On November 5, 2026, the Nasdaq Composite fell 5.52%, and the Philadelphia Semiconductor Index dropped 12.25%—its largest single-day decline since March 2020. The "Magnificent Seven" lost approximately $1 trillion in market value in a single day; S&P 500 constituents collectively shed about $3 trillion; and the VIX surged roughly 70% to 28.18. Overnight, "the AI bubble has burst" went from a research hypothesis to a headline verdict.

But treating a one-day decline as the end of a cycle is the easiest—and most error-prone—judgment to make. To be precise: if "burst" means a 2000-style simultaneous occurrence of valuation collapse, earnings falsification, and credit rupture, the evidence does not support it. A significant portion of the valuation bubble has indeed been squeezed out, but through a pattern of "earnings rising, multiples falling." The real risk has moved—from P/E ratios to cash flow statements and off-balance-sheet commitments, from the stock market to the credit market.

It's not that the bubble has burst; the bubble has changed form and pushed the reckoning to the 2027–2029 refinancing window.

01 Defining a Bubble "Burst"

Debates are often unproductive because the two sides are not using the same yardstick. A complete bubble burst requires three conditions to hold simultaneously: collapse of the valuation center, falsification of earnings expectations, and financing disruption triggering passive capacity reduction. In March 2000, Cisco's trailing P/E was 135x and its forward P/E about 200x; the Nasdaq subsequently fell about three-quarters and took roughly 15 years to return to its prior high. More importantly, most companies at that time had no profits, and telecom companies had laid "dark fiber" far exceeding demand—asset prices, corporate earnings, and credit conditions all broke down together.

Measured against this yardstick, only one condition has broken today—and not completely.

Valuations are not at extremes. Currently, the S&P 500 forward P/E is about 18.5x, and the Nasdaq 100 PE-TTM is 27.08x; the Philadelphia Semiconductor Index forward P/E has compressed from about 28x to about 16x, touching the valuation low of this AI cycle. Market concentration resembles 2000, but pricing multiples do not.

Earnings have not turned. In Q3 2026, S&P 500 constituent earnings grew over 30% year-over-year, though the beat margin was only 1.9%, far below the five-year average of 7.5%—"high growth, low surprise," not disappearing growth. In the first three quarters, the U.S. tech sector posted a price return of +22.1% and EPS growth of 55.7%, while P/E ratios actually contracted 21.6%.

02 A "Valuation Reset," Not "Demand Disappearance"

The shape of the 2026 U.S. AI market was "high-level oscillation + year-end valuation compression." As of November 6, the Nasdaq 100 was still up 6.89% year-to-date and the S&P 500 up 7.76%, but the Nasdaq 100 had retreated 13.56% over the past month, and the Nasdaq Composite stood 10.18% below its 52-week high; the S&P 500 hit an all-time high of 7,844.52 on October 6, then gave back most of its gains over two months.

More telling is the decline gradient. The Philadelphia Semiconductor Index had fallen over 38% from its June 22 peak, entering a technical bear market; upstream, Intel fell 46.25% over 20 days, Micron dropped 32.93%, and AMD fell 26.16%, while downstream Microsoft fell only 0.83% and Amazon 1.19%. The decline diminishes progressively along "chips—hardware—cloud providers," which is a pattern of capital expenditure expectations being repriced, not terminal demand evaporating. The trigger was specific: on November 3, Democrats won a landslide in the U.S. midterm elections, raising market concerns about tighter data center approvals and AI regulation—Goldman Sachs estimates this could affect about 45% of planned U.S. data center capacity; combined with the 10-year Treasury yield briefly hitting 5.34% and the 30-year at 5.70% (highest since 2002), discount rates pressed hardest on the longest-duration assets.

Valuation divergence also shows this is not indiscriminate rupture. At the November 6 close, Google's PE-TTM was 14.87x, Oracle 18.20x, Meta 20.82x, Nvidia 23.60x, and Microsoft 27.71x; the truly high multiples were Palantir at 112.94x and AMD at 97.41x. On the other end, Micron had more than doubled year-to-date with a PE of only 9.28x—cyclical earnings depressed the denominator, not sentiment inflating the numerator. Nvidia's market cap of $4.51 trillion, about 23% below its 52-week high, had a PEG clearly below 1, versus the roughly 3.1x PEG of tech stocks at the 2000 bubble peak—the magnitude difference is stark.

03 The Real Cracks: Cash Flow, Debt, and Off-Balance-Sheet Commitments

If there is a bubble in this cycle, it is no longer in P/E ratios but in cash flow statements.

The scissors gap between investment and cash flow. In Q3 2026, Amazon, Google, Microsoft, and Meta had combined cash capital expenditures of $200.54 billion, up 108.0% year-over-year, while combined operating cash flow was $175.50 billion, up only 10.4%; the four companies' combined free cash flow swung from +$7.84 billion in Q2 to -$25.04 billion, with Microsoft's +$18.63 billion the only positive, Meta about -$2.1 billion, Alphabet about -$9.1 billion, and Amazon about -$33.5 billion. That is, every $1 of operating cash required about $1.15 of spending, versus 0.77 twelve months earlier. JPMorgan estimates 2026 capex will be about 93% of hyperscalers' operating cash flow, versus only 33% in 2023.

Revenue remains strong, but payback periods are lengthening. During the same period, AWS revenue was $45.984 billion (+39.3%), Azure $33.165 billion (+48.2%), and Google Cloud $30.709 billion (+102.6%), totaling $109.86 billion, up 55.7% year-over-year, accelerating from Q2's 47.9%; the global cloud infrastructure services market reached $167.7 billion in Q3, up 55% year-over-year, accelerating for 12 consecutive quarters. The four companies, including Oracle, had about $2.5 trillion in undelivered orders. Demand is real; the problem is the steeper investment slope: the four companies' 2026 capex guidance totals about $720–750 billion, and the market expects growth to decelerate to about 40% in 2027 and about 10% in 2028. UBS estimates cumulative AI capital expenditure of about $4.1 trillion over 2026–2028, while IDC projects the global generative AI market will only approach $150 billion in 2027; in McKinsey's survey, nearly 80% of enterprises deploying AI have not achieved net profit improvement, and 95% of generative AI pilots have no measurable financial return. In Q3 2026, Google, Amazon, Meta, and Microsoft's ROIC were 3.43%, 2.82%, 2.82%, and 6.89% respectively, all declining quarter-over-quarter.

The financing structure has changed—this is the most fundamental difference from 2000. As of early November, the five major cloud providers had issued about $296 billion in bonds year-to-date, equivalent to 36% of 2026 capex; Morgan Stanley expects AI-related debt issuance of nearly $570 billion for the full year. Off-balance-sheet, there are over $1.1 trillion in uncommenced lease commitments and $1.9 trillion in other unconditional purchase commitments (about $685 billion maturing before end-2027), not counted in capex but approximating debt. According to the Bank for International Settlements, 55.2% of capital flowing into AI companies in 2021–2025 came from other AI companies—"demand" and "financing" are hard to separate. The credit market has already priced it: Nvidia's 5-year CDS rose from about 45 basis points in July to about 81–89 basis points in November, CoreWeave briefly hit about 1,289 basis points; S&P downgraded Meta's outlook from "stable" to "negative" on October 30.

The three-layer superposition explains why "cloud revenue broadly beat expectations" nonetheless resulted in Alphabet falling 12.64% in a single day, shedding about $500 billion in market value. The pricing function has changed: previously it valued investment scale, now it values the ability to convert investment into free cash flow. Microsoft was resilient due to positive FCF; the others were punished for FCF turning negative—the difference is not in growth rates but in the sign of cash flow.

04 China: Fell Earlier and Cheaper, But What's Cheap Is the "Shovel Sellers"

In the same global de-valuation cycle, Chinese assets fell first and deeper, but the pricing logic is different. As of November 9, the A-share AI sector (02003800.PT) was down 13.75% year-to-date, with PE-TTM of 36.83x at the 3.1st percentile of the past 5 years, PB of 2.32x (8.4th percentile); the Hang Seng Stock Connect AI Theme Index was down 52.33% year-to-date, 58.8% below its January 13 high; the Hang Seng Tech Index was down 26.14% year-to-date, with PE-TTM of 21.39x at the 24th percentile since 2010. From November 2–6, Hang Seng Tech rebounded 8.61% over five days, while the Nasdaq 100 fell 4.91% over the same period.

But "cheap" is easily misread.

First, the low PE percentile is driven by denominator expansion. CPO optical modules rose 26.78% year-to-date with 114% amplitude, PE-TTM of 27.54x at the 0.25th percentile of the past 5 years, yet PB at the 65th percentile; semiconductors had PE of 42.30x at the 19th percentile, PB at the 64.6th percentile, and PS at the 71st percentile. Earnings have indeed materialized: in H1 2026, Zhongji Innolight's revenue was 41.778 billion yuan (+182.49%), net profit attributable to shareholders 13.651 billion yuan (+241.70%); Eoptolink's net profit was 7.529 billion yuan (+90.98%); TFC Optical's net margin was 42.59%. This is the "low PE" of cyclical growth stocks at peak earnings—using it to argue valuation safety is dangerous.

Second, profits are concentrated among shovel sellers; the application end remains empty. Of the 1,075 constituents in the A-share AI sector, 407 are loss-making (38%); of the 39 in the Hong Kong AI theme, 27 are loss-making (69%), with negative PE (-89x), priceable only on PS, and the 7.81x PS sits at the 86.6th percentile of the past 5 years—after falling 52%, it remains one of the most expensive AI assets in the entire market on a revenue basis. The Huawei computing theme was -24.31% year-to-date, the Ascend theme -22.96%; the domestic substitution narrative has also undergone valuation digestion. The real bottleneck is not computing hardware but the lack of a replicable profit paradigm downstream.

Thus China and the U.S. form a mirror image: the U.S. risk is "too much money borrowed, too slow to pay back," while China's risk is "money earned, but only in a small segment at the very upstream of the supply chain." The November 5 Philadelphia Semiconductor crash instantly transmitted to A-share semiconductors and Hong Kong AI hardware, showing the pricing anchor remains in the overseas computing chain; meanwhile, in early November, Hong Kong internet platforms led gains while AI hardware lagged, showing capital beginning to pay for "low valuation + own cash flow." This is not decoupling; it is divergence.

05 "Not Burst" Does Not Equal "Won't Burst"—It May Just Be Delayed

The "alarm" has not been lifted:

One, interest rates have shifted from tailwind to headwind. On November 6, U.S. employment data (165,000 new jobs, 4.1% unemployment) beat expectations, and the market-implied probability of a 25 basis point rate hike in December rose to about 70%; the average coupon on U.S. corporate bonds is about 7.1%, and high-yield bonds near 11%. When expansion shifts from "cash-flow-driven" to "debt-issuance-driven," interest rates are no longer a valuation parameter but a condition for financing survival.

Two, the "second derivative" problem. The market does not need to wait for absolute capex to decline before repricing; as long as growth decelerates from 94% to 40% and then to 10%, upstream entities that expanded capacity based on high growth and relied on debt support will be the first to falter. Moody's identified 2027–2029 as the debt maturity peak, with about 17% of the 607 private credit borrowers it tracks using PIK (payment-in-kind). This is the ultimate judgment window for "whether the bubble bursts," not November 5.

More useful than guessing the top is watching four verifiable indicators: (1) Nvidia's FY2027Q3 earnings after the U.S. market close on November 17, with market expectations of revenue around $108.8–109.0 billion, EPS around $2.47, and gross margin holding at 74%, with management having already warned of a decline to 71%–72% in Q4; (2) whether the four major cloud providers' 2027 capex guidance shows a growth inflection point; (3) whether Nvidia's 5-year CDS breaches the 200 basis point warning line and whether leading cloud providers' FCF can recover; (4) Anthropic's IPO pricing (market discussion around $1.1–1.3 trillion), the first time the model layer is publicly priced by the market.

06 Conclusion: A Layered Answer, and What Remains After the Bubble

Viewed in layers, the answer is much clearer:

Hyperscalers and leading chipmakers: Not burst; it is valuation convergence. Orders, profits, and cloud growth have not systematically deteriorated, and P/E ratios have fallen to the teens to twenties;

AI semiconductor hardware: Already in a technical bear market, a valuation reset at peak earnings;

Highly leveraged computing rental and pure theme plays: Clearing is underway, and credit spreads are already a substantive alarm;

Primary market model layer: Not yet publicly priced; Anthropic's IPO is the first judgment point;

Chinese AI assets: Valuations cleared before earnings, PE percentiles extremely low but earnings quality uneven; hardware has delivered, application emptiness unchanged.

Using "speculative growth" rather than "bubble burst" to describe the current moment: if high valuations persist long enough, they will stimulate real capital accumulation in computing power, electricity, data centers, and other physical assets; even if valuations decline, the infrastructure already built remains in the economy. The dot-com bubble burst did not prevent the internet from reshaping commerce; the Nasdaq took about 15 years to return to its prior high, but the fiber laid down was eventually fully utilized. This cycle will likely be similar—valuations will disappear, but capacity will not. The difference is that in 2000, bad debt remained mainly with telecom companies and shareholders, whereas this cycle's debt is embedded in private credit, lease commitments, SPVs, and circular guarantees—longer transmission, more opaque.

The most precise way to put it might be this: the valuation bubble already burst once in October–November 2026, but the credit bubble has yet to be tested. The bubble hasn't truly popped—it has simply migrated from the stock market to the bond market, from the question of "is it overpriced?" to the question of "can it be repaid?"

- END -