Market Update

Summary

• Mr. Jin sees durable alpha in industry insight and remains structurally bullish on AI and semiconductors. Two weeks ago, he flagged liquidity risk but did not recommend selling everything: seeing a risk does not mean you can trade it accurately. In his quant experience, a single price-and-volume factor at retail trading frequencies rarely delivers a durable edge. Over the past 5–8 years, most of the returns from selling puts still came from “the market going up”; getting out ahead of a couple of selloffs may simply be survivorship bias.
• His core view is that data and semiconductor usage is becoming denser, and the market is not yet ready for semiconductors to become the largest industry. The progression runs from the internet moving paper records onto computers, to mobile internet driving more access, to AI prompting high-frequency usage; demand growth does not depend solely on chatbot revenue. To him, semiconductors are a binary bet on whether compute is sufficient: if you truly believe orders will be cut in 2027–2028, there is nowhere to hide—not at TSMC, ASML, or anywhere upstream or downstream.
• He attributes this selloff to liquidity transmission from rate volatility and rejects inferring a demand peak from falling share prices. The sequence was rising rate volatility, stress in credit, sovereign credit and FX, then transmission from credit into CDS and equities. He thinks hedging and forced-liquidation pressure on the rates side has eased. The fading tariff effect over the next 3–6 months also supports lower rates, but credit spreads have yet to recover, and further disruption to oil prices from Iran could reopen the “upstream water source.”
• He does not think CDS alone can force CoreWeave into another Lehman. By his account, the company has about $46B in PP&E, $30B in net debt and $8–10B in cash; operating cash flow is positive and roughly covers interest. About half its financing is also backed by GPU collateral or cloud-provider support. He acknowledges that rising CDS can hurt financing and create a negative feedback loop, but sees reduced investment as the more likely downside, with the gap absorbed by other buyers.
• Enterprise post-training is a source of demand he thinks the market has not fully priced in. General models must absorb enterprise data and know-how to go from smart “PhD students” to productive employees, and that training has to continue. His estimate of Microsoft-related demand has risen from about $2B at the start of the year to roughly $20B now. Meanwhile, short-term rental and weekend compute prices have not shown the oversupply signals he wants to see, so he does not accept a slowdown in Anthropic’s growth alone as proof of excess compute.
• His long-term view of Labs is shifting from model capability toward data accumulation and direct operations in high-value use cases. Users pay while also supplying post-training feedback, potentially enabling Labs to enter businesses such as search and quantitative trading themselves. He has sold Google: he initially worried about search-ad revenue, but now questions whether search itself remains a monopoly. “In search, you can’t even be sure who will deliver the best results in the end.”
• Mr. Jin expects Anthropic’s IPO could break below its offer price, while stressing that model companies and semiconductors can have different outcomes. He says 35x or 40x sales is not what determines whether the trade works; more important are the urgency signaled by a 25% price cut and the allocation between long-term investors and hedge funds. “No matter what price it lists at, it could break below the offer” does not mean he thinks the entire compute cycle is over.
• He is relatively confident that chips, power and financing will be ready for 2027, but remains uncertain about 2028 and China’s overseas supply chain. Based on what he knows, power constraints will not broadly force project cuts in 2027; for 2028, he “doesn’t know.” A potential obstacle for Chinese optical communications companies could be restrictions on building factories overseas. Even if InnoLight trades at only 11–12x next year’s forward PE, as the interviewer suggests, its industrial capabilities and “reliability of supply” still need to be assessed separately.

Deep dive

1. Seeing risk does not mean you have alpha in trading it

• Mr. Jin warned two weeks ago that liquidity could deteriorate, but did not recommend selling everything because the balance of upside and downside risks was relatively symmetrical at the time. His starting point was: “Don’t make a blind move.” Seeing a risk indicator does not mean you can consistently turn it into returns.
• In his quant experience, single factors such as implied volatility, fear indices and order-book signals rarely provide reliable alpha at frequencies retail investors can execute. Millisecond trading, multi-factor portfolios, or access to prime-broker and market-maker positioning can offer an edge.
• He says the only simple strategy that worked reasonably well over the past 5–8 years was selling puts, and those returns still came from “the market going up.” Selling calls, in his view, does not make money; even if you keep doing it, that becomes clear when you look back. Avoiding two selloffs in a year and thinking you are good at it often reflects survivorship bias.
• Fundamental calls may take two or three years, or even five, to play out; you cannot expect the price to validate them every day. He compares industry insight to seeing your hole cards in Texas Hold’em: if you entered to capture fundamental alpha, “then hold it and don’t move.”

2. The underlying shift in the AI cycle is denser data usage

• Mr. Jin links the three cycles this way: the internet moved paper records onto computers and connected them; mobile internet drove more frequent access to accumulated data; AI is bringing data and semiconductors into contact more closely and more often.
• So he focuses not just on “how much work AI can do,” but on the ongoing expansion of data, semiconductors and then energy. Even if the price per token falls, that does not directly imply lower semiconductor demand.
• He thinks we are at the “end or beginning.” People rushed in on imagination at first, then began questioning the opportunity just as real demand and capital started to arrive. He admits he had also doubted the scope of the AI opportunity, but now thinks CSPs will move faster to meet enterprise demand.

3. The selloff began with rates and credit; demand doubts came later

• Mr. Jin stresses that rising rates three or four weeks ago were not, by themselves, the most dangerous signal. The danger came when rate volatility rose at the same time and crossed a critical threshold. Once oil prices became linked to rates, war-driven oil volatility was transmitted into the vast rates derivatives market.
• The shock spread to corporate credit, sovereign credit and FX; South American currencies were already weak before equities fell. He thinks the panic in European bond markets two or three weeks ago was even worse than what we are seeing in equities now.
• A bearish demand narrative effectively assumes cloud providers are unreliable customers and that orders to TSMC and further upstream cannot be trusted, then prices in cuts for 2027–2028. He disagrees: “Don’t reverse cause and effect.” A decline followed by demand doubts is different from demand deteriorating first.

4. Rates have begun to stabilize, but credit has yet to recover

• Mr. Jin thinks this month’s CPI was worse than last month’s, yet rates reacted more moderately, mainly because the convexity-hedging pressure he refers to has eased. After the jobs report, a rate hike in October was “off the table”; it was not a few Fed officials’ remarks that changed expectations.
• He compares concentrated swaption liquidations to a gamma squeeze in equities and thinks this squeeze is close to ending. Rates are more mean-reverting, so at current elevated volatility, “I would definitely sell swaptions now.”
• His view is that “the upstream water source has been shut off,” but that does not mean liquidity is abundant: some credit spreads and French, Italian and Spanish government-bond spreads remain high. Unless Iran causes another shock, he does not yet see the source of stress worsening.

5. Hedge funds may hunt leveraged positions where liquidity is thin

• Mr. Jin uses “hunting elephant” to describe short-term trading: with the midterm elections about two weeks away, the market also has to raise cash for financing needs, while buyers are reluctant to put money to work ahead of earnings. Capital may therefore choose to apply pressure in the periphery.
• The cluster of after-hours news on Thursday and Friday made him suspect that someone was trying to flush out known, large leveraged positions. He explicitly describes this as his observation and judgment, and thinks the neocloud segment—the most vulnerable part of the AI chain—could become a target.
• The real concern he discussed with former Goldman colleagues was that rising CDS could raise financing costs and feed back into a crisis. That negative feedback loop deserves attention. But he thinks shorting may be relatively safe in the near term and very risky over a longer horizon; it is not enough to conclude the whole industry has changed for the worse.

6. The key in 2008 was that the underlying assets had already deteriorated

• Looking back at 2008, Mr. Jin says mortgage delinquency rates were already rising. Losses hit equity first, then mezzanine, while CDOs were synthesized from mezzanine tranches. He says CDOs and super-senior tranches could ultimately both lose around 40%; the underlying cash-flow problem was real.
• Lehman also held large amounts of hard-to-price, interconnected derivatives, while the government allowed leverage to be unwound in a concentrated way. He thinks the situation would not have developed that way if liquidity had been provided then. CoreWeave today has neither the same scale and interconnectedness nor the same deterioration in its underlying assets, so he says 2008 is not a direct comparison.

7. CoreWeave’s credit needs to be assessed through its assets and financing structure

• On CoreWeave, Mr. Jin says about half its financing is senior unsecured, with spreads up to around 15%; the other half is secured, collateralized or supported by cloud providers. The debt should not all be treated as carrying the same risk.
• He puts PP&E at about $46B. Most sites are leased, and the assets are largely production equipment such as chips, interconnects and power systems. Net debt is about $30B, and cash is about $8–10B. He suggests listeners compare the secondhand prices of H100s and A100s to judge how much the assets could actually recover.
• The company is not in good shape, but he emphasizes that operating cash flow is positive and “roughly covers interest”; he also thinks the assets can cover the debt. To him, that is the key difference from CDOs whose underlying collateral was already suffering cash-flow losses.
• Asked whether CDS could directly force the company into default, he says, “I don’t think that’s possible.” In his downside scenario, the company reduces investment and transfers or sells projects in packages. With compute and power still scarce, he thinks other buyers could absorb the demand.

8. The semiconductor bet is on a compute shortage; positioning differs between the US and Korea

• Mr. Jin reduces the semiconductor trade to “zero or one”: if you think compute is insufficient, that is one; if you think it is already in oversupply, that is zero. If orders are truly cut, everyone from design companies to TSMC and ASML is affected; moving to another part of the chain does not avoid the same demand call.
• He chooses “one” because orders at the front end and demand further upstream remain full, and says aggregate forward PE has returned to its level before GPT launched. He also thinks overall leverage and positioning in the US market are low; a few highly leveraged “elephants” do not mean the whole market is leveraged.
• In Korea, he still thinks memory is in short supply but does not accept the scenario of margin rising to 90%. Leverage in Korea rebuilds quickly after deleveraging; by comparison, he thinks US positioning offers a more attractive risk-reward.

9. The 3–6 month rates call is a bet on the tariff effect fading

• Mr. Jin uses no hike in October and a possible hike in December as his framework, and thinks the tariff effect fading over the next 3–6 months is relatively clear. Last October’s government shutdown may distort year-over-year data, so he pays more attention to December and January.
• He explains that the lagged impact of tariffs may peak in March or April this year, then ease month by month; the impact on core inflation is “about 0.8,” a large component. To keep generating inflation, there would have to be new tariffs; the original price increase cannot be counted again.
• The same applies to oil: holding at $100 is different from continuing up to $150; the latter depends on whether Iran keeps disrupting supply. He says he does not know, but thinks that is unlikely. In a supply-driven inflation shock, a rate hike “can’t produce another cow.” He thinks rates have risen too far recently and have room to fall.

10. Enterprise post-training turns PhD students into productive employees

• Mr. Jin’s analogy is that even the smartest PhD student cannot work for a company just by rereading “one employee handbook plus four skills books” every day. Models need to absorb a company’s own data and know-how, so post-training demand cannot be estimated from chatbot growth alone.
• Asked whether post-training means CSPs choose verticals based on customer data, Mr. Jin first says, “No, no.” He then explains that companies can train models on their own data with Microsoft or Amazon. His estimate of Microsoft-related demand has risen from about $2B at the start of the year to roughly $20B now, and he suggests watching how Microsoft and Amazon discuss corporate AI demand.
• Companies may feel more comfortable giving data to Microsoft, which has served them for years. As he puts it: “You’re more willing to give it to Microsoft. It has been providing your SaaS for years; if it were going to steal your data, it would have done so already.” Cloud providers’ deployment capabilities and customer use cases are also competitive advantages.

11. Feedback from high-value use cases builds more defensibility than general capability rankings

• Mr. Jin thinks users pay to use a model while also contributing post-training feedback, creating a two-way loop between the model and data. That makes coding’s first-mover advantage hard to catch and leads him to think ByteDance, not DeepSeek, is the underestimated Chinese model company: what matters is already having a foothold in high-value use cases.
• He uses his own BN order-routing needs as an example: multiple servers must finish their calculations and arrive at the same AWS location in the same millisecond. Early on, Anthropic had no idea about this routing know-how; now, with a simple description, it can do it “pretty well,” and he says it outperforms an experienced team working on it for a week.
• He also says that after he labeled data from web novels for a smaller model and trained it, the model wrote better than a general-purpose model. Post-training is not a one-off project: enterprise data grows and workflows change. He also acknowledges that using a smaller model after post-training could mean inference demand is lower than expected.

12. He sold Google after changing his view of search’s monopoly

• Mr. Jin initially thought that AI not reading ads would hurt Google’s search revenue, but that search capability itself would remain a permanent monopoly. He has now changed his mind: “In search, you can’t even be sure who will deliver the best results in the end.”
• By his comparison, total usage of other search engines last year was about 0.5% of Google’s; by the start of this year, Anthropic’s search usage had reached about 20%. He guesses that even 70%–80% now would “not be surprising.” He mentions Brave and gives searching for a Hong Kong MTR exit as an example: sometimes OpenAI works better than Google Maps.
• His further thesis is that Labs could enter high-value businesses itself once it has accumulated enough feedback, including quantitative trading. Even Jump, Tower or Jane Street may not have an edge against Labs if it can screen factors and has more compute; he presents this as a future possibility, not as a claim that Labs have already entered those businesses at scale.

13. Slower Anthropic growth is not yet proof of compute oversupply

• The interviewer raised a specific counterargument: if Anthropic’s MoM growth falls to about 5%, that implies 80%–90% annual growth if sustained, while compute plans are set to double over the next two years. Before CSP post-training revenue is disclosed, does the market’s negative equity story have a fundamental basis?
• Mr. Jin thinks Anthropic’s price cuts are a negative signal and separates two effects: the per-token price may be under pressure from OpenAI’s price war, while the impact from open-source models is “not that severe yet.” Supply constraints and deployment pace may also limit revenue; neither is directly equivalent to compute oversupply.
• He wants to see rental prices confirm the thesis: on-demand short-term rentals should at least be near long-term contract prices for the market to look oversupplied. Short-term rentals remain tight, and weekend prices have not fallen. He therefore thinks training demand remains strong: “Those saying you shouldn’t train models so quickly” are mostly still training privately.

14. Anthropic’s IPO risk lies in financing needs and the buyer mix

• Mr. Jin says, “No matter what price it lists at, it could break below the offer.” The semiconductor trade comes down to whether compute is sufficient; model companies each face their own use cases, competition and monetization challenges. Microsoft and Amazon having strong enterprise channels does not mean independent Labs can monetize just as easily.
• Asked whether a price cut could improve the trade, he says there is no fundamental difference between 35x and 40x sales. Sentiment can push a valuation to 50x or 100x, or crush it to 20x. The price cut first signals that the company “really wants to raise” and is desperate.
• He cares more about allocation: IPOs where quality long-term investors cannot get enough shares and the issuer is reluctant to allocate to hedge funds tend to do better. After a price cut, it is hard to choose the buyers. He cites Facebook at the time: broad demand for the offering and sales to retail investors did not guarantee a strong initial trading performance.

15. Restrictions on Chinese models could drive more companies to build their own compute

• On possible restrictions on the use or export of Chinese models, Mr. Jin thinks the existing open-source models “can’t be put back in the box”; this will not affect things now or even over the next six months. But he thinks the probability of future policy action is significant, because it is difficult for the decision-making system to oppose the view that “leading technology needs to be controlled.”
• On the demand side, he thinks this could push companies such as Microsoft from “keeping a model of their own as a backup” to having to build their own models. Since enterprise post-training is the main source of demand, companies cannot assume DeepSeek will remain open source forever; they will still need to build their own models and compute.

16. The potential headwind for Chinese optical communications is reliability of overseas supply

• Mr. Jin says the policy risk he has heard about is that Chinese companies building factories in places such as Thailand could be seen as making concessions to the US and face resistance at home. He is not certain, but worries this could affect companies such as InnoLight continuing to serve as first-line suppliers to US CSPs.
• Even if production ultimately remains in Chinese factories, it could be subcontracted through other suppliers because buyers look not only at capacity but also at “reliability of supply.” He thinks packaging and testing demand will remain strong next year, with shortages at every stage; near-term results may not be affected immediately, but others could build the new overseas capacity.
• The interviewer asked about InnoLight at roughly 11–12x next year’s forward PE, the alleged 65% FCC rule and Dongshan Precision’s Hong Kong IPO. Mr. Jin says he is not good at mapping A-share valuations and did not give a clear view on whether Dongshan should be compared with AAOI. He was more certain about InnoLight’s packaging capabilities, while stressing policy risk on the China side.

17. Chips, power and financing are lined up for 2027; 2028 remains unknown

• Asked whether US power approvals and a cap on BE installations would force projects to be cut, Mr. Jin says, “Not next year; I don’t know about the year after.” Looking from 2026 to 2027, he understands that the major CSPs have already secured power and alternatives.
• He breaks the preparation into three parts: the source of chips is broadly clear; if power is short, there are existing reserves that are not yet connected to the grid, gas turbines or BE; financing and operating cash flow are also in the plans, with about $420B of debt issuance planned.
• He argues against drawing a direct parallel with the 2000 internet bubble: some companies then invested on a whim and bought servers indiscriminately, while large companies today invest after modeling demand and resources. That is why he has some confidence in 2027, not a guarantee about 2028.

18. Value Huahong, SMIC and China’s premium separately

• Mr. Jin thinks Huahong and SMIC are both good companies, but says customers in the past asked them to manufacture GPUs and CPUs by using DUV “as if it were EUV,” with repeated exposures compressing margins. He says the company has recently dropped off the “far ahead” list; margins at the two companies could improve in the relevant areas.
• He still thinks foundries should ultimately be valued on PB: Huahong is at around 3.x times and GFS at about 2x. The interviewer explained the valuation by calling them the “only two seedlings” in China. Mr. Jin agrees that this includes a “China premium,” but does not think that means alternatives in other markets can be ignored.
• As a comparison, he proposes a hypothetical Intel breakup: if the CPU business is valued at 20x earnings of 5, that implies 100, and the foundry’s implied PB would be close to zero. This is his valuation framework, not a claim that CPU earnings have already been realized.

19. The gap in CRWV’s results is still unexplained, so it should not automatically become a major position

• Asked whether CRWV should be treated as a lottery ticket or a major position, Mr. Jin only answers with his own choice: “I definitely wouldn’t make it an important position.” After selling Google, he also did not move all the proceeds directly into semiconductors; by comparison, he is willing to hold a larger position in INTC.
• The issue with CRWV is that some investors used TSMC supply-chain orders to project strong results, but the company did not report what they expected. He says he is “not sure” whether capacity has yet to show up in the numbers or is being discounted along the way. He currently leans bullish, but the gap remains.
• He cautions that after a broad selloff, other stocks with more certain fundamentals may have fallen further; opportunities should not be ranked by their own drawdown alone. If this is a leveraged-position hunt, he thinks reports of a large fund blowing up could be a buy signal: “Generally, these selloffs end once someone has been killed.”

20. Intel EMIB’s bottleneck: distinguish packaging yield from substrate supply

• In the final round, the interviewer worries that EMIB’s substrate and small-silicon-die structure is prone to warping and could hurt yield. Mr. Jin corrects him: “Yield is good; there’s no problem.” He thinks the issue is supplier substrate yield, which leaves too few substrates available and limits shipments.
• Improving that yield requires supplier investment, and Intel as the customer also has to spend money to push it along. He compares it to HBM: “You’re the customer; spend the money and you can improve it.” Based on what he has heard, supply is currently okay, but better substrate yield still depends on the supplier.