Herman Jin: Chip Shortages Favor Laggards; Buy Optics, ChangXin, YMTC

Summary

• Herman’s core thesis is that semiconductors are forming productive capacity directly for the first time in 40 years. It is comparable to the steam-engine and internal-combustion revolutions: the 2000 tech revolution had extremely low utilization—the utilization rate of 5G base stations was only 5-10%—and semiconductors were merely software platforms. TSMC did not believe in AI and lacked the courage to expand from 2023 onward; “if TSMC does not expand, nobody dares to expand.” Capacity is at least three years behind, while token demand “is infinite” on the demand side. The AI trade is “just getting started… end of beginning.”
• Semiconductor shortages favor the catch-up players. “When a major semiconductor shortage hits, the one that profits is always the laggard”—AOI in optical modules, Intel in wafers, AMD in GPUs ($210 bought as a large position), and ChangXin and YMTC in memory. Intel’s five steps—avoiding bankruptcy, internal use, external use, profitability, and taking share from TSMC—“will all be completed.” 18A yields climbing 1-2 points a month is a one-way function; at 6-7x PB in 3 years, Intel could be worth more than $1T. Only Intel and TSMC can make 2nm globally; “Samsung is garbage.” But surpassing TSMC is “a fantasy.”
• Optical modules are a buy-across-the-board trade. Optics account for roughly 20% of NVDA rack procurement, yet the combined market caps of all optical companies are below Micron’s. Each generation penetrates more deeply from scale-in to scale-across; Google’s new GPU architecture could double optical demand, while the 800G-to-1.6T upgrade can lift gross margins organically without capacity expansion. Memory is a commodity—“it is difficult to value an oil company on a PE basis”—and although it will certainly make new highs, he does not buy it: “fools, or smart people who understand fools, can make the most money.”
• Interbank liquidity should remain workable through June. Unwinding ESLR theoretically releases $4.5T of bank exposure, or $1.5-2T in practice; money-market lending is about $3.1T, swap spreads are low, and the White House’s stance is essentially, “I want to make it go up.” He has been cutting positions and “sold out of quite a few winners,” but “do not short it under any circumstances.”
• A meaningful, multi-month correction may arrive around year-end or early next year. Wall Street’s final question is whether OpenAI and Anthropic are “stepping on each other’s feet,” with hyperscaler FCF declining and CDS spreads rising from near zero in 2024 to roughly 15 bps on average. Add oil prices acting like a variable tax, with inflation transmission lagging by 6 months; weakening employment; and the liquidity drain from the OpenAI, Anthropic, and SpaceX IPOs, and “the probability of resonance is fairly high.” If the two companies each list at $2T and the models come in slightly below expectations—the real reason would be insufficient compute, but the market will interpret it as model failure—“the whole market will crash on you.” If the market really falls for several months, “buy decisively, decisively.”
• China’s semiconductor industry is strong, but Huawei has pushed it down the wrong path. Using DUV multi-patterning to replace CPUs and GPUs is an “patriotic business” that doubles costs and halves yields. The correct strategy is memory: DUV is sufficient, and DDR is a standardized commodity. With national support—including zero-interest loans of hundreds of billions of RMB and NDRC pressure on production lines to support ChangXin—ChangXin and YMTC are “the two most investable companies in China.” Buy them at market prices when they list on the STAR Market.
• BTC is only an S allocation for liquidity hedging. He buys around $60-70K but feels “not the slightest excitement” at $120K; it should be held for years. “You might as well buy some semis—I can give you 2 names at random and both will outperform bitcoin.” “All altcoins will go to zero.” Crypto is a “very, very manipulated market.” Circle is a “money-market fund that pays no interest,” with its value driven mainly by regulatory arbitrage and deep political ties: “I might as well buy TSMC.”

Deep dive

1. Abundant Liquidity Is the White House’s Stance; This Rally Started With Forced Buying, Then FOMO

• Herman’s opening framework is that liquidity “needs to be monitored continuously,” but the evidence is clear: SOFR is low, average daily money-market lending is $3.1T versus $2.9-3T last year, and term premium minus swap spread is low across both 5-year and 10-year maturities, signaling abundant bank-side liquidity. Treasury issuance should be light in Q2, and heavier issuance after July does not imply a reversal: the Treasury Secretary, likely Bessent, said on May 6 that issuance proceeds would be put back into repo. Herman reads the White House between the lines: “I don’t really care if abundant liquidity creates a bubble… As for the market, I think it should go up. I want to make it go up.”
• Unwinding ESLR—announced last December and implemented on April 1—removes Treasuries from the denominator, allowing banks to hold Treasuries as collateral. The theoretical increase in exposure is $4.5T, versus $1.5-2T in practice. The Fed has 2 rational paths: slowly release liquidity through regulation, or combine QT with regulatory easing. Herman guesses the latter—rate cuts plus QT—but says, “I don’t even know whether he will do QT.”
• The anatomy of the rally: in the 3 weeks before March 31, systematic strategies—CTA plus vol control—bought $180B globally, in waves of $40B, $80B, and just over $30B. “If you are trying to pull the market up, the money you put in at the beginning works best.” Once the market was pulled back into place, 2 things became formal catalysts: compressed semiconductor valuations, with Nvidia earnings up 70%—“$180 is no different from $90 last year”—and Anthropic proving that “AGI has already been achieved.” The second phase was broad-market FOMO.
• The short-term technical setup lacks the same certainty of buying, so you cannot “make a snap decision.” But call skew is high, creating the potential for a gamma squeeze, while low put vol suggests someone is willing to take the other side. Still, “all vol exists to get blown through”; once it is breached, the selloff accelerates. Calls are concentrated in single-name semiconductor exposure, including Micron and Intel, and a long squeeze would also hit semiconductors.

2. The Year-End/Early-Next-Year “Resonance” Correction—Buy Decisively Into the Drop

• Wall Street’s final question is whether OpenAI and Anthropic are “stepping on each other’s feet”: the money they raise is spent on CSPs, while the hyperscalers actually paying the bills are seeing free cash flow decline. CDS spreads have risen from 2024 levels—when Microsoft CDS was roughly in line with banks and essentially zero, making 2024 “the healthiest stock market”—to about 15 bps on average, with even Meta gradually moving higher. Add oil prices: “There is no fundamental difference from a tariff; it is effectively a tax,” with at least a 6-month lag before reaching PCE, plus weakening employment. “If resonance develops, I think the probability of that resonance is fairly high.” He stresses that this would not be a crash, but would last “at least several months,” and says, “Never, I never said” it would definitely happen this year.
• IPOs create a liquidity-drain risk in the same window. The market is currently in a comfortable state of isolation: TSMC, the “gatekeeper,” is not expanding and is supporting margins across the sector; without OpenAI and Anthropic being public, there is no linked financial data with which to challenge the story. The host compared it to being most attractive before taking off one’s clothes; Herman’s refinement is that institutional investors already know the data, “it just has not received enough attention from the retail sentiment side.” Once public, weaknesses will be magnified. The disappointment will not be because of the model but because of insufficient compute—“the market’s first interpretation will definitely be that the model is bad.” GPT-4.6 was called “senile,” but “it has no compute for you to use.”
• His own crypto high-frequency trading provides experiential evidence of the compute shortage: roughly 95% of his programs may be written by AI. GPT-5.5 is better at long context and planning; “somehow, one reason may be that OpenAI has relatively sufficient compute.” His workflow is to have 5.5 plan and write the document, then hand it to 4.6 or 4.7 for execution. But “if you have more people and keep using it this way, I don’t know whether it can continue to be maintained.” Eventually, context windows and deep thinking will have to be cut.
• His conclusion is straightforward: if that correction arrives, “buy decisively, decisively,” because the semiconductor story is “just getting started… end of beginning.”

3. Semiconductors Are Forming Productive Capacity Directly for the First Time, While the Capex Valuation Paradigm Is Changing

• The fundamental difference from the 2000 tech revolution is utilization. Submarine cables then had “terribly low” utilization, and 5G base stations were only 5-10% utilized. For the past 40 years, semiconductors were merely software platforms on which people wrote PowerPoints and bought Google traffic to create “false or real productivity.” But “at least starting with 4.6, including the 5.5 I have been using recently, you are making semiconductors form productive capacity directly.” In his view, this is the third revolution after the steam engine and internal combustion.
• The second reason is collective disbelief on the supply side. TSMC has not made large-scale AI capex investments since 2023. “The capacity leader everyone watches is TSMC. If TSMC does not expand, nobody dares to expand.” ASML also does not dare to expand, leaving capacity at least 3 years behind. Token demand “is infinite on the demand side.”
• Wall Street has not updated its framework. It has always hated capex and preferred the asset-light pharmaceutical model—“turning toward the virtual, throwing the dirty work to the mainland for 20 cents, while wafers are made in Hsinchu.” But in the token-factory model, “if you do not do capex, you have no capacity.” The new scorecard is return on production: did you invest, do you have capacity today, and how quickly do you convert investment into output? ChangXin’s fab reached the roof in 6 months—faster than Taiwan.
• There are 2 ways this mindset changes. The adjustment path looks like Oracle: push up CDS, drive the stock to the bottom, force an equity raise, and then give the company a valuation once capex starts generating revenue. The natural path looks like Amazon, where return on production speaks for itself. Among neoclouds, Herman chooses Oracle: it has kept investing in capex and is already 3x ahead of CoreWeave, likely the “CoV” in the original remarks, whose projects keep getting delayed. “800 volts of power, 300 megawatts, and site selection like buying land in Monopoly—it is not that easy.”

4. Buy Optical Modules Across the Board; Packaging Takes the Baton From Process; Nvidia’s Moat Is Integration

• The optical math is simple: in an NVDA rack, procurement value is roughly 10% CPU, 40% GPU, 30% memory, and 20% optics. “The combined market caps of all optical companies are below Micron’s,” which is a fundamental reason to own the group. Herman’s urbanization analogy runs from elevators inside buildings—scale-in—to corridors between buildings—scale-up—to neighborhoods—scale-out—and cities connected to cities—scale-across. Optics keep penetrating deeper, expanding both the market and share. Google’s new GPU architecture “could increase optical demand 2x,” and each subsequent generation should require more.
• Optics have product leverage. In the 800G-to-1.6T upgrade, most components are similar and only the wafer becomes more expensive: “I do not need to expand capacity to raise my gross margin.” His summary is: “For semiconductors, enter whenever there is a shortage; for optical companies, buy across the board. Almost none of them should be too bad.” AOI is the case study: Herman first encountered it at 16, and Covid’s shutdown of Chinese capacity nearly bankrupted it. It has made real capex in recent years. The founder’s statements have been “decent” line by line; what he said in 2023 about 2025 was indeed achieved. “He speaks very confidently… so it is very easy for this company to miss expectations.”
• The next leg of compute improvement is packaging. Process improvement is logarithmic; packaging improvement is exponential. “Going forward, most of our compute gains will come from packaging.” TSMC controls CoWoS: “My wafer is baijiu; CoWoS is the bottle. You cannot sell baijiu without a bottle.” Google’s proposal to have Intel handle packaging while MediaTek sends wafer orders to TSMC “does not really work” and cannot reach scale. ASE benefits from TSMC, but packaging expansion is still “really hand-built” and slow.
• Nvidia’s underappreciated moat is integration, from ODMs—it learned the hard lessons of liquid cooling alongside ODMs—to procurement. “Isn’t HBM so scarce because Jensen bought it all up?” Like Apple before it, Nvidia’s rack-level coordination produces a large compute uplift. “That is why it is worth $5T.”

5. Memory Is a Commodity; It Can Make New Highs and Still Be Uninvestable—That Is Money From Fools

• The first principle of valuation is straightforward: fabs are valued on PB—they are “actually a real-estate leasing business,” with TSMC’s most profitable assets being old DUV lines whose depreciation has been fully written off—while memory is valued on PS. “It is difficult to value an oil company on a PE basis.” DDR is a standard product; “Samsung DDR and SK hynix DDR are the same.” HBM is simply stacked DDR, and when HBM rallies it pulls up DDR, which represents more than 75% of shipments. Gross margins rise from the 30s to above 80s, but “three-quarters of the earnings increase comes from higher DDR prices, and that is understating it.” The same thing happened during the bitcoin-mining cycle, which is why the market historically never assigned memory a PE.
• This time memory received a PE because the visible, certain shortage cycle is expected to last longer—industry participants say shortages will persist until 2030—and because a group of people who “completely do not understand semiconductors” entered the market. A pension-fund manager was squeezed out of a short, scanned Bloomberg, and said, “Hey, Micron is at 6x PE. Buy it.” Herman sees many trades around him built on that logic, and “it will not be disproven next year.”
• His positioning is explicit: “We make money from the industry, so I do not buy this.” If you make money from information asymmetry and from Wall Street investors who do not understand semiconductors, buy memory: it monetizes fastest and will first be re-rated to 10x PE. NAND and Sandisk are the same story; “I think NAND is also a commodity.” But “fools, or smart people who understand fools, can make the most money.”

6. Intel and AMD: In a Major Shortage, the Losers on the Mainline Become the Winners

• The catch-up law is clear: “When a major semiconductor shortage hits, the one that profits is always the laggard.” In normal years, nobody wants to spend money testing a PDK that nobody has used, or explain it to TSMC. Now capacity is insufficient no matter what, so customers are willing to spend. Apple might give $30B—“what is that money to Apple? It does not matter.” Intel’s real opportunity comes from TSMC’s conservatism in prior years, “not because Trump did marketing for you.”
• The original entry thesis was valuation asymmetry. Chinese DUV fabs trade at PB above 1 even before shipping a wafer, while Intel’s internal EUV yields were already in the 60s to 70s. “This is an unfair trade.” There were only 2 paths: bankruptcy, which would not happen because the pieces were valuable, or 18A working, which should take the company to $400-500B. Herman built a large position after Trump sought to remove Lip-Bu Tan: “I knew explicitly that Apple would invest in him. The market did not know; I knew.”
• Looking ahead, the 5 steps—avoid bankruptcy, internal fab use, external use, profitability, and materially taking TSMC share—“will all be completed.” 18A internal yield is 85%; externally, Tan may believe it is 65%. Yield is a one-way function, climbing 1-2 points a month. At $110-120, Intel trades at about 2x PB, versus 2-3x for 28nm GlobalFoundries and 10x for TSMC. In 3 years, a larger book value multiplied by 6-7x PB could mean “more than $1T,” before accounting for server CPUs in Huaqiangbei already rising 3-5x in price. The path will be highly volatile—“there is a 100% chance it runs into problems”—and surpassing TSMC is “a fantasy.”
• AMD is the GPU version of the same trade. Herman bought a large position at $210: “AMD is a loser in GPUs… but there is a shortage.” CUDA may be the bottleneck because everyone developed and debugged around it together; shortages are pushing customers to help AMD debug, and Meta and OpenAI procurement in Q4 may help it get there. AMD is also gaining CPU share: “If Intel is losing share and still rising that much, how could AMD be worse?” The map is: AMD is the GPU laggard, Intel the CPU laggard, AOI the optical-module laggard, and ChangXin and YMTC the memory laggards.

7. China’s Semiconductors: Huawei Took the Industry Down the Wrong Road; Memory Is the National-Level Escape Route

• Herman lays out the thesis before opening fire: China’s semiconductor industry is “very, very strong,” and Wuhan’s advanced packaging is “absolutely no worse than TSMC.” The glass ceiling is EUV, like the technology lock in The Three-Body Problem. Huawei is “the biggest troublemaker in China’s entire semiconductor industry”: Ren Zhengfei pushed domestic CPU/GPU substitution, forcing SMIC to use 14nm multi-patterning for advanced processes. Costs double, timelines double, and yields halve to the power of the number of exposures, producing “no commercial value”—an “patriotic business” sustained by the argument that because Huawei is sanctioned, nobody else can use anything else. He adds a pointed historical observation: China’s semiconductor roots and earliest capital came largely from Taiwanese investors, including the Huada and Lip-Bu Tan lineage.
• The correct strategy is memory. DUV is sufficient, DDR is a globally standardized commodity, and production can be expanded indefinitely. Zheng Shanjie, likely the NDRC chairman referred to in the captions, gave ChangXin a RMB12B loan when it was at its weakest and saved it. The current national effort is even stronger: banks can lend RMB100B at zero interest against a fab’s purchase of several DUV tools; the NDRC will not approve automotive-grade lines unless they cooperate with ChangXin, which sent a 1,000-person team to help bring them online. “They are no longer looking at economic returns.” The only bottleneck is equipment: DUV was easy to buy last year; now the whole world is buying it, and you cannot buy it either.
• The speed of monetization is extraordinary. ChangXin raised only RMB40B when it financed, yet “this year it is making RMB10B a month,” with expected annual earnings of RMB70B. “This has never happened in Chinese history, and even Huawei could not do it.” Last year DDR was still sold at a 20-30% discount, with every wafer losing money and subsidies filling the gap. This year yields have improved and are fully competitive with Samsung and SK hynix; DDR5 is no problem. Therefore, “ChangXin and YMTC are the two most investable companies.” When ChangXin lists on the STAR Market, “just buy it at the market price; it will not reach its full value on day 1.” Herman entered pre-IPO and says the filing has already been submitted. One additional comparison: Hua Hong is more attractive than SMIC. China should invest in mature nodes; trying to turn mature nodes into advanced nodes is “evil cultivation,” and evil cultivation will not succeed.

8. Crypto: BTC Is an Allocation, Not a Belief System; Altcoins Go to Zero; Circle’s Story Has Been Heard a Hundred Times

• His BTC history is unvarnished: he bought at $20-30K, sold everything at $70K, and bought back this cycle around $60-70K. If it reaches $100-120K, “I will not be excited in the slightest.” Its role is an S allocation for liquidity hedging. His family office manages only its own money, and crypto high-frequency trading has less than $200M in daily cash turnover, so liquidity is continuous; “when it falls, we can hold it very well.” But he is not “as convicted as we are on semiconductors.” If you buy, hold for years; otherwise do not hold it. “You might as well buy some semis—I can give you 2 names at random and both will outperform bitcoin.”
• His verdict on altcoins is unchanged: “I still believe all altcoins will go to zero; they are worth not a single cent.” Blockchain technology has no barrier—“I could probably generate Sui with GPT-5.5 in a few days”—and the ecosystem is “all multi-level marketing once you strip it down.” Listings require roughly 15% of the total token supply to be given to exchanges, with the only metric being “how much can be extracted.” On-chain exchanges are “street stalls next to a Macau casino”: Hyperliquid sees your order 900 milliseconds late, trades against you, and then hedges on Binance. His view of BTC’s real demand is blunt: the dark web moves $200-300B a year, all in bitcoin, according to an Israeli intelligence contractor’s 2016-17 remarks. “Do not underestimate that money.” Crypto is a “very, very manipulated market.” MSTR is one player, not the price setter; watch NAV and market beta, and “you cannot do any short-term trade.”
• Circle was challenged by listeners twice, and he dismantled it twice. “At its core, it is a money-market fund that pays no interest.” Coinbase and Binance take 50% of the money, and “the exchanges taking 99% of the stablecoin profits is for certain.” Its value rests on 2 layers of regulatory arbitrage—not paying interest, so it is not a security, plus license-based monopoly—and deep political ties: “If you believe Lutnick will remain evergreen, you can buy it.” As for the stablecoin-buying-Treasuries narrative, stablecoins at $100-200B are “a drop in the bucket” against $3.1T of daily SOFR trading and do not even reach the level worth considering. The machine-to-machine payments story is “something I have heard a hundred times; I do not believe it at all.” Decentralization only makes settlement slower; the real use case is gray-market payments, and that belongs to USDT. “For a company with this kind of competitive advantage, I might as well buy TSMC.” His standing caveat remains: “Not believing it at all does not mean you are wrong—you need to understand what kind of money you are making.” References to Lumentum, indium phosphide, and II-VI in the original captions are uncertain and should be treated as likely entities.