed hike, and a broad sell-off in AI infrastructure stocks.
- Is AI a bubble in 2026?
- It depends on the layer. Demand for the technology is real: data-centre vacancy is about 1.4 percent and cloud backlogs have more than doubled. Listed valuations are high but not extreme on forward earnings, while the Shiller CAPE is near dot-com levels. The clearest stress is in the debt carried by neoclouds, joint ventures and lenders below the largest tech companies.
- Was Oracle downgraded to junk?
- No. On 9 July 2026 S&P cut Oracle from BBB to BBB-, the lowest investment-grade rating, citing heavy AI spending and a projected free-cash-flow deficit. A further one-notch cut would take it to speculative grade.
- What is circular financing in AI?
- It describes deals where the same companies are investors in, lenders to and suppliers of one another, such as a chipmaker investing in or guaranteeing a customer that then buys its chips. The BIS found that such deals were 16 percent of AI-to-AI investments since 2021 by count but 46 percent by value. The sales are real, but they no longer cleanly measure end demand.
- How could an AI correction affect people who do not own tech stocks?
- AI-related debt sits in bond indices, insurance portfolios, pension funds and target-date retirement funds. The IMF's downside scenario also estimates US output 1.8 percent below baseline by 2028 if AI investment and asset prices fall sharply, because AI construction is now a large share of US growth.
Written 25 October 2026 from an owner-supplied English transcript of the Grand Angle video "Le premier domino de l'éclatement de la bulle IA est tombé" (youtube.com/watch?v=wQaJmQWfb9U). CoreWeave's 14 and 15 October share and CDS moves come from Reuters reporting; the 1,194.5 basis-point intraday CDS record is ICE data as reported. Revenue growth and the $104.2 billion backlog are from CoreWeave's 11 August Q2 2026 release. The A3 Moody's rating on the $8.5 billion loan was reported by DatacenterDynamics and others; the Meta backing was reported by Bloomberg, not stated in CoreWeave's filing. Bond prices and yields are Tradeweb data via Reuters. CBRE's H1 2026 North America Data Center Trends report supplies vacancy and preleasing; H100 rental prices are Silicon Data index readings, the October figure as reported. Cloud backlogs come from Microsoft's SEC filings and the Alphabet and Amazon Q2 2026 earnings releases. Adoption figures: Apollo's Torsten Slok via Forbes (11 September 2026), McKinsey's State of AI 2025 and OpenAI's August 2026 Enterprise Signals. Valuations: FactSet Earnings Insight (23 October), Morgan Stanley via press coverage, Shiller's CAPE data. OpenAI's valuation, run-rate and cash-burn projection are FT and other press reports, not company disclosures. Depreciation lives are from company 10-Ks. Capex and free-cash-flow projections are Morgan Stanley and Bank of America estimates as reported. The Oracle downgrade, its free-cash-flow figure and the OpenAI share of its contracts come from S&P's 9 July action as reported. Hyperion's structure and audit treatment are from Meta's FY2025 10-K as reported by the WSJ; the $300 billion guarantee figure is an FT tally. The Nvidia, OpenAI and SB Energy agreement is from the 17 August announcement as reported by Reuters and Axios. Circularity figures are from BIS Bulletin No 137 (1 October 2026); the IMF scenario is from its 13 October World Economic Outlook and Global Financial Stability Report coverage. Rates: the Federal Reserve's September decision, Treasury yield data and futures-implied odds as reported. The one-third-of-growth estimate is ING's, via the WSJ; target-date assets are from Sway Research. Corrections to the source: Oracle was cut to BBB-, not BB+, and remains investment grade; CoreWeave's loan was rated A3, not Ba3; its backlog was $104 billion, not $99 billion; OpenAI's reported target valuation is $1.4 trillion, not $140 billion; the $2.20 H100 price dates from early 2026, not 2022; the AI share of S&P 500 gains since ChatGPT (about three quarters) is a J.P. Morgan estimate, not Morgan Stanley's; Alphabet's record yen deal was a bond, not a loan; December hike odds were about 80 percent, not 90 percent; the FT bank story dates from May 2026; the $20 billion leveraged-loan figure is a forecast. Not independently established and therefore omitted or labelled: the tech sector's January to August price, earnings and multiple changes, the 37 percent hedge-fund tech exposure, the $161 billion and $276 billion generative AI revenue figures, the 20 percent hyperscaler return on capital, the 9 percent summer yield on CoreWeave bonds and the quote about AI being the only pro-cyclical impulse left. This is editorial analysis, not investment advice; TaskNorth has no commercial relationship with the companies, institutions or creator cited.
Trying to work out which AI tools fit your task? Describe the outcome and get a Blueprint: the tools, the prompt, and the steps, with pricing we verified ourselves.