Positions have collapsed, but the story continues: Understanding the underlying logic of this round's AI stock crash

Original source: Yanwai Zhiyi

The U.S. added 172,000 non-farm payroll jobs in May, more than double market expectations. This should have been a reassuring report—the economy hasn't stalled, and employment is still expanding. But its impact landed more like a trigger.

On Friday, June 5, the Nasdaq Composite closed down 4.18%, its worst single-day performance since April 2025; the Philadelphia Semiconductor Index plummeted 10.26% in a single day, erasing over $1 trillion in market value from the chip sector—the bleakest day since the circuit breakers in March 2020. Three days later, at Monday's open, South Korea's KOSPI plunged over 8%, breaking below 7,500 points and triggering a circuit breaker that halted trading for 20 minutes. China's A-share market opened lower and extended losses, with the Shanghai Composite Index dipping below 4,000 points intraday. Japan's Nikkei and Taiwan's stock market each fell nearly 4%.

An employment figure that wasn't even excessively hot capsized tech stocks from New York to Seoul within 48 hours. The severity of the move doesn't match the trigger, and that's the real puzzle worth pondering about this sell-off.

The real issue isn't those 172,000 jobs, but the positions it hit

Blaming this plunge entirely on rate hike expectations doesn't hold up.

After the non-farm payrolls report, the 10-year Treasury yield indeed jumped to 4.5%, and the 2-year yield rose to 4.17%, hitting a high since February 2025. Market expectations for a rate cut this year were completely erased, with futures even starting to price in a rate hike. Rising interest rates disproportionately hit high-valuation growth stocks—that logic is sound. However, on the same day, the Dow Jones Industrial Average only fell 1.35%, the small-cap Russell 2000 index even turned positive intraday, and nearly half of the S&P 500's components were still rising. If it were a systematic repricing of all assets due to rising rates, the market wouldn't behave this way.

The selling pressure was highly concentrated, focused on the stocks that had risen the most and were most crowded over the past two months.

Just how crowded these stocks were was flagged by Wall Street analysts well before the crash. Data from Goldman Sachs showed that global hedge funds' net leverage ratio surged from below 70% to over 80% in less than two months, approaching the 85th percentile of the past five years. Net allocation to the information technology sector increased by 853 basis points in a single quarter, the largest quarterly increase on record. The semiconductor industry alone accounted for 19% of global hedge funds' total exposure, an all-time high, and this proportion had more than doubled since the start of 2026.

David Chew, a strategist at Citigroup, put it bluntly three days before the crash: bullish bets on the Nasdaq 100 had been stretched to extreme levels, "significantly increasing the probability of profit-taking and long-position unwinding upon any negative catalyst." Citigroup's famous "Bear Market Checklist" triggered 11.5 out of 18 indicators on June 5, the highest reading since the 2008 financial crisis.

When a sector is being bet on by global capital in the same direction and with the same leverage, it ceases to be a diversified investment and becomes a trade. In any trade, someone has to press the sell button first.

The non-farm payrolls report was the reason to press that button.

This crowding is self-reinforcing. In recent years, trend-following quant funds, risk-parity strategies, and a huge volume of zero-day-to-expiry options have become dominant forces in the U.S. stock market. Their common trait is to follow the trend. Once prices break below key levels and volatility spikes, these models mechanically reduce positions, and the selling itself pushes prices lower and volatility higher, feeding into the next round of selling. Humans decide *whether* to sell; machines decide *how fast*. On June 5, the VIX volatility index surged about 34% in a single day, climbing back above 20—a clear signal that this positive feedback loop had been activated. As for the exact volume of programmatic selling that day, the market doesn't have public data; this is a structural judgment call rather than a precise amount that needs to be calculated.

Watch how individual stocks fall, and you'll see this is deleveraging

Those who rose the most are falling the hardest this time.

May was a celebratory month for chip stocks. The Philadelphia Semiconductor Index gained over 60% year-to-date at one point, rising on 22 of the last 23 trading days. Micron was up as much as 154% this year, SanDisk gained nearly fivefold, and AMD surged 40% in May alone, hitting an all-time high. When money floods in one direction, valuation becomes an afterthought.

On June 5, the bill came due. Marvell, which had surged about 25% in a single day after Jensen Huang's comment about a "potential trillion-dollar company," gave back over 16%, leading the chip stock decline. Micron fell about 13%, while AMD and Intel each dropped around 11%. These were all names at the top of the recent gainers list.

Nvidia, on the other hand, only fell about 6% that day. Its market cap dropped below $5 trillion, losing nearly $280 billion in a single day—sounds scary, but it was relatively resilient amidst the carnage. The general held the line; the ones that collapsed were the soldiers charging ahead with the heaviest leverage. This type of decline, where the leader holds up while the periphery crumbles, is a classic sign of deleveraging, not a fundamental collapse. If AI demand were truly the problem, Nvidia would be the first to be questioned.

The direct trigger was Broadcom. On June 3, it provided guidance for Q3 AI chip sales of $16 billion, below market expectations of $17.2 billion. More critically, it did not raise its full-year AI chip target. In a market accustomed to "guidance hikes every quarter," "maintaining the status quo" was interpreted as bearish. A string already pulled extremely taut was given a single, gentle pluck, and it snapped.

Why South Korea triggered the first circuit breaker

The selling pressure cascaded down through the most heavily held positions, with South Korea as the first stop.

At Monday's Seoul open, Samsung Electronics and SK Hynix both fell about 10% intraday. Together, these two companies account for nearly half of the KOSPI's total market capitalization—such is the extreme concentration of South Korea's stock market bet on AI. The KOSPI was one of the best-performing markets globally this year. What goes up the most naturally has the most room to fall. When global capital needs to raise cash and reallocate, heavily held, liquid Korean tech stocks become the most convenient "ATM." Traders at BNY Mellon and Lucerne Capital Management used similar language: these are the most heavily held positions globally, making them the first to be sold to generate liquidity.

South Korea's domestic leverage acted as an amplifier. As of June 4, retail investors' margin debt in South Korea was still at an all-time high of 37.74 trillion won. The Korean won weakened to near 1,560 against the U.S. dollar on Monday, signaling accelerating foreign capital outflows. Some brokerages suspended margin trading due to exhausted credit limits. South Korea's finance minister, central bank, and financial regulator issued a joint statement that day, pledging to intervene in the foreign exchange market.

Continuing down this chain are Taiwan's stock market, home to TSMC, and China's A-share computing power chain. Their distance from the epicenter varies, and the logic behind how they are traded differs.

Which assets to watch this round and how

Asset reactions have a sequence; don't lump them all together.

The first and most sensitive is HBM, the epicenter. High Bandwidth Memory is the current hardest bottleneck for AI computing power. SK Hynix holds over half the global market share, followed by Samsung. These two are the primary targets of this trading round. How they move essentially defines the upper limit of sentiment for the entire chain. Watching their order books is more accurate than watching any AI concept index.

The next circle out includes the high-conviction U.S. chip stocks in the Philadelphia Semiconductor Index—Micron, Broadcom, Marvell. Their reaction is primary; their decline is determined by the progress of position liquidation, not by fundamentals. Further out, TSMC is the choke point for process technology and advanced packaging, a quasi-primary node in the chain.

China's A-share optical modules, CPO, PCB, and server stocks are part of the chain's spread. They have a real performance correlation with global AI capital expenditure, and orders do follow Nvidia and Broadcom's deployment. However, expectations constitute a higher proportion of their pricing, making them more elastic. A sneeze at the epicenter can easily turn into a cold here. As for a batch of high-flying thematic stocks and crypto assets that rose broadly on the AI narrative, they are more followers of sentiment and leverage, lacking direct fundamental support. Caution is most warranted at these levels. In contrast, the pricing of A-share computing power chain is more influenced by local capital and policy expectations and may not perfectly synchronize with U.S. stocks. But as long as the global AI narrative remains the anchor for pricing, it's hard for this chain to stay entirely outside this liquidation process.

Across assets, the decline in gold and Bitcoin follows the same logic: real interest rates are rising, putting pressure on non-yielding assets. On June 5, gold dropped over $100 in a single day, falling below $4,370 and wiping out all its year-to-date gains. Silver weakened in tandem, and Bitcoin briefly dipped below $60,000. The U.S. dollar index strengthened. A strong dollar combined with rising real interest rates creates consistent pressure on non-yielding assets. By Monday, Bitcoin had rebounded above $63,000, with risk sentiment recovering ahead of the stock market—a notable signal.

If net leverage begins to fall, the VIX peaks and declines, and HBM and computing power order prices remain tight, then this round is likely a violent but phased liquidation. The pullback might even create entry opportunities. But if the CPI on June 10 exceeds expectations, combined with the FOMC meeting (June 16-17) dot plot shifting towards rate hikes, the nature changes. A systemic upward shift in the interest rate floor would require a complete rewrite of the valuation logic for high-valuation AI assets. Deleveraging would just be the opening act; in this scenario, one must acknowledge it and pivot to a more cautious stance. Another contrarian signal to watch: if core stocks like Nvidia or SK Hynix start to lower guidance or signal loosening capital expenditure, that would be the moment when fundamentals truly deteriorate.

What crashed was the crowded trade, not the AI story

In the same week the Philadelphia Semiconductor Index lost a trillion dollars in value in a single day, TSMC Chairman C.C. Wei stated at a shareholder meeting that global chip supply will be "unable to meet AI demand for several years," and demand for advanced process nodes in 2026 will exceed supply by 25% to 30%. Jensen Huang said in Seoul on June 7 that memory shortages "will last for several years." Nvidia and SK Hynix just announced a multi-year memory collaboration, with SK Hynix's entire 2026 HBM production capacity already sold out. SK Group Chairman Chey Tae-won even extended the timeline for the shortage to 2030. DRAM prices rose about 90% quarter-over-quarter in Q1 this year.

On the demand side, there is no sign of weakening. What crashed was not the AI story itself, but the way it was being bet on.

Jensen Huang had only one sentence in response to the Friday crash, telling reporters in Seoul: "We are just getting started. No matter what happens in the stock market, you should be happy because now you can buy the dip." Of course, he has a vested interest, but his point highlights a core divergence: if you believe AI is infrastructure akin to the internet, then a position shakeout is a window to buy. If you suspect the returns on this wave of computing power investment will ultimately fail to materialize, then Friday was the beginning of a crack.

Some reject the term "stampede" entirely. In their view, Friday's market action looked more like a rotation: money didn't leave the market; it just changed seats, moving out of crowded semiconductors and into banks, industrials, and value stocks. The Russell 2000's contrarian rally is circumstantial evidence. According to this interpretation, it's a healthy rebalancing disguised in scary packaging. Goldman Sachs strategist Mueller-Glissmann also said a bit of consolidation "isn't necessarily a bad thing" and that speculative leverage and options positions needed to be worked off. This viewpoint reminds us not to read a deleveraging event as an apocalypse. However, it equally fails to answer a key question: before positions are truly cleared and volatility truly peaks, no one can guarantee there won't be a second wave.

Neither assessment can be definitively confirmed yet. What's certain for now is that the crowded trade global capital built up over the past two months is being forcefully unwound. This process is violent, cascading, and not yet finished. The fact that South Korea was still triggering circuit breakers on Monday is itself proof that the liquidation is far from over.

For the coming week, watching three things is sufficient: the CPI data on June 10, the first FOMC meeting chaired by [Federal Reserve Chair Nominee] [注: 此处 '沃什' 按理应为 '鲍威尔' (Powell) 或其他主席，但原文是沃什，按规则保留其名] from June 16-17 and its dot plot, and whether the South Korean stock market can stabilize. These will tell you whether this is ultimately a healthy deleveraging or the opening act of a longer correction.

This article is source from: Yanwai Zhiyi