AI Trends

Global chip stocks have pulled back nearly 40% from their peak; where does the AI bubble debate stand now?

TMTPost · AGIUpdated Nov 15 16:18

AI summary

The Philadelphia Semiconductor Index has pulled back nearly 40% from its peak of 14,634 points on June 22, and Nvidia has lost nearly $1.4 trillion in market value since its high in early October, fueling the AI bubble debate. The article identifies tiers of decline: Goldman Sachs' AI memory stock basket is down 51% from its June high, CoreWeave has pulled back about 70% from its high, while cloud giants with cash flow have been relatively resilient. In the first nine months of this year, capital expenditure by Microsoft, Google, Amazon, Meta, and Oracle totaled $556.9 billion, already exceeding their operating cash flow of $542.9 billion over the same period; the gap was filled by bond issuance, and Oracle's 5-year CDS once rose from about 58 basis points to 295 basis points.

Why read this

AI score 63

The article uses three sets of data—chip stock pullbacks, cloud vendors' free cash flow, and credit spreads—to break down the mismatch between AI capital expenditure and returns.

Topics

Tags

Body · AI translation

After the U.S. market close on November 17, Eastern Time, Nvidia will release its financial results for the third quarter of fiscal 2027. The earnings call will no longer be hosted by Jensen Huang, who has led the company for 33 years, but by interim Chief Executive Officer and Chief Financial Officer Colette Kress. On November 12, Nvidia announced that Jensen Huang had resigned as president, chief executive officer and board member, effective immediately, and that veteran director Harvey Jones would serve as the company's first-ever board chair. On the day the news was announced, Nvidia fell more than 10% in premarket trading at one point, closed down 4.34%, and shed nearly $182 billion in market value.

The signals from this earnings call will go beyond the results themselves: whether the enormous capital expenditure that the giants have poured into AI can yield returns is a debate that has lasted for more than a month, and Nvidia sits at the center of that debate. This earnings report is the first formal response since the debate erupted.

The debate was set off on November 5. At the close that day, Eastern Time, the Philadelphia Semiconductor Index plunged 12.25%, its largest single-day drop since March 2020. Nvidia fell 9.3%, shedding about $447 billion in market value in a single trading day. The triple-leveraged semiconductor ETF SOXL fell 38.7% in one day. The next day, Seoul's KOSPI index widened its decline to more than 8% within two minutes of the open, triggering a circuit breaker, and fell below 5,000 points during the session.

Extending the observation window to one month reveals even larger declines. Since its high in early October, Nvidia has shed nearly $1.4 trillion in market value, and in the two trading days after the apology letter was exposed it shed more than $420 billion more, leaving its market value at $4.07 trillion at the close on November 10. The Philadelphia Semiconductor Index has pulled back nearly 40% from its peak of 14,634 points on June 22. Oracle has fallen more than 30% this year. CoreWeave, regarded as a bellwether for AI infrastructure, had retraced about 70% from its June 2025 high as of November 10, before its earnings release.

But looking only at the broad market may lead to the opposite conclusion. The Nasdaq has pulled back about 12% from its high, just entering technical correction territory. In the week of the plunge, the equal-weighted S&P 500, which represents ordinary stocks, even set a record high.

The breakdown came three and a half months earlier than the index

The calm in the broad market is to some extent determined by its weighting structure: the top ten constituents of the S&P 500 account for nearly 40% of its weight. On November 5, Nvidia fell 9.3%, TSMC ADRs and Broadcom fell more than 10%, while small and mid-cap AI stocks such as memory names generally fell more than 20%. Multiplied by their respective weights, the S&P 500 fell only 4.36% that day.

The money did not leave U.S. stocks. In October, the equal-weighted S&P 500 outperformed the market-cap-weighted version by 5.7 percentage points, the largest monthly gap since April 2009. That month, AI-related stocks fell more than 13%, while non-AI sectors set new highs.

The timeline is more telling. The earliest warning came from South Korea. On June 19, the KOSPI hit a record 9,385 points intraday, then fell sharply, with a maximum drawdown of nearly 44% by the end of July. What was sold off were Samsung Electronics and SK Hynix, which account for half of the index, as the market doubted the returns on AI investment and worried about slowing HBM (high-bandwidth memory) demand. Almost simultaneously, on June 22, the Philadelphia Semiconductor Index set a record high of 14,634 points, and Goldman Sachs' AI memory stock basket also peaked in June.

The Nasdaq did not hit its high until October 6. From June to October, the upstream of the AI chain had already fallen from its highs; only giants such as Nvidia and Microsoft were still propping up the index. The election results on November 5 set off the selloff: the Dow fell 3.44%, the S&P 500 fell 4.36%, and the Nasdaq fell 5.52%. The cover provided by the heavyweights failed, and the broad market followed with a catch-up decline.

Within the chain, the declines are clearly stratified. Memory fell the most: Goldman Sachs' AI memory stock basket is down 51% from its June high, the largest drawdown for that basket since ChatGPT was launched.

Next are the new cloud providers that build data centers with high leverage. CoreWeave has retraced more than 70% from its June 2025 high and is down about 32% this year. Data center power facilities, computing chips and optical modules have also mostly fallen more than 20% from their highs, with the hardest-hit names down nearly 40%.

Only the cloud giants at the end of the chain held up. As of November 4, the day before the plunge, the Magnificent Seven were still up about 9.6% overall this year. Apple rose 5.31% that day, overtaking Nvidia in market value to return to the global top spot.

The more asset-heavy a segment is, the more it relies on financing, and the further it is from ultimate cash flow, the harder it has fallen; the giants with cash flow in hand have been the most resilient.

At 14 times earnings, Nvidia's decline is not unjustified

Nvidia shows the biggest contrast. In the fiscal quarter ended at the end of July, Nvidia's revenue grew 106% year on year to $96.2 billion, yet its share price fell back to the level at the start of the year. According to well-known investor Josh Brown, its price-to-earnings ratio is now only about 14 times. Many therefore call it cheap, but placing Nvidia back in the chain, its decline is not unjustified.

First, Nvidia's revenue is other companies' capital expenditure. In fiscal 2026, U.S. customers contributed about 69% of its revenue. The giants and new cloud providers issue debt to build data centers, and that money ultimately flows to Nvidia. When customers' financing costs soar, the market worries about whether they can still pay for the next round of orders. Nvidia itself is also investing downstream: in March this year it invested $4 billion in optical communications companies Coherent and Lumentum, and this kind of circular investment is precisely the focus of the bubble debate.

Second, a low P/E ratio does not necessarily mean cheap. For cyclical stocks, the lowest P/E ratio often comes exactly when profits are at their peak and about to top out. Josh Brown warns that, given Nvidia's current situation, if results come in just 5% below expectations, the stock could fall 40%.

Third, historical patterns do not favor the bulls. After reviewing the past four cycles, Bank of America found that declines in the chip index typically lead the downturn in semiconductor sales by 6 to 9 months. By that reckoning, the real earnings downgrades may not show up until the first quarter of 2027.

The latest escalation in the controversy came a week before the earnings report. On November 8, Jensen Huang sent an apology letter to all employees, admitting to infidelity toward his wife. According to reports, the woman with whom he had an intimate relationship had served as an executive at Megaspeed, a Singapore data center company that is an important Nvidia customer in Southeast Asia. On the same day, Nvidia's board announced the launch of an independent investigation, prompted by reports that Huang had become a target of "foreign intelligence activity." On November 10, the independent directors hired former federal prosecutor Jin Junxian and the law firm WilmerHale to conduct the investigation. On November 12, Huang resigned.

It should be noted that most of the $1.4 trillion in market value erased occurred before the controversy came to light. The cracks in the chain came from doubts about financing and demand; the controversy was simply the latest weight pressed on top.

The earnings report will therefore take on an extra layer of observational value. The market consensus expects third-quarter revenue of $109.01 billion, up 91.2% year on year, and net profit of $60.65 billion. The bigger suspense lies in the earnings call: Morningstar analysts are watching whether Nvidia maintains its forecast of 70% revenue growth for fiscal 2028, as well as its judgment that global AI spending will reach $3 trillion to $4 trillion by 2030. Whether the new management will tighten investment and financing support for AI customers also remains to be seen.

Orders are full, yet the stock still falls

There are two more sets of contrasts in the chain. One is that orders are full, yet the stock still falls. Lumentum's CEO said orders for data center optical components are sold out for years. SanDisk beat expectations for six consecutive quarters, yet its stock still plunged, down more than 60% from its high. CoreWeave's third-quarter earnings, released after the market close on November 10, also beat expectations, with revenue up 168% year on year and backlog rising to $132.5 billion; the stock rose 6% after the report but is still down about 32% for the year. On the other side, Liz Ann Sonders, chief investment strategist at Charles Schwab, citing Baird data, noted that data center projects canceled in the third quarter were nearly more than in 2024 and 2025 combined.

The other is how much they rose versus how much they retraced. Goldman Sachs's storage basket rose more than fourfold in the year before its peak. SanDisk was listed at $36 per share when spun off, peaked at $2,335, and after a 22% single-day plunge on November 5, it is still up nearly threefold this year. AMD fell more than 30% in a month but has still doubled this year; by November 9 it had already fully recovered all the ground lost on the day of the plunge.

Over the past two years, when looking at cloud giants' earnings, the market asked only one question: did capital expenditure go up? If it did, the stock rose. In the earnings season at the end of October, that logic failed. The four major cloud providers' combined 2026 capital expenditure guidance is about $720 billion to $745 billion, nearly 80% higher than about $410 billion in 2025. Except for Microsoft, which slightly lowered its guidance due to delays in third-party capacity delivery, the others maintained or raised their guidance.

The market's reaction, however, turned cold across the board. For Google's three earnings reports this year, the stock reaction went from a gain of nearly 10% to a drop of 7.1%, and then to a 12.6% plunge on October 29, the largest single-day drop since its listing, wiping out about $500 billion in one trading day.

In the first nine months of this year, capital expenditure at Microsoft, Google, Amazon, Meta and Oracle totaled $556.9 billion, exceeding their operating cash flow of $542.9 billion over the same period. Four of the Magnificent Seven saw free cash flow turn negative this year, the first time in a decade.

By company, Google's third-quarter capital expenditure rose 114% year on year to $51.4 billion, and free cash flow swung from positive $24.55 billion a year earlier to negative $9.07 billion. Amazon's free cash flow over the past 12 months was negative $42.36 billion. Meta's free cash flow was negative $1.157 billion. Microsoft was the only giant to keep positive free cash flow, but it still fell 27% year on year.

The gap was filled by bond issuance. According to S&P Global, large U.S. cloud service providers and related companies issued about $225 billion in bonds in the first nine months of this year, up 973.7% year on year. Microsoft issued $25 billion in bonds, its first in nine fiscal years. Meta announced it would issue up to $95 billion in additional shares.

The most aggressive borrower is Oracle. S&P downgraded its rating from BBB to BBB- in July, the lowest investment-grade tier. The company's total debt has been pushed to nearly $169 billion over two years, capital expenditure doubled to $55.7 billion in fiscal 2026 and is expected to rise another 65% to $91.9 billion this fiscal year. Market consensus expects its free cash flow to be negative $23.7 billion this fiscal year. Its $300 billion computing power contract with OpenAI underpins $553 billion in remaining performance obligations and also ties the company's fate to a customer that is not yet public.

The credit market's pricing is unambiguous. A year ago, Oracle's 5-year credit default swaps (CDS, which can be understood as default insurance) were priced at about 58 basis points; in mid-October they briefly rose to 295 basis points, on par with bonds of much lower-rated companies. On November 2, Oracle returned to the bond market, launching a $10 billion new bond issuance.

CoreWeave's CDS broke through 1,000 basis points in mid-October, and its bond yields briefly approached 15%. Riskier AI companies prefer convertible bond financing. According to Barclays statistics, convertible bonds fell 6.4% in October, the worst October since 2008.

The stock market is still debating whether there is a bubble, while the bond market is already pricing default risk.

Voters are voting

The election put the second layer of scrutiny on the table. In the November 3 midterm elections, the Democrats won both the Senate and the House. What made the market more nervous was the local level. According to Ballotpedia statistics, there were 41 data center-related local ballot measures across 10 states, and the results of 40 were unfavorable to data center construction. Voters passed 27 restrictive measures and rejected 13 proposals favorable to developers. The only exception was Harwood, North Dakota, where local voters rejected a proposal to add additional review for industrial projects.

The biggest variable is Texas. Democrat Gina Hinojosa defeated Governor Abbott, who was seeking re-election. During the campaign she promised to pause approvals of new data centers, and on November 6 she reiterated on CNN that data center construction must be paused before protective measures are introduced. NBC exit polls showed that 75% of Texas voters oppose building new data centers in their own communities. According to Wells Fargo statistics, Texas accounts for 27% of planned U.S. data center capacity. Goldman Sachs estimates that the Democratic victory puts 45% of planned U.S. data center expansion capacity at risk of a short-term pause.

On November 4, when the election results came out, power, storage and data center stocks were the first to fall, while the broader market was still rising; Apple rose 5.31% that day. The next day, Qualcomm issued weak guidance and warned about rising memory costs. The two forces combined, and the selloff spread to the entire chip sector, evolving into the scene at the beginning of this article.

Buffering factors also exist. The new Congress will not take office until January 2027, and any restrictive bill would have to get past the presidential veto; Trump has clearly sided with opposing regulation. UBS estimates that if data center construction is paused, about 30% of the capacity originally planned to be added in 2028 would be affected, which amounts to a slowdown rather than a halt. The market's panic may be running ahead of policy.

The giants are still increasing investment, and orders remain full. After the election, OpenAI reiterated that its target of $600 billion in computing spending before 2030 remains unchanged, and CoreWeave's backlog stood at $132.5 billion. AI demand has not broken. What has broken is the belief that "as long as you spend money, it will rise." AI is moving from the storytelling stage to the audit stage.

Whether the giveback turns into a collapse depends on the three readings to be delivered next: whether the free cash flow of the cloud giants can turn positive again, whether the credit spreads of Oracle and CoreWeave will continue to widen, and whether the data center restrictions brewing in Texas and other states will actually take effect.

After the market close on November 17, Nvidia's earnings report will provide an answer. Every phrase in interim CEO Cress's growth guidance will be magnified and scrutinized.

Source

TMTPost · AGI