Did the $6 Trillion Stock Sell-Off Burst the AI Bubble?

Did the $6 Trillion Stock Sell-Off Burst the AI Bubble?

US stocks lost about $6 trillion in 45 days. A widely read analysis calls it a correction, not a burst AI bubble. We checked each claim against filings and market data.

ZharfAI ResearchNovember 21, 202623 minutesSource: The Kobeissi Letter

Has the AI bubble burst? That is the question The Kobeissi Letter, a subscription market newsletter with a large following on X, set out to answer on November 20, 2026, in a long post titled "-$6 Trillion In 45 Days: Is This A 'Bubble' Or An Opportunity?" Its answer is no. It describes the fall in US stocks since early October as a normal correction, driven by six pressures at once, and argues that so much bad news is now priced in that the bigger risk is missing a rebound. The newsletter says it has been "highly bullish on AI" since November 2022, and its conclusion matches that view.

X counted more than eight million views of the post within a day, and the question it asks matters well beyond Wall Street: the price of AI services, the pace of data-center building and the money available to AI companies all depend on how this sell-off ends. This article follows the post section by section. For each claim we went to the primary record: company filings, Nvidia's and Anthropic's own releases, FactSet's earnings reports, the Federal Reserve, US energy data and Cboe's volatility history. We report what checks out, what does not, and what the post leaves out.

Status at the US market close on Friday, November 20, 2026: the S&P 500 closed at 6,910.39, 11.6% below its October 6 record of 7,818.93, which is what traders call a correction. The Nasdaq Composite closed 18.1% below its record. The newsletter went up 13 minutes after the opening bell on Friday and uses Thursday's closing figures, when the S&P 500 was 8.2% below its record. Friday was worse: the S&P 500 fell 3.7% and the Nasdaq 5.3%.

How we checked: we opened every source listed at the end of this article. Where a figure could only be found in search snippets or social posts, we say so or leave it out. The four charts are ZharfAI's own, drawn from public data: the companies' SEC filings, the US Energy Information Administration and Cboe. The newsletter's charts are not reproduced. The cover is a 2017 photograph by Arild Vågen of the New York Stock Exchange seen from the steps of Federal Hall, not a photograph of this week's events; it is cropped and resized, and remains licensed under CC BY-SA 4.0. This is news analysis, not investment advice.

The argument in brief

The post explains the sell-off with six forces that arrived together:

  1. Spending. Amazon, Alphabet, Meta and Microsoft are spending record sums on AI infrastructure, and investors worry about their cash flow in 2027 and 2028.
  2. Politics. Democrats won the November 3 midterm elections. In the newsletter's view this ended the "Trump put," the belief that the president would step in to support a falling market, and raised fears of harsh AI regulation.
  3. Interest rates. The Federal Reserve, under Chair Kevin Warsh, is raising rates while inflation stays above its 2% target, which makes the AI build-out more expensive to finance.
  4. Nvidia and AMD. Nvidia lost its chief executive, Jensen Huang, and AMD was hit by a short seller's report about China.
  5. Sentiment and cash. Fear rose sharply, and Anthropic's record IPO pulled cash out of the market just as hedge funds were selling technology shares.
  6. Valuations. After the fall, technology shares are cheaper relative to expected earnings than at any time since March 2020.

Its conclusion: there is no bubble, the market is in a "normal correction," even another 10% fall would be "healthy," and quality technology companies now offer an attractive "risk-reward setup." It is sceptical of the calls, which it says are spreading, for a fall of 30% or more.

Forty-five days, in dates

The S&P 500 last closed at a record on October 6. What followed:

  • October 28. The Fed held rates steady, but two officials dissented in favour of a hike.
  • November 3. US midterm elections; Democrats won control of both the House and the Senate, according to AP race calls.
  • November 5. The VIX, the market's main gauge of expected volatility, jumped 70% in one day.
  • November 12. Nvidia announced that Jensen Huang had resigned, effective immediately.
  • November 17. Nvidia reported its quarterly results.
  • November 18. Anthropic priced its IPO at $83 a share.
  • November 19. Anthropic's shares began trading on Nasdaq; a short seller's report on AMD came out the same day.
  • November 20. The S&P 500 fell 3.7%, Nvidia 14% and AMD more than 13%.

The headline figure, $6 trillion lost in 45 days, is plausible, but we could not find who measured it. Forty-five is the correct count of calendar days from October 6 to November 20. The newsletter's own chart ends at Thursday's close, so the figure does not include Friday's fall. The nearest sourced number we found was $5.7 trillion lost through November 5, from an index of total US market value cited by The Edge Malaysia from Bloomberg. The S&P 500 closed at 7,353.73 that day and at 7,323.63 on November 18, nearly the same level, so a loss of about $6 trillion by November 18 is consistent with it. The post also says US stocks lost "an average of -$225 billion per day, for 31-straight trading days." There were 31 trading days from October 7 to November 18, but 31 days at $225 billion comes to about $7 trillion, not $6 trillion, and the market rose on several of those days, so "straight" cannot mean 31 days of losses.

AI spending is real, and so is the cash squeeze

The claim: combined capital spending at Amazon, Alphabet, Meta and Microsoft rose by $36.9 billion in the third quarter of 2026, to a record $201.9 billion. It has more than doubled in a year and is up 356% since early 2024. "Big Tech has now spent $1.6 trillion on CapEx over the last 3 years."

What the filings show: the quarterly numbers are right. We rebuilt the series from each company's reports to the SEC, counting cash purchases of property and equipment. The third-quarter total is $201.9 billion, $36.9 billion more than the quarter before, 108% more than a year earlier and 356% more than the first quarter of 2024.

Quarterly capital spending at Amazon, Alphabet, Microsoft and Meta

Combined cash purchases of property and equipment, mostly data centers and servers, in billions of US dollars; Q1 2024 to Q3 2026

Capital spending, billions of US dollars
QuarterAmazonAlphabetMicrosoftMetaTotal
Q1 2024$14.9B$12.0B$11.0B$6.4B$44.3B
Q2 2024$17.6B$13.2B$13.9B$8.2B$52.9B
Q3 2024$22.6B$13.1B$14.9B$8.3B$58.9B
Q4 2024$27.8B$14.3B$15.8B$14.4B$72.3B
Q1 2025$25.0B$17.2B$16.8B$12.9B$71.9B
Q2 2025$32.2B$22.5B$17.1B$16.5B$88.2B
Q3 2025$35.1B$24.0B$19.4B$18.8B$97.3B
Q4 2025$39.5B$27.9B$29.9B$21.4B$118.6B
Q1 2026$44.2B$35.7B$30.9B$19.0B$129.8B
Q2 2026$54.2B$44.9B$35.8B$30.1B$165.1B
Q3 2026$77.3B$51.4B$39.4B$33.9B$201.9B

Data: each company’s 10-Q and 10-K filings with the SEC. Some quarters, including every fourth quarter, are derived from year-to-date totals. Microsoft’s fiscal quarters are aligned to calendar quarters and exclude finance leases. Chart: ZharfAI.

The three-year total is not right, at least for these four companies. From the fourth quarter of 2023 to the third quarter of 2026, their cash capital spending adds up to about $1.14 trillion. Adding Meta's and Microsoft's leased data centers brings it to roughly $1.2 trillion. The newsletter's own chart covers the same four companies, so it does not explain the gap to $1.6 trillion.

The more important figure is free cash flow: the cash a business has left after paying for its investments. The newsletter's chart shows it turning negative for the four companies combined, and the post calls "a brief period of negative free cash flow" the "short-term cost of a long-term buildout." The company figures are starker than "brief" suggests. In the third quarter, three of the four spent more on investment than their operations brought in.

Free cash flow in the third quarter of 2026

Cash from operations minus capital spending, billions of US dollars. The four combined: about −$26.1B

* Amazon reports free cash flow only for the trailing twelve months (−$42.4 billion); this quarter’s figure is ZharfAI’s calculation from Amazon’s cash-flow statement, using Amazon’s own definition (purchases of property and equipment net of proceeds and incentives). Others: Alphabet’s and Meta’s Q3 releases and Microsoft’s CFO on its earnings call. Chart: ZharfAI.

Amazon's free cash flow for the past twelve months was minus $42.4 billion, against plus $14.8 billion a year earlier, and it still plans about $220 billion of cash capital spending in 2026. Meta raised its 2026 spending forecast to $135 to $145 billion. Microsoft is the exception: it stayed positive and says it expects to remain so this fiscal year.

When spending exceeds cash flow, the difference is borrowed or drawn from savings. Through November 4, US investment-grade companies had sold $1.88 trillion of bonds in 2026, more than in the record year of 2020, and the biggest cloud companies accounted for $370.7 billion of that, according to LSEG data reported by Reuters. The European Central Bank warned on November 18 that this is how losses in AI could spread from shares to bond markets; we covered that warning in our November 19 news brief.

Earnings. The post says S&P 500 companies have reported earnings growth of 31.9% for the third quarter, "the strongest YoY growth rate in at least 4 years." The 31.9% was FactSet's figure on November 13; FactSet's November 20 report puts it at 32.9%. But it is not the strongest in four years. FactSet calls it the third quarter in a row above 25%, and the newsletter's own FactSet chart shows the previous quarter at 52.2%. We found no source for its other claim, that earnings have beaten estimates for 13 consecutive quarters.

What the post leaves out is where the weakness is. In aggregate, earnings came in only 3.5% above analysts' estimates, the smallest margin since late 2022. The seven largest technology companies, which FactSet calls the "Magnificent 7," missed their estimates by 4.9% in aggregate and grew earnings 27.0%, slower than the other 493 companies in the index for the first time since late 2022. The companies doing most of the AI spending are now the soft spot of the earnings season, not the strong one.

The midterms, the "Trump put" and an AI ban

The claim: Democrats "swept" the midterms, ending the "Trump put," and "calls for an AI ban have gained traction." Markets have priced in "aggressive regulation and potential bans on AI" when the new Congress takes over in January 2027.

What checks out: Democrats did win both chambers, according to AP race calls reported by NPR. The "Trump put" is the newsletter's own reading of market psychology. It can be argued but not checked.

What does not: the "AI ban" the post points to is a November 4 post by Representative Yassamin Ansari, which the newsletter shows as an image. She called for "a pause on advanced AI development, regulation of this out of control industry, and a ban on artificial superintelligence." That is a call to pause frontier development and ban superintelligence, not to ban AI. The newsletter then argues, reasonably, that "the US banning AI will not happen," but that rebuts something nobody in its evidence proposed. It offers no measure showing that markets have priced in an AI ban.

The Fed, bond yields and oil

The claim: "the Fed is hiking interest rates aggressively," inflation has been above the Fed's 2% target "for 67-straight months," and oil has fallen about 22% this month as the Iran war "nears its 300th day."

Rates: "aggressively" overstates it. Kevin Warsh became Fed chair on May 22, 2026. Under him, the Fed has raised rates once this year, by a quarter of a percentage point on September 16, to a range of 3.75% to 4.00%. On October 28 it held, with two members dissenting in favour of another hike. By November 12, futures traders put the chance of a hike at the next meeting, on December 8 and 9, at about 91%, Reuters reported. The inflation streak is accurate: by the Fed's preferred measure, the PCE price index, prices have risen faster than 2% a year in every month from March 2021 to September 2026, which is 67 months. September's rate was 3.6%.

What matters more for AI companies than the Fed's overnight rate is what they pay to borrow for years, and the post barely mentions it. The 10-year Treasury yield was 5.04% on November 19 and has been at or above 5% since November 12. It was below 4.7% in late August.

The post also says investment-grade bond sales are expected to reach a November record of $175 billion. Its chart credits that figure to BofA Global Research. It is a forecast, and we could not open the BofA research to check it.

Oil: the fall is real. By the US Energy Information Administration's daily data, Brent crude fell from $120.15 a barrel on October 30 to $92.94 on November 17, the latest published day, a drop of about 23%. The US benchmark, WTI, fell about 17% over the same days.

Brent and WTI crude oil prices

Daily spot price, US dollars a barrel; September 1 to November 17, 2026, the latest published day

Data: US Energy Information Administration, via the St. Louis Fed’s FRED (public domain). Chart: ZharfAI.

The war count is off. By Al Jazeera's account of its first four weeks, the war began on February 28, 2026, so November 20 was its 266th day, not close to its 300th. As of mid-November, Al Jazeera and Reuters reported, full-scale fighting was on hold without a settlement, and talks were continuing through Oman and Qatar. The newsletter suggests that cheaper oil could let the Fed ease in 2027; that is a forecast, and the Fed's October statement still said "inflation remains elevated."

Nvidia, AMD and China

The claim: Nvidia "has lost Jensen Huang," its stock is down 40% in weeks, and its forward price-to-earnings ratio is down to 9, even though it guides about $750 billion in revenue for fiscal 2028. AMD "is now being accused of smuggling chips into China."

Nvidia: on November 12, Nvidia announced that Huang had resigned as president, chief executive and director, effective immediately, without giving a reason. Chief financial officer Colette Kress became interim CEO while remaining CFO, and Harvey C. Jones became the company's first chair of the board. The board had opened an inquiry into reports about Huang's personal conduct days earlier, AP reported. On the November 17 earnings call, Jones said the board had "unanimously accepted his resignation."

The stock fall is, if anything, understated. Nvidia closed at a record $239.24 on October 6. Thursday's close of $145.98 was 39% lower; Friday's $125.53 was 47.5% lower. CNBC put its forward price-to-earnings ratio at 8.4 after Friday's fall.

The $750 billion is not guidance. Nvidia's results release gives no figure for fiscal 2028. On the call, Kress raised the company's expected fiscal 2028 growth to 75%, from 70%. Applied to this fiscal year's three reported quarters plus the fourth-quarter forecast, about $424 billion, that implies roughly $740 billion. That is our arithmetic, not Nvidia's number.

The post leaves out how strong the latest quarter was, and one new risk. Revenue for the quarter that ended October 25 was $115.8 billion, up 103% from a year earlier, of which $107.6 billion came from data centers. Nvidia forecast about $130 billion for the next quarter, assuming no data-center sales in China. Then on November 20, Reuters reported that the SEC "is exploring potential accounting issues at Nvidia," according to CNBC. What exactly the SEC is looking at, and whether anything comes of it, is not known. The low price-to-earnings ratio reflects doubt about future earnings, which is the newsletter's own point: "the market does not believe the E."

AMD: the post misdescribes the report. On November 19, Hunterbrook Media, whose sister fund Hunterbrook Capital disclosed a short position in AMD, alleged that IEIT Systems, a Chinese server maker placed on the US Entity List in March 2025, kept receiving newer AMD firmware, and that near-identical firmware appears at affiliated companies that are not listed. It also traced $67 million of AMD's 2025 licensing income to two Chinese joint ventures blacklisted in 2019. The report does not allege smuggling, and says its findings "don't establish on their own a violation of export laws." AMD told CNBC the claims "misrepresent AMD's business practices and relationships" and that it complies with US export controls. Its shares fell 7.7% that day and more than 13% the next.

Valuations: the claim that technology shares are the cheapest since the pandemic relies on a Bloomberg chart of the S&P 500 technology index, and FactSet's figures agree. FactSet puts the sector's forward price-to-earnings ratio at 16.3, against a five-year average of 25.4, and the whole S&P 500's at 17.1, against 19.8. But part of the "cheapness" comes from rising estimates, not only falling prices. Between September 30 and November 19, the S&P 500 fell 6.2% while analysts' earnings forecasts for the next twelve months rose 4.4%. Those forecasts assume the AI spending continues. If it slows, the earnings that make these shares look cheap shrink too.

Fear, the Anthropic IPO and the rush for cash

Volatility: the post says the VIX is up more than 100% from its October low and "posted its biggest daily gain since February 2018 on November 5th." The first part is true: it closed at 14.84 on October 9, at 28.44 on Thursday and at 35.89 on Friday. The second part is not. On November 5 it rose 69.7%. By Cboe's daily history, that is the third-largest one-day rise since 1990. December 18, 2024, at 74.0%, was larger.

The US stock market’s “fear gauge” (VIX)

Daily close, October 1 to November 20, 2026. Higher means options traders expect bigger swings over the next month

Data: Cboe Global Markets, VIX daily history. Chart: ZharfAI.

The IPO: "Anthropic's IPO just raised over $100 billion in the biggest IPO in history." It was the largest IPO on record, but "over $100 billion" needs a footnote. Anthropic priced its shares at $83 on November 18, above its $70 to $78 range. Its closing announcement gives about $93.5 billion of gross proceeds from the public offering, including the underwriters' extra shares, of which $91.6 billion went to the company. Separately, Anthropic sold $26.0 billion of shares at the same price in a private placement to AMD, Broadcom, Google, Micron, Microsoft, Nvidia and TSMC. Only the two together, $119.5 billion, exceed $100 billion. The previous record was SpaceX's June 2026 IPO at $85.7 billion, according to CNBC.

The post cites $400 billion of institutional orders and $100 billion of retail orders. Those were non-binding indications of interest, reported by CNBC from unnamed sources, and the chart the newsletter itself uses labels them "nonbinding." The shares opened at $111.50 on November 19, closed at $97.17 and fell another 8.5% on Friday. The post does not mention that Anthropic lost $104 billion in the first nine months of 2026, as Reuters reported from the IPO filing. Nor does it mention who bought the $26 billion placement: chip and cloud companies, several of which also sell Anthropic the computing power it runs on. Money flowing both ways between AI companies and their suppliers is one of the financing patterns central banks have begun to flag.

The newsletter argues the IPO drained cash from the market. That may have added to selling on November 18 and 19, since buyers had to sell other holdings or use cash. It cannot explain a slide that began in October. The $93.5 billion raised is about 1.5% of the $6 trillion of lost market value.

Hedge funds and everything else: the post says hedge funds posted "their largest trailing 6-week net outflow in over a decade." Goldman Sachs prime brokerage data, reported by several outlets, showed hedge funds selling technology shares at their fastest pace since the data began in 2016, so "over a decade" here means the whole record. Other assets fell too. Gold was 31.5% below its January record by November 16, and bitcoin fell below $60,000 on November 20, from about $81,000 in late October.

Correction or bubble?

The newsletter is right about several things. The AI build-out shows up in real money: Nvidia's sales doubled in a year, S&P 500 earnings are growing about 33%, and the companies doing the spending are among the most profitable in the world. It is also right that some fears are overdone. An outright ban on AI is not on the table, and oil prices have fallen sharply.

But the case for "no bubble" is weaker than the post makes it look, for five reasons:

  • The weak spot is AI itself. Three of the four biggest spenders had negative free cash flow last quarter, the largest technology companies missed earnings estimates, and more of the build-out is paid for with debt.
  • "Cheap" depends on the spending continuing. Low price-to-earnings ratios rest on forecasts that assume the investment wave keeps going. Nvidia's earnings are other companies' capital spending.
  • There is no "Fed put" either. Nobody should count on the central bank cutting rates to rescue the market: the Fed is expected to raise rates again in December, and 10-year borrowing costs are above 5%.
  • The company shocks are not resolved. Nvidia's leadership, the reported SEC review and the AMD allegations are all open questions.
  • The post was overtaken within hours. Friday's fall took the S&P 500 past the 10% correction line and the Nasdaq close to 20%.

A bubble, in the usual sense, means prices that no plausible future cash flow can justify. The data so far does not settle the question either way. Demand for AI is real and growing fast. Whether $200 billion a quarter of spending will earn its cost back is not yet visible in the buyers' cash flow. The newsletter itself expects no clarity on that before the first half of 2027. "Correction" and "bubble" are both forecasts; this sell-off is the market arguing about which one is right.

For companies that buy AI services rather than shares, the practical meaning is narrower. Prices of models and cloud capacity are set by providers that now face investors demanding returns, and one of them, Anthropic, is newly public. Expect pricing and terms to move. Measure the cost of each task you automate, and keep a second supplier you could switch to. For more on how fast AI capability is moving underneath all this, see our analysis of AI research automation, and on the hardware side, Nvidia's Rubin Ultra memory trade-off.

What to watch next

  • December 8–9: the Fed's last meeting of the year, and whether it raises rates again.
  • Late January to February 2027: the next quarterly reports from Amazon, Alphabet, Meta and Microsoft, and their spending plans for 2027.
  • Nvidia: the board's choice of a permanent chief executive and any official word on the SEC's review.
  • AMD: any response from the US Commerce Department to the Hunterbrook allegations.
  • Anthropic: its first quarterly report as a public company.
  • Oil and bond yields: the talks over the war, and whether the 10-year yield stays above 5%.

Claim by claim

The newsletter's claimWhat we found
$6 trillion lost in 45 daysPlausible; no primary source named. 45 calendar days is correct. Excludes Friday's fall.
$225 billion a day for 31 straight trading daysDoes not add up: 31 × $225 billion ≈ $7 trillion, and the market rose on several of those days.
Big Tech capex $201.9 billion in Q3, +$36.9 billion, +356% since Q1 2024Correct, from SEC filings.
$1.6 trillion of capex in three yearsNot reproduced: about $1.14 trillion for the four named companies.
Q3 earnings growth 31.9%, strongest in four years32.9% in FactSet's latest report; the previous quarter was higher (52.2%).
Markets priced in AI bansThe cited post calls for a pause and a ban on superintelligence, not on AI.
The Fed is hiking aggressivelyOne quarter-point hike in 2026; a December hike is expected.
67 straight months of inflation above 2%Correct (PCE, March 2021 to September 2026).
Oil down about 22% this monthCorrect for Brent (−23% to November 17); WTI −17%.
Iran war nearing its 300th dayDay 266 on November 20.
Nvidia down 40%; forward P/E about 9Correct on Thursday; −47.5% and 8.4 after Friday.
Nvidia guides about $750 billion for fiscal 2028Not a dollar guidance; implied by a 75% growth outlook (≈ $740 billion).
AMD accused of smuggling chips into ChinaThe report alleges firmware access by a blacklisted firm and China revenue; it does not allege smuggling.
VIX's biggest daily gain since February 2018Third-largest since 1990; December 2024 was larger.
Anthropic raised over $100 billion, biggest IPO everBiggest ever at $93.5 billion; over $100 billion only with a $26 billion private placement.

Frequently asked questions

Has the AI bubble burst?

Not by any agreed definition. US stocks are in a correction, and AI shares have fallen hardest, but earnings are still growing fast. What remains unproven is whether the record spending on AI will pay for itself.

Why did Nvidia's stock fall so much?

Several things at once: its chief executive resigned on November 12, investors grew worried that the AI spending that drives its sales will slow, and on November 20 Reuters reported that the SEC is exploring potential accounting issues at the company. Its latest quarterly results were strong.

How much did Anthropic's IPO raise?

About $93.5 billion from the public offering, the largest on record, plus $26.0 billion from a private placement to seven chip and technology companies: $119.5 billion in all.

Is this article investment advice?

No. It checks the facts behind one widely shared analysis. The newsletter's view that the risk-reward is attractive is its own; we make no recommendation.

Source notes

Sources checked on November 21, 2026. Market levels are closing prices for November 20, 2026. Where we could read a news agency's story only in a syndicated copy, we link the copy.

TagsAI BubbleStock Market Sell-OffNvidiaAnthropic IPOBig Tech CapexFederal Reserve

Related Posts

Name one process for a discovery call

If this note maps to a real system in your organisation, start with the services page or a shipped case study.