Market Overview July 14, 2026

Summary

• The dominant theme over the past 2 weeks has been the unwinding of the momentum trade. All AI-linked assets—semiconductors, memory, new cloud names (likely CoreWeave and Nebius), foundries, and optical modules—sit in the same highly levered asset pool, so a broad deleveraging hits everything indiscriminately. Jin’s core view is that storage companies’ “earnings will be very good for the next 2 years”; what is falling is valuation, not fundamentals, and this kind of violent valuation swing “will keep happening in semiconductors over the next 1-2 years—this is definitely not the first time, nor will it be the most violent.”
• The trigger was a financing-driven liquidity drain. SK Hynix raised roughly KRW43T through an ADR offering (~$25B), and “that money had to be pulled out of the long-leg momentum trade that believed in AI.” Levered positions were triggered, setting off a stampede on top of profit-taking after the market had surged 3-5 sigma in the short term. He repeatedly stressed that “stocks are fundamentally about supply and demand”: semiconductor earnings are excellent and companies have done almost no external financing, so once the supply-demand balance is distorted, a single large financing can ignite the whole move.
• The deleveraging is not over, but it may not need to be. The first wave of long-short leg unwinds is “possibly less than half done”; fully invested holders still need to cut positions, while underinvested holders may cut again after a sharp enough pullback. But the further deleveraging drives prices down, the lower the payoff to selling becomes—at KRW1.8M, some capital will decide that even another KRW300,000 decline over a few months still leaves room for a rebound to its entry price. The discipline for catching the falling knife is simple: “know that you will not catch it all the way down,” and ask only whether the current price is acceptable.
• Two signals show that long leverage has not been fully cleared. Software names—ServiceNow, possibly CrowdStrike, and Palantir—have rebounded as hardware collapsed, suggesting the most violent deleveraging may be over but the parity trade has not fully unwound. Options skew still favors paying up for calls; compared with the pandemic, when “calls were almost completely ignored and everyone was scrambling to buy puts,” the market is nowhere near a fully cleansed, reverse-positioning state.
• The condition for the next broad selloff is a rebound in correlation. Software rising while hardware falls has pushed S&P correlation to a historical low, which is why VIX has not rebounded. If correlation rises broadly, VIX rises, negative gamma compounds the move, and CTAs sell—around $100B globally on the downside, including more than $50B in the US—the result would be a broad selloff, “and that would be the absolute bottom.” He nevertheless leans toward this case failing on timing because capex and earnings catalysts are too close.
• Passing the model test has made him bullish in the short term. Anthropic says it generated $1B of operating profit in a quarter; GPT’s new model, “So,” is “both cheap and useful… and trounces China’s open-source models,” while its “Ultra Code” mode calls on substantial hardware and “Honeys” is exceptionally strong. On capex, Google’s internal number for next year is roughly $250B versus investment-bank expectations of about $180B; once prices have fallen, capex could readily become a catalyst.
• The real tail risk is in CDS, not equities. Hyperscalers will need to raise several hundred billion dollars in the future and another $500B next year, yet Amazon’s roughly $26B financing was enough to push hyperscaler CDS spreads about 20 bps wider—“that magnitude is already terrifying.” His warning was explicit: CDS risk needs to be hedged; without a hedge, “this really could blow up.” His road-condition metaphor says it all: the road was flat before, but “the road ahead may be icy, so you need to drive carefully.”

Deep dive

1. The mechanics of the stampede: SK Hynix’s $25B drained long-side liquidity

• Jin’s read on the selloff: storage fundamentals have not broken—“I never thought storage deserved a PE multiple… earnings are unlikely to deteriorate much”—but the momentum trade was unwound. Every AI-linked asset that had rallied—semiconductors, memory, new cloud, power, foundries, and optical modules—belonged to the same highly levered pool. When that pool is sold down in a deleveraging, no one survives: “It is the same logic as Goldman Sachs getting crushed during the 2008 financial crisis.” TSMC Co-op was rumored to be in trouble, but future use of higher-quality light sources (200, 400; 200, 400×4) would actually be better for Lumentum; Lumentum fell anyway.
• The direct trigger was SK Hynix raising roughly KRW43T (~$25B) through an ADR offering, pulling cash out of the AI-believing long leg and triggering levered positions. The setup was a short-term move of 3-5 sigma. His first-principles view is that “stocks are fundamentally about supply and demand”（股票本质还是要看供需的）: semiconductor earnings were excellent, companies had raised almost no external financing—Intel, Lumentum, Coherent, and Corning all absorbed their strategic financing internally—and scarce supply combined with the AI narrative had completely distorted the supply-demand balance. One large financing was enough to set it off.
• For Chinese investors, the comparison is the exact same quant stampede seen in the CSI 2000: long small caps, short the index, and when a concentrated book is unwound, old-guard assets rally while emerging assets collapse. If ChangXin is financed next, the same pool of capital will be pulled out again, taking the CSI 2000 down as well. “So when you see small companies with very good earnings falling like hell, well, they should fall like hell.”

2. Two post-mortems: parity trade underestimated, “dirty positions” not confined to storage

• Jin offered an unusually public post-mortem on 2 mistakes. First, he had seen gross margin reach as high as 300% and net margin around 50%, with both declining steadily from April onward, and attributed that to disagreement between longs and shorts. In hindsight, it was likely a large parity trade: funds were long semiconductors while shorting software on the view that software was a substitutable asset, or shorting Microsoft, retail, and Apple. Asset-side leverage was extremely high.
• Second, he had repeatedly said that positioning in storage was “very dirty”—Korean leverage and leveraged ETFs—but the actual situation was that every AI-related asset had dirty positioning, not just storage; storage was simply the dirtiest. That is why a storage stampede triggers indiscriminate selling everywhere else.

3. Is deleveraging over? Three waves of unwinds and the math of catching the knife

• The unwind comes in 3 waves: the long-short leg goes first—“possibly less than half done”—followed by fully invested storage and AI holders cutting positions, and then underinvested holders who may cut again after a sufficiently large pullback. But as deleveraging pushes prices lower, the payoff to shorting, selling, or hitting stop-losses shrinks. Those reducing positions will buy back below; those who are not fully invested may turn around and add. Every semiconductor company is due to report earnings within 2-3 weeks, and “the earnings we can see are all very good.”
• The discipline for catching a deleveraging knife is: “If you think about catching it, know that you will not catch it all the way down.” The only question is whether the current price is acceptable, which is the same logic Buffett uses. If SK Hynix falls to KRW1.8M, some capital will naturally think: even if it falls another KRW300,000 over the next few months, the next rebound should at least bring it back to my entry price, so I will not lose money.
• He also warned against mechanically demanding that leverage go to zero: “You cannot carve a mark on the boat to find a sword.” A stock may have risen 5x or 10x, but earnings and earnings expectations may have risen even more, while the arrival of new models has fundamentally changed the underlying picture. The exception is if you no longer believe in AI at all.

4. The dashboard: software-hardware divergence, call skew, and record-low correlation

• There are 2 signs the cleanout is incomplete. Software—ServiceNow, possibly CrowdStrike, and Palantir—is still rising as hardware falls, although the gains are narrowing. That suggests the most severe deleveraging may be over, while the long-short parity unwind is not. Options skew still favors buying calls at a premium, another sign that long leverage has not been fully washed out. A truly reversed and cleansed setup looked like the pandemic, when calls were almost completely ignored and investors scrambled to buy puts; the mid-2025 rebound and the April tariff-war rebound had the same reverse structure.
• The macro switch is correlation. Software rising while hardware falls has pushed S&P correlation to a historical low—“correlation is at its lowest point now”—leaving single-stock volatility huge while aggregate VIX remains subdued. Once the unwind ends, there are 2 paths: normalization, with correlation staying low; or a broad rise in correlation, followed by higher VIX and another leg down across the market. If a broad selloff combines with negative gamma—dealer hedging that sells into the decline—and CTA selling, with stale data pointing to roughly $100B globally and more than $50B in the US, the stocks hit hardest would be those still rising today. “That would be the absolute bottom.” He leans toward this case not necessarily working on this timeline.

5. The model hurdle is cleared; capex turns from a problem into a catalyst

• Two near-term developments have shifted him bullish. Anthropic says it is profitable, with $1B of operating profit in one quarter, easing the concern over whether models can make money. He tested GPT’s new model, “So” (his wording): “It is both cheap and useful… it trounces China’s open-source models.” He finds China’s open-source models roughly comparable to Claude for coding, but GPT’s “Ultra Code” mode draws heavily on hardware and “Honeys” is extremely strong. His framework is that a model is the combination of data, “Honeys” engineering, the model itself, and compute—4 things stacked together.
• The capex logic has reversed. For a month, he had argued that investors should reduce positions and get out before July earnings because when prices were high, “capex could never come out right.” Now that prices have fallen, capex can become a catalyst. Several companies’ capex is running very high: Google’s internal estimate for next year’s capex is roughly $250B, versus about $180B from investment banks; even the most bullish buy-side view reaches only around that latter figure. Once the numbers are made explicit, semiconductor stocks that have already sold off can get a catalyst.
• Three short-term uncertainties remain on the table: the Fed—“Wash” has made hawkish comments, raising the question of whether it could simply hike once on Friday, although “it is really just trying to scare you,” and the size of subsequent hikes may be smaller; Iran—whether Trump throws caution to the wind and goes all the way or caves; and credit, covered in the next section.

6. The real tail risk is in CDS: “The road may be icy, so drive carefully”

• A month ago, he was already saying to buy CDS as a hedge. The true extreme risk is not the stock price but hyperscalers needing to borrow heavily to fund capex, pushing CDS and financing spreads wider. “A wider interest spread is a very dangerous thing,” as shown by the reaction in Nvidia and semiconductors after Oracle’s stock ran into trouble. 2 questions determine whether the market is willing to lend: how much capex you are committing, and what the ROI will be.
• The sensitivity to financing size is unsettling. Amazon raised only about $26B, yet hyperscaler CDS spreads widened roughly 20 bps in a straight line over the past month. “That magnitude is already terrifying,” especially with several hyperscalers set to raise several hundred billion dollars and “another $500B next year.” His conclusion was unambiguous: “CDS risk needs to be hedged; if you do not hedge it, this really could blow up”（CDS你的问题需要对冲的，你不对冲的话，这个是真的有可能会爆的）.
• His closing road metaphor and the full framework are straightforward: these companies all have strong earnings. “It is not that you cannot drive; it is that the road ahead may be icy, so you need to drive carefully”（往前的路上可能是结冰的路，你要开车要小心）. Semiconductor earnings have not been impaired; this is a positioning problem—the momentum trade is unwinding, and the process may not be finished or need to finish. The real risks ahead are macro, through higher VIX and correlation, and credit, through CDS. The former is short term; the latter must be hedged.