Summary

Momentum-factor deleveraging is the main driver of this selloff, and it is not over. Goldman says it “could last at least another 2 weeks.” Deleveraging has pushed the market back to April levels, but gross-margin leverage remains above 3x, still far above year-end levels; “leverage generally doesn’t get fully flushed out.” Dip buyers emerge once prices get low enough, so the marginal damage is diminishing.
He starts by admitting he got it wrong. In early May, he said positioning did not support a further rally, yet the market kept rising in June—“I did get it wrong.” The mistake was reading gross margin above 3x as a standoff between longs and shorts; in reality, it represented quant positions long semiconductors and short other stocks. Momentum “mindlessly goes long the stocks that have gone up the most,” and nearly all of those were semiconductors: “you lift semis all at once, then dump them.”
Leverage has not been flushed out. Semiconductor leveraged ETFs fell from $157B to $94B; based on the stock-price declines, they should have sold $78B, but leverage actually fell only $64B, implying roughly $15B of dip-buying. Gross-margin leverage remains at the 95th percentile of the past 5 years, while hedge-fund leverage is at the 82nd percentile of the past 3 years—“even aggressive deleveraging still has room to run.”
The real landmine is credit, not the five companies themselves. Google’s 2026 capex consensus is $186B versus an internal estimate we learned is roughly $230B; the 4 hyperscalers plan to raise about $300B this year and $500B next year. Compute ROI remains very high—“Amazon would still raise money for capex with its CDS at 120bp”—but the risk is massive supply crowding out the entire IG pool: a rapid rise of tens of basis points in credit spreads “would be catastrophic for the real economy.”
That leads to a conclusion opposite the Street’s consensus. Kevin Warsh can likely run QT and probably will not do QE, but “he needs to let money flow out”—through deregulation, freeing banks to lend, or a guarantee vehicle; whether it comes from the White House or the Fed, “money ultimately has to be put out.” He thinks next Tuesday’s meeting is very unlikely to produce a hike: CPI is a lagging indicator, downside risks outweigh upside risks, and as a recent Fed arrival he needs to establish credibility and may sound relatively hawkish.
There are 2 signals for putting on a large position. One is a higher VIX—currently just 18 and at a historical low—while an index catch-up selloff clears the landmine ahead; semis have already taken their hit and may then fall only in line with the index. The other is momentum flipping to short semis and then getting blown up by strong earnings in reverse: the momentum index is already down 35 points, so “it cannot push another 35 points lower.” The rebound will not be V-shaped; history says 4-6 months, and “the real rally rotates, so there is less fear of missing it.”
Kimi is not a DeepSeek moment. DeepSeek was about cutting training costs and the possibility that less compute would be needed; Kimi is an open-source model, and the main race has already returned to inference infra. A 2.8T model divided by 4 still requires more than 700G; it needs 10 H100s, and 8 is not enough. “There is no way to deploy it on a local machine; it has to go back to the cloud.” Better open-source models are positive for compute investment—“AI keeps rolling forward”（AI还是滚滚向前）—provided the credit problem does not blow up, in which case “AI valuations are not high right now.”

Deep dive

1. Revisiting the May-June misread: June’s rally was momentum-driven, not a sign of abundant liquidity

He opens by owning the mistake. In early May, he said, “Based on the positioning I calculated from market fundamentals, it really did not support a continued rally.” Yet the market kept rising in June—“I did get it wrong.” The quarter-end stock-bond rebalance in June failed to have an impact, which he initially attributed to exceptionally strong liquidity: “That was a bit of a blind spot on my part. That wasn’t the reason.”
The real oversight was in reading the leverage data. Net margin was below 1% and gross margin was above 3x; he assumed that meant a standoff between long and short positions, with the market preparing to choose a direction. In reality, gross margin reflected views across different stocks: long semiconductors and short other stocks. “It was effectively a quant long in semiconductors,” not an outright long book.
Momentum’s mechanism, in his words: “It mindlessly goes long the stocks that have gone up the most, then mindlessly shorts the ones that have fallen.” The stocks that had risen the most were almost entirely semiconductors, so “you lift semis all at once, then dump them.”

2. How far has the flush gone? Back to April levels; Goldman says at least another 2 weeks

Deleveraging has brought the market back to April levels; momentum had been following the rally since March. But “you can’t say all the leverage has been flushed out”—the market is merely “at a somewhat more balanced level.” Gross margin remains above 3x, “still much higher” than at year-end. Citing Goldman, he says deleveraging could last at least another 2 weeks, although the impact is fading: once prices are hammered low enough, “a lot of people will come out to buy the dip.”
The leveraged-ETF math is straightforward. Semiconductor leveraged ETFs—2x-long vehicles holding Hynix, Micron, Intel and similar names—fell from $157B to $94B. Given the stock-price declines, they should have sold $78B, but actual leverage declined only $64B, implying that roughly $15B of capital bought the dip during the deleveraging. Volume picked up last Friday; he was specifically watching whether money would come in to absorb the selling.
Overall leverage remains elevated. Gross-margin leverage is at the 95th percentile of the past 5 years, and hedge-fund leverage is at the 82nd percentile of the past 3 years. “If you delever aggressively, there is still room to go.” The macro and micro pictures are diverging: “Macro risks are clearly asymmetric, with more room to run on the downside; but at the micro level, we can see with great confidence that these companies’ earnings are very strong.”

3. Momentum will not V-rebound: rebalancing lags, and history says 4-6 months to recover

Mechanically, the MSCI momentum index looks at the prior 12 months while excluding the most recent month, so it has an inherent lag. A sharp rise in factor volatility triggers forced unwinds and rebalancing—“that is the main reason for the first leg of the selloff we are seeing now.” After a month of declines, many stocks have already been removed from the momentum basket, and the end of deleveraging does not mean they will be bought back: “If you’ve been killed by 30-something percent, you may no longer be the biggest winner, so it may not buy you back.”
The historical sample is tiny. Momentum-index corrections of more than 20% have occurred only 11 times since 1980, and a full recovery to the prior high typically takes 4-6 months. Quant strategies look only at price and volume: “If you go up the most, it buys you; if you fall the most, it shorts you.” Fundamentals must push the companies back into favor before momentum will buy them again.
Deleveraging can produce multiple violent dead-cat bounces, but a genuine long rebound “should not be every stock rising together; it should be one sector after another.” March followed that pattern: memory first, then CPUs led by Intel/AMD, then optical and electronic components. “The real rally rotates, so there is less fear of missing it”（真正的上涨它是会轮动，所以不是那么怕错过）—if you miss Intel’s earnings, you can still buy AMD or optical names.

4. The real focus is credit: hyperscaler issuance will squeeze the IG pool, and the Fed ultimately has to release liquidity

The trigger was Hynix’s $26B financing. “It was mainly that financing that blew up the whole momentum factor.” When a trend has become extremely crowded, a sudden burst of financing can reverse it; the impact of its exchangeable ADRs has not yet come through. The scale of bond supply is unprecedented: “One year’s issuance exceeds the previous 5 years, and next year’s will exceed the previous 10 years.” Hyperscaler CDS had been low largely because there was no supply.
Google reports earnings tomorrow. The market consensus for 2026 capex is $186B, but “we learned” that its internal estimate for compute demand is roughly $230B—not a forecast from the CFO. The 4 hyperscalers plan to raise about $300B this year and $500B next year. Anthropic will likely raise money in October; OpenAI has already raised substantial sums, and Anthropic will need to raise at least $200B-$300B next. “U.S. equities are not actually short of that $20B—the issue is that there is enormous funding demand in front of you.”
The key distinction is that the trigger is not the CDS of the five companies themselves. Their on-paper ROI is “much higher than investing in manufacturing or other industries”; “Amazon would still raise money for capex with its CDS at 120bp.” Amazon’s spread has risen from 30-something bp to 60-70bp, while major banks’ 5-year financing costs are around 80-120bp. The real danger is the entire IG pool being squeezed: capital providers prioritize AI, crowding out financing for smaller banks, pharmaceuticals and infrastructure. “If credit spreads across the entire credit market rise by tens of basis points in a short period, it is catastrophic for the real economy.”
The conclusion is very different from the Street’s. Most people think Kevin Warsh will come in and run QT; he says Warsh can run QT and probably will not do QE, “but he needs to let money flow out” through broader deregulation, by taking the restraints off banks, or through a guarantee vehicle. “Whether it is the White House or the Fed, money ultimately has to be put out.” At next Tuesday’s July meeting, probably on the 28th, most expect a hike; he thinks it is very unlikely. The Fed’s internal disagreement is fairly wide. CPI is “a very lagging indicator,” last month’s reading was driven by oil, and shelter and labor costs have little room to keep rising—one tied to rents, the other to services. Employment is not that strong, so “the downside risk to CPI exceeds the upside.” He has just joined the Fed and needs to establish his own credibility, so he may speak relatively forcefully.

5. Why buyers are absent—and 2 signals for putting on a large position

The first reason investors are reluctant to buy is that VIX is only 18. That is not especially low compared with the lows of 2024 and 2025, but over a longer horizon “it is still at a historical low.” Portfolio NAV volatility “certainly does not look like 18%,” indicating that correlations have been compressed to extremely low levels. His trade is to go long stock correlation and long bond correlation.
The second is the asymmetric risk from CTA strategies. A higher VIX means a falling index, which means CTAs sell—“they sell $140B”—and negative gamma kicks in. Individual stocks may already have fallen sharply, but the index landmine has not been cleared. “When everyone sees that individual stocks have already fallen so much, they do not really dare to buy.”
The first entry signal is a higher VIX and an index catch-up selloff. Semiconductors have already taken their hit; if “the index falls 10 points, semiconductors may also fall only 10 points,” making the risk-reward more attractive and clearing a landmine ahead. The second is momentum flipping to short semis after the unwind: “Semiconductor earnings are genuinely very strong… a blowout earnings print can get the trade blown up in reverse.” The momentum index is already down 35 points, so “you cannot push it another 35 points lower; even another 10-plus points would be difficult.”
The third is a long-term position. “If you believe AI is not ending, and the credit problem does not explode, AI valuations are not high right now.” The condition is that “the credit problem does not blow up,” which depends on the Fed not allowing it to happen.

6. Kimi is not a DeepSeek moment: 2.8T can only go back to the cloud, and open source is positive for compute

Kimi is “extremely useful and very powerful,” and compute capacity was already sold out by the weekend after its release. The key difference from DeepSeek is that DeepSeek was a training-side cost-reduction story—“we may not need that much compute”—which primarily hit Nvidia. Kimi is an open-source model, and the main race has now returned to inference infra.
The deployment math is simple: Kimi is a 2.8T-parameter model; “even divided by 4, you still need more than 700G.” It requires 10 H100s, and 8 is not enough. “There is no way to deploy it on a local machine; it has to go back to the cloud.” By contrast, GLM can be deployed on 2 large machines. For cloud providers, 3-5 open-source models are better than 2 large private customers. “Open-source models getting better is actually a positive signal for compute investment,” although the market may need time to digest that.
The near-term risks are a decline in Anthropic’s earnings and a shortage of “greater fools”—in other words, a shortage of people willing to provide the money. “The biggest risk in semiconductors right now may be credit, and credit crowding out non-AI.” The broader tone remains optimistic: “AI keeps rolling forward”（AI还是滚滚向前）. Lower token costs are good for the entire AI ecosystem; “what is really behind semiconductors is compute, not entirely the models.”

2026: AI’s Breakneck Run, Capital Drain and Our “Engels Pause” — A Crossover with Jinbo Finance Mark Cuban on the AI Bubble: Who Actually Gets Wiped Out?