Hosts: Mr. Z, Victor, 168X
Guest: Herman Jin ( @ShanghaoJin )
In 2026, the semiconductor industry entered a new bull market. The continued explosive growth in AI model capabilities drove rapid increases in token demand, leading to shortages across the entire supply chain, from wafer foundries and advanced packaging to optical modules and memory. Meanwhile, US macro liquidity remained ample after the SLR was lifted, and systemic strategies such as CTAs quickly reversed their large-scale sell-off at the end of March and began buying, further fueling a strong rebound in semiconductor stocks.
In this episode of 168X, we invited Herman Jin ( @ShanghaoJin ), former executive at Goldman Sachs' FICC (Fixed Income, Currencies & Commodities) and founder of Zen Family Office, to discuss macro liquidity, the semiconductor industry chain and AI computing power demand, and the logic behind Bitcoin, stablecoins, Intel's catch-up strategy, as well as China's semiconductor and memory strategy.
Herman Jin currently runs Zen Family Office, managing his and his partners' personal funds. They engage in high-frequency trading in the cryptocurrency market and long-term investments in the stock market, using the cash flow generated from high-frequency trading to continuously invest in long-term assets such as semiconductors. Herman was also one of the earliest investors to recommend buying Intel, possessing deep industry knowledge and a strong network in the semiconductor sector. His core assessment of the demand for AI computing power is: for the first time in 40 years, semiconductors have directly created productivity; this is not a software revolution, but a transformation on par with the steam engine. There will be shortages of light and storage, and these shortages will persist for a very long time.
Table of Contents:
I. Liquidity Outlook for 2026: Interbank funding is ample, but a correction is possible by the end of the year.
II. Why semiconductors are directly contributing to productivity for the first time in 40 years
III. Investment Logic for Optical Modules: Acquire Market Share, Increase Gross Profits, Buy Everything You Can.
IV. CapEx Mindset Shift: Wall Street's Mindset Hasn't Been Moved Yet
V. AI Model User Experience: Claude 4.6 vs GPT 5.5
VI. Bitcoin: A liquidity hedging tool, but all altcoins are worthless.
VII. The True Value of MicroStrategy and the Stablecoin Circle
8. Intel's Logic for Catching Up: In an Era of Shortages, Laggards Actually Benefit the Most
IX. TSMC, Advanced Packaging, and the Real Source of Computing Power Enhancement
10. The debate over storage valuation: DDR is a commodity and should not be valued using PE ratios.
XI. China's Semiconductor Strategy: Focus on Memory, Not Advanced Manufacturing Processes
12. Cloud Computing Centers: Oracle is the most committed, while CoreWeave has been consistently delayed.
13. Audience Q&A: AMD's heavy investment rationale, NVIDIA's competitive advantage and short-term market trends
I. Liquidity Outlook for 2026: Interbank funding is ample, but a correction is possible by the end of the year.
Mr. Z: You previously focused primarily on FICC (Fixed Income, Currencies, and Commodities), and also viewed assets like Bitcoin and stocks from a macro perspective. In the context of the semiconductor bull market of 2026, what are your predictions for overall market liquidity? And what are your expectations for US stocks and bonds over the next year?
Herman: Liquidity is a difficult thing to predict over a very long period of time; you have to constantly manage and monitor interest rates. Let me first share my views on liquidity, and then discuss the trend from March to now.
Actually, we mentioned last June that the SLR (Supplementary Leverage Ratio) would be deregulated, leading to relatively ample overall market liquidity. Looking back, the impact of the SLR deregulation is already evident; the SOFR (Secured Overnight Financing Rate) has remained relatively low, with daily trading volume of underlying repos around $3.1 trillion, compared to approximately $2.9 to $3 trillion last year. Liquidity this year is relatively abundant.
You can also make comparisons, for example, by subtracting the swap spread from the term premium. The swap spread represents the cost of holding a bond, essentially how much interest you can earn by holding a bond at a short-term interest rate. These spreads are currently at relatively low levels, including for five-year and ten-year bonds. This means that liquidity on the banking side is relatively ample.
In a market with ample liquidity, absorbing that liquidity in the short term would require measures such as bond issuance by the Ministry of Finance. We observed relatively low bond issuance in Q2, suggesting that market liquidity may remain ample. However, bond issuance is expected to increase after July.
However, this doesn't mean liquidity will reverse directly. On May 6th, Treasury Secretary Bessent said that the portion of bond issuance would be reinvested in repos to maintain market liquidity stability. Normally, when you issue bonds, you take money from the Bank Reserve, issue them as Treasury bonds, and then deposit them into the Treasury's account in the Federal Reserve—essentially absorbing liquidity. He's now saying he'll put it back.
I think the White House's attitude is quite clear: I don't care too much about a lot of liquidity potentially creating a bubble. My current attitude towards market liquidity is that I want it to increase; I want the market to rise. This is a signal. He didn't say it explicitly, but from his words and statements, I think that's the attitude he's conveying.
Looking back, the market didn't rise in March, so all CTA (Commodity Trading Advisor) positions began to be adjusted. CTAs are designed to buy high and sell low, and there are Short Term, Mid Term, and Long Term CTAs, along with many Vol Control Strategies, collectively known as Systematic Strategies. In the three weeks leading up to March 31st, these systematic strategies sold $180 billion globally. This led to a market sell-off. Of course, the Iran issue played a role, but the market was already stagnant, and the CTA sell-off exacerbated the situation. However, the decline was much smaller than I expected after the Spring Festival, mainly because semiconductor valuations were simply too low. For example, NVIDIA's earnings grew by 70%, but at $180, it was essentially the same as $90 last year. The entire semiconductor sector was suppressed and failed to rise before April.
However, starting from March 31st, CTAs began to reverse course. I divide this round of gains into two phases: The first phase was "forced buying." CTAs bought 40 billion in the first week, achieving the highest efficiency and the most aggressive price increase because the initial investment was most effective. They bought 80 billion in the second week and over 30 billion in the third. After establishing momentum, CTAs became less important within two to three weeks, no longer significantly impacting the market. The second phase was the entire market's leveraged FOMO. The market truly reflected two things: first, the low valuation of semiconductors; and second, the popularity of Anthropic at the time, with everyone knowing the model's effectiveness, and we believed AGI had been achieved. The most hyped-up move was the initial surge in semiconductors.
Moreover, despite the rising Treasury bond yields—which are partly pricing in inflation expectations, rising oil prices, and geopolitical uncertainties at the long end—interbank liquidity itself, as seen in SOFR and the Financial Conditions Index, remains quite ample. Market liquidity remained ample for the following weeks.
Incidentally, the Strait of Hormuz remains closed to Iran, which is also a risk factor.
Then, I'd like to add a few points about the Fed. I still believe that, at the end of the day, if you look back at the US cycle from 2008 to 2012, banks gradually increased regulation while continuously implementing QE to release liquidity, but the multiples of banks decreased exponentially. This process took a very long time. Now, they may be considering a second path, including the Fed's mentioned interest rate cuts and QE, followed by deregulation. Because the SLR was suddenly deregulated, especially with the direct removal of Treasury bonds from the Risk-Based Capital component, banks can directly invest in Treasury bonds and engage in collateralized lending, increasing the overall flow of money within the interbank market. This is a multiple effect; the expected increase in bank exposure was $4.5 trillion, but the actual increase may be between $1.5 trillion and $2 trillion.
Mr. Z: So short-term liquidity is not a problem, but what about the future?
Herman: I'll give you a hypothesis that I've been thinking about: there might be a meaningful pullback, not immediately, but perhaps at the end of this year or the beginning of next year.
Wall Street hasn't really believed in AI yet. The entire semiconductor industry, I think, only started to believe in AI last year; before that, they weren't very skeptical either. So where does the final skepticism lie? The final skepticism lies with Anthropic and OpenAI. Are they playing both sides? That is, their money is raised through financing, and then that money is spent on CSPs (Cloud Service Providers). But the free cash flow for the hyperscalers that are actually spending that money is starting to decrease.
You can look at CDS (Credit Default Swap). The growth in 2024 was the healthiest the stock market has seen in the past two decades. Back then, Microsoft's CDS was on par with banks, even lower, practically zero. But now, the overall CDS averaged over 15 basis points. All Hyperscalers, starting with Meta, are seeing their CDS gradually rise.
Is it possible that a correction is brewing? I think it's possible. Because at that point in time, it will coincide with rising oil prices leading to a weakening economy. Oil prices are essentially no different from tariffs; they're like a tax being levied. This tax will affect inflation for at least six months, and then the chain reaction will spread to PCE and GDP. If the economy doesn't look so good, employment declines, and large companies experience a resonant free cash flow—I think the probability of such a resonant reaction is quite high—then there will be a renewed round of skepticism towards AI.
If this pullback occurs, it won't last just a day or two, or even a week or two; I think it will take at least several months. But if it does happen, I still think it's crucial to buy in. Because I believe that the semiconductor industry and the entire AI sector are still in their infancy, just beginning, the end of the beginning, only just finishing the first phase.
Mr. Z: You mentioned IPOs. There are several big IPO deals this year, such as OpenAI, Anthropic, and SpaceX. Can the US stock market absorb the liquidity from these IPOs, or will it explode?
Herman: If it's going to explode, the timing overlaps with several things I just mentioned, and the Strait of Hormuz hasn't even been opened yet. We've never experienced such a large IPO, plus there might be existing shareholders cashing out later. But at the same time, if OpenAI and Anthropic think their stock prices will fall after listing, they might not go public. They're not that short of money right now.
There are two reasons why the market is relatively safe right now: First, TSMC is the gatekeeper of the entire semiconductor industry; its lack of capacity expansion leads to high profit margins for everyone. Second, Anthropic and OpenAI are not publicly listed, so you can't use their financial data to question whether they are "stepping on their own feet." Moreover, the models are indeed very good, and revenue is constantly increasing, and increasing rapidly. The entire US stock market is watching this segment of the semiconductor industry being hyped, and this process is actually the most comfortable.
But if we assume that Anthropic and OpenAI both go public, each worth two trillion dollars, and if one of their models fails to converge properly, or if they don't meet expectations due to insufficient computing power, causing their stock prices to fall, it could cause everything to collapse together.
Mr. Z: So it's a kind of hazy beauty. It might be the sexiest when you haven't taken off your clothes yet, but once you're naked, it might just explode?
Herman: These deals are no different from semi-public IPOs. This data is public knowledge to many institutional investors; everyone knows it. It's just that it hasn't received enough attention from retail investors. When it does receive enough attention, the flaws will be magnified. And expectations will keep rising because Anthropic has performed so well this year. With such high expectations, it will inevitably be disappointed. The reason for the disappointment isn't a flawed model, but insufficient computing power. But the market's first interpretation will definitely be that the model is flawed. For example, you might say Claude 4.6 is malfunctioning, but it's not malfunctioning; it simply lacks the computing power available to it.
Not only these two, but SpaceX is also going public. The three companies will absorb a lot of liquidity, and we certainly face that risk. However, on the other hand, interbank liquidity is still relatively ample. The pool is deep, but you have to withdraw a lot of water, and what happens after that might resonate with various current market factors. Therefore, the probability of further turbulence is quite high.
Then I'd like to share my own experience. When market sentiment picks up, you might feel valuations are a bit high, emotions are too high, and you might want to exit some positions. I've actually sold some too, I've been selling continuously, and I've missed out on quite a bit. But I haven't sold everything. Never short semiconductors when liquidity might be good. Before June, liquidity was still acceptable; overall interbank liquidity was definitely still good.
II. Why semiconductors are directly contributing to productivity for the first time in 40 years
Mr. Z: Since we're on the topic of semiconductors, could Herman share his thoughts on why he's so optimistic about them?
Herman: I'll start from the top. Over the past 40 years, starting with the technological revolution of 2000, what's the difference between the technological revolution of 2000 and now? The utilization rate of technology in 2000 was very low; now it's very high. Back then, the utilization rate of submarine fiber optic cables you laid was extremely low. Even now, the power utilization rate of 5G base stations is only 5% to 10%, so its capacity isn't actually that strong.
The last wave of internet technology revolution relied on semiconductors, but advancements in semiconductors didn't directly generate productivity. Instead, semiconductor advancements transformed into software platforms, allowing people to develop software and apps. The productivity generated by those using these software and apps could be either illusory or real.
For example, you design something, outsource the manufacturing to China, and they'll do it for you for two cents. But then you buy Google traffic, do a lot of marketing, and create a bunch of PowerPoint presentations—all of that adds up to "value." All that value adds up to $100, but fundamentally, all semiconductors don't directly produce productivity.
However, starting with Claude Opus 4.6, and including the GPT 5.5 I've been using recently, you've made semiconductors directly a productive force. Achieving this productive force is like the steam engine revolution. We've only seen this twice before: the steam engine revolution and the internal combustion engine and oil revolution. This is based on data: the efficiency of a steam engine directly correlates to the amount of clothing it can produce—it's all data.
Semiconductors have essentially become a direct productive force, for the first time in the past 40 years. This is the first reason.
The second reason is that the entire semiconductor industry, including TSMC, doesn't believe in AI. Therefore, from 2023 to the present, it hasn't made large-scale investments in AI development. From TSMC's wafer fabs to packaging—which is essentially all done by hand—capacity cannot be increased. This includes memory capacity, photonics capacity, and even the final raw material capacity; none of these have increased.
The reason it hasn't been upgraded is that the entire semiconductor industry is driven by the top players, and the top player's production capacity is TSMC. If TSMC doesn't expand, no one else dares to. TSMC has been relatively conservative over the past three years, resulting in a capacity lag of at least three years. Even today's investment is far from sufficient. This is because we see that the real demand for tokens is limited to their token-related functions, which is essentially unlimited on the demand side. However, on the supply side, including TSMC and everyone else combined, capacity supply remains relatively conservative. The capacity invested in today will also take a long time to materialize.
III. Investment Logic for Optical Modules: Acquire Market Share, Increase Gross Profits, Buy Everything You Can.
Mr. Z: So, going back to the question we just discussed, why are you so optimistic about optical modules?
Herman: To give you a rough estimate, in NVIDIA's NVL72 rack, the CPO (Consumer Product Owner) accounts for about 10%, GPUs for 40% (procurement amount), storage for 30% (because storage prices have risen significantly and will likely rise even higher), and optical modules for about 20%. However, the combined market capitalization of all optical module companies is less than Micron's. This is a fundamentally unreasonable situation.
Of course, Micron uses a lot of DDR components; over 75% of its shipments are DDR. And I don't think DDR deserves a PE valuation. Therefore, based on a reasonable valuation, both the market share of optical modules and the overall market capitalization need to increase.
Let me put it in a less technical way: imagine you're building a data center right now. This represents a massive scaling improvement, much like a city undergoing urbanization. During urbanization, you need buildings; as buildings get taller, you need elevators—that's scale-in. The passageways between buildings are scale-up, the connections between blocks are scale-out, and the connections between cities are scale-across. All of these things require light.
From Scale-Across, which inherently utilizes light, to Scale-In and Scale-Up, light continues to expand its applications, becoming increasingly specialized. Because previously, copper was used for short-range applications, light is now taking market share away from copper. Therefore, it's not only growing in the market but also gaining market share, making it a business that absolutely benefits.
Take Google's new TPU solution, for example; it may double the demand for optical modules, and this will increase with each subsequent generation. It's like infrastructure; it will continue to grow.
Both optical and storage will face shortages, and these shortages will persist for a very long time. At least for now, unless you invest for five or ten years, you'll likely see a market that isn't a seller's market. When both are in short supply, I would choose optical over storage. The reason is that storage still has a strong commodity aspect; it's difficult to value it like an oil company with a PE ratio. But with CPO, GPU, and optical modules, each product is different; they have product characteristics and iterative development.
Another advantage of this technology is that, for example, during the iteration from 800G to 1.6T, the gross profit margin increases with each iteration. From 800G to 1.6T, revenue more than doubles, but some of the purchased components are similar; only the wafers may have become more expensive, so the gross profit margin continues to improve. It doesn't need to expand production capacity to increase the gross profit margin; as a product, I increase the gross profit margin and increase revenue, and the growth is rapid.
Therefore, my attitude towards semiconductors is "to the point of shortage," meaning there should be a shortage of everything. As for optical module companies, my approach is "to the point of buying everything," meaning if you buy everything, things shouldn't be too bad.
Mr. Z: If we delve deeper, what are your thoughts on companies like AAOI or Lumentum that do CPO?
Herman: It's difficult for me to recommend individual stocks. For example, with AAOI, our cost price was so low that it would be too much pressure for me to publicly express my bullish or bearish opinion at the current price. AAOI actually has a deep connection with us. Back when we were working on CB, we had contact with them when it was 1.6 yuan. At that time, the market value was only 90 million yuan, and they were almost bankrupt because their production capacity in China was shut down.
In the semiconductor industry, a common way to evaluate a company is by whether it has started production and conducted CapEx transactions in the past few years. Wall Street dislikes CapEx transactions, so if you're conducting CapEx transactions, everyone will definitely push down your valuation. AAOI has indeed conducted a lot of CapEx transactions in the past; it may have a factory in Ningbo and another in the United States.
As for this company, all I can say is that every word its founder says is decent. What he told us in 2023, we didn't meet in 2024, but we did in 2025. His direction wasn't wrong; he just tends to be very confident. There's definitely a lot of room for error in the production and new product introduction process, so he often misses expectations—that's just his personality. But the direction is definitely correct.
IV. CapEx Mindset Shift: Wall Street's Mindset Hasn't Been Moved Yet
Mr. Z: You just said that Wall Street used to dislike CapEx, but now they want to see how efficient your capital utilization is and how quickly you can generate returns after investment?
Herman: Yes, I think there's a trend everyone needs to pay close attention to: many of the top hyperscalers in the US, including Google, have already realized that our lightweight model will be changed. In the future, we'll have to use the token factory model, the one Jensen Huang talked about.
Previously, I used software to develop the software and then sell countless patterns. But now, I'm constantly selling tokens through inference with users. You've become a factory, so you absolutely must do CapEx. If you don't do CapEx, you have no production capacity.
This is one of the reasons why I think there might be a significant market downturn at the end of this year or the beginning of next year; Wall Street's mindset hasn't fully shifted. Too many CapExs create problems with free cash flow. But those who truly understand the market, like Larry Ellison, are making very determined investments because they know this is the future.
After adopting a factory model, all companies look at several points: First, have you made any investments in the past, do you have the production capacity today, and can you sell your products? Whether you can make money in the future depends on your return on investment.
For example, at our wafer fab for chip manufacturing, it might only take six months from capping to machine installation, even faster than in Taiwan. But in Phoenix, Arizona, where there's no infrastructure, how long would it take? Therefore, the speed and ratio of production to output determine the company's future value.
Mr. Z: So how will this transition period proceed? What event made Wall Street realize that "now is the time to CapEx"?
Herman: I think there are two types. The first is an adjustment, which is a drop. After the drop, it's discovered that these CapEx companies with good valuations are undervalued. For example, Oracle's cash flow will improve between 2025 and 2028. You don't have it now; you're doing CapEx. I'll first inflate your CDS, then drive down your stock price, forcing you to raise funds with stock at the bottom. After that, the price can't be driven down further, and then you realize the performance is still good, so it rebounds.
The second type is natural, like Amazon, where the market itself looks at your return on investment, whether you've gained market share, and whether you've made money.
I'm not saying there will definitely be this adjustment this year—I never said that. But there might be an opportunity. I say "an opportunity" because I'm very committed to the idea of semiconductors directly becoming productive, and I think it will take a long time.
V. AI Model User Experience: Claude 4.6 vs GPT 5.5
Mr. Z: Herman mentioned AI models. What are your impressions of the few models you've used so far?
Herman: I think GPT 5.5 is relatively easier to use, especially in long contexts, where it performs planning better. I think one reason for this is that OpenAI has sufficient computing power.
Claude 4.6 worked very well when we first started using it. Because we use it directly for production-level tasks, such as quantitative high-frequency trading, we get immediate positive feedback on order placement speeds (two milliseconds vs. 1.4 milliseconds) and concurrency issues. We see immediate positive feedback in our trading, with daily trading volumes reaching hundreds of millions of dollars, or even more. Any issues with trades or backend maintenance are all handled by AI. Currently, about 95% of our program is written by AI.
My feeling about versions 4.6 and 5.5 is that 5.5 might be slightly better in context. If you want to use 4.6 or 4.7, you should first use 5.5 to complete the planning, write it into a document, tell 4.7 all the skills and MDs to be called, and then let 4.7 write them. The effect is also acceptable.
Version 5.5 is indeed impressive, and I've found it to work very well. However, I don't know if it can be sustained with a larger user base, because ultimately the model depends on the output token, and it might cut some context or Deep Thinking functionality. This is unavoidable due to insufficient computing power.
VI. Bitcoin: A liquidity hedging tool, but all altcoins are worthless.
Victor: Herman mentioned earlier this year that the surge in gold and silver prices was driven by a geopolitical bubble, but stocks and the US dollar were fine, and BTC wasn't afraid of a bubble. From the perspective of working in the Family Office, how do you view BTC as part of an asset allocation strategy?
Herman: Let me share my thoughts on BTC and what we're doing. Our Family Office manages personal funds—the personal funds of our two partners. We do high-frequency trading on crypto and invest in stocks. Our daily trading volume on Binance is currently less than $200 million, much less than before, because liquidity is so poor right now. We trade on every exchange. Because it's our own money, we're more aggressive in our investments; we're more willing to buy when BTC prices drop, and we're more willing to hold onto our positions. This includes a lot of stock investment, because we make money every day from high-frequency trading, and we keep investing in stocks. I don't care if stocks drop, because we don't need the money.
So, when you're investing and managing a portfolio, you can't resist the urge to sell or reduce your holdings to mitigate drawdowns. But we don't have that need. We essentially have a cash flow plus a deposit, with the deposit simply held in some risky assets. We hold on very well when prices fall. Sometimes we're tempted to sell when prices rise, but we hold on very well when prices fall.
I actually bought Bitcoin quite early on, around $20,000 or $30,000. But I sold it all when it reached $70,000 and then mostly invested in US stocks. Later, when it dropped to below $60,000 or $70,000, I bought some more, but haven't touched it since. Those Bitcoins can also be held in the account as collateral; they're just sitting there.
You're saying I'm extremely bullish on Bitcoin and that it will definitely do this or that? I'm not that convinced. I think Bitcoin is a liquidity hedge in the long run, an asset worth watching, but I'm not as committed as I am to semiconductors. But you can't put all your money in semiconductors, so you put some in Bitcoin.
Why is it okay to buy Bitcoin? First, it's relatively unchangeable, and you can't lock my account. USDT allows account locking, and Ethereum doesn't have as high acceptance. All the money I know of among criminals is in Bitcoin. Don't underestimate this amount of money; it's easily two or three hundred billion a year. I was talking to a contractor for an Israeli intelligence company who was collecting data on the dark web for Mossad. He bought countless Bitcoins in 2016 and 2017. He said all dark web transactions were in Bitcoin, and while some have been transferred to USDT, Bitcoin remains the most secure option.
So, can Bitcoin be directly replaced? No. That money won't go into a bank account. Bitcoin has value.
As for all altcoins, I still believe they will all go to zero, worthless. I got a lot of flak when I first said this back in 2020. Where is the technology behind all blockchains? If you publish the code, GPT 5.5 can probably generate one for you in a few days. The difference lies in the people involved, whether or not there's a pyramid scheme team helping you recruit. If your foundation is this, then there's no barrier to entry.
Victor: You mentioned before that Solana and BNB are different from other altcoins. What's your opinion now?
Herman: I'm not saying they have cash flow, I'm saying they have bookmakers who will pump and dump. These are casino chips, and whether a casino does well or not depends on whether you think it can do it well. But essentially, you have to have some competitive advantage over others.
Let me tell you what the crypto ecosystem is like. If you want to list on Binance, you have to give 15% of your circulating supply to the exchange; it's pretty much the same on every platform. Most other coins are locked up; maybe a third of the active coins are held by exchanges. The core metric for every listing is how much you can profit from, and which exchange you can profit from. How much can you profit from listing on Perp? How much can you profit from listing on Spot? It's all the same.
Then, because of its strong monopoly, some people started setting up makeshift trading platforms, and all those on-chain exchanges are like those makeshift platforms. For example, we now trade on Hyperliquid, which has a 900-millisecond latency. If a platform sees your order and the market moves against you, it will execute the trade; if it moves in your favor, it won't execute the trade and will hedge on Binance. All these so-called on-chain exchanges are pretty much the same.
So I don't buy anything except Bitcoin. I'm not Bitcoin Maxi, I'm Semi Maxi. Bitcoin is an allocation for me.
VII. The True Value of MicroStrategy and the Stablecoin Circle
Victor: What's your take on MicroStrategy? You mentioned before that they could hold out indefinitely, selling BTC when NAV falls below one and raising funds by selling stocks when it rises. But recently they've been hinting at testing the waters by selling BTC?
Herman: MicroStrategy is merely a player in the BTC market, not a dominant player or a price determinant. The crypto market is a highly manipulated market, manipulated day by day, hour by hour, and minute by minute. Liquidity itself has a directional impact, but some people manipulate the market at short, medium, and long-term levels. All major markets are concentrated areas of manipulation.
MicroStrategy offers two parameters: NAV and Market Beta. You need to understand that this market ultimately has fundamental demand and can indeed hedge some liquidity, but it's constantly being manipulated. Therefore, you can't engage in any short-term trades. If you want to buy Bitcoin, your holding period should be at least in years; otherwise, don't hold it at all. You might as well buy some semiconductor stocks; I can randomly give you two that will outperform Bitcoin.
Victor: What about stablecoins? What's your opinion on stocks like Circle?
Herman: Right now, a lot of people around me are hyping Circle. I feel like there's a divide, just like there's a huge divide between the semiconductor industry and the non-semiconductor industry.
Circle is essentially a non-interest-bearing money market fund. 50% of its funds go to Coinbase. This means that, to date, all stablecoin payment scenarios rely on exchanges. Binance's terms with stablecoins are that Binance takes 99% of the profits, for a certain period.
What is Circle's competitive advantage? Its value comes from two parts, both of which are policy arbitrage. First, once you pay interest on a token, it becomes a security, subject to SEC regulation. So if I don't pay interest today, it's not a security, and I pocket the interest. Second, I have the license, others don't. Would you be willing to invest long-term money in a company that profits from policy arbitrage? I certainly wouldn't. Moreover, Circle is deeply intertwined with politics; just look at its shareholder list.
(Audience follow-up question: If the US elite and Wall Street are pushing for a stablecoin boom, won't that weaken Circle's competitive advantage? This includes stablecoins used in Southeast Asian cross-border payments, AI agent economies, and 24-hour trading. Also, during the conflict with Iran, while other futures markets were shut down over the weekend, trading volume on Hyperliquid exploded. Wall Street itself has also discovered the advantages of using stablecoins and blockchain for financial infrastructure, such as 24-hour trading, rapid confirmation, and settlement. Furthermore, Coinbase has implemented the x402 protocol on AWS Bedrock. Many people aren't speculating on Circle's reserve returns, but rather on these larger trends.)
Herman: I've heard these stories 100 times, and I don't believe them at all.
First, you mentioned using stablecoins for machine-to-machine transactions. This increases the ease of circulation, not the transaction scenario. The effect is no different from Alipay and WeChat Pay. Paying to your Binance account via Binance Pay is definitely the fastest. Centralized ledgers are always faster; decentralization simply makes things slower.
Secondly, you mentioned using stablecoins to buy US Treasury bonds. The total size of stablecoins is probably between $120 billion and $200 billion. SLR released $4.5 trillion in this wave. SOFR's daily trading volume is $3.1 trillion. Your money is just a drop in the ocean and doesn't even reach the level they're considering.
Third, the only real advantage of blockchain currency is in gray payments. Black payments use Bitcoin, gray payments use USDT. I agree with that, but I don't agree with your Competitive Advantage. Why Circle and not USDT? Why can't Trump's son create another one? If there's no need for regulation, a Singapore license would also work.
If I were faced with all these problems, I definitely wouldn't believe Circle's story. I'd rather buy TSMC stock.
8. Intel's Logic for Catching Up: In an Era of Shortages, Laggards Actually Benefit the Most
Mr. Z & Victor: Herman was basically one of the first people online to advocate buying Intel, predicting a five-fold increase in three years by 2024, and that prediction has now been met. Could you analyze your thinking back then?
Herman: At the time, it was about valuation. Intel had invested in so many wafer fabs, while some Chinese wafer fabs using DUV process equipment hadn't even produced a single wafer yet, but their PB ratio was already 1. Intel's EUV yield was internally known to be around 60-70%, but its PB ratio was 1. I think that's unfair, an unfair trade.
Whether Trump comes to power or not, the outcome will be the same. Intel only has two paths: one is bankruptcy, which won't happen, as it would be valuable even if dismantled. The second is that the 18A series will come to fruition, and if it doesn't fail, it should be worth four or five hundred billion. So that's how I saw it back then, and that's how I bought it.
The major position was established around the time last year, when Trump was trying to oust Lip-Bu Tan (Chen Liwu). At that time, I knew for certain that Apple was going to invest in it, and I knew for certain that Apple was having its chips manufactured there. This was industry information; the market didn't know, but I did.
Mr. Z: So what's Intel's perspective now?
Herman: There are too many consumer goods now. The news of Apple's investment has already come out, and Apple may provide more R&D funding later, but basically, everything else that can be consumed has already been consumed. Next, it will have to fulfill those promises.
Intel's problem isn't a lack of technology; in fact, their technology is quite strong. However, none of their major customers have ever developed on their new PDK before. Giving them something new to develop with is bound to cause various problems. Furthermore, Intel has historically been an internally focused company, not a very client-oriented one.
But I think there are four or five steps to climb:
Not bankrupt
Wafer fabs can supply internal use
Can be used by outsiders
Wafer fabs can be profitable
It can substantially pose a partial challenge to TSMC.
All five stages will be completed. No matter how strong Taiwan becomes, and I know Taiwanese people look down on Intel, in the semiconductor industry, once a major shortage occurs, the one who profits is always the lagging follower. It's like the massive shortage of optical modules benefited AAOI. Now, with the massive shortage of wafers and packaging, Intel benefits. Because previously, no one would use your PDK, but now, with insufficient production capacity, customers are willing to pay to try it. Apple can only provide a maximum of 30 billion, which is nothing to Apple; providing resources and money could potentially have a much greater capacity effect.
Looking back at its valuation, it's now over double its price-to-book ratio (PB), but GlobalFoundries is also over double its PB. We normally value TSMC at seven times PB, but now it might be ten. Intel is one of only two companies in the world capable of producing 2nm chips; Samsung doesn't count, as Samsung has proven its yield rate is only 30%. Giving Intel double its PB is reasonable, in my opinion.
There will be ups and downs, going from two times to zero, zero to three times, three times to five times and back to two times, back and forth. But you know, after three years it will definitely reach six or seven times. Because yield is a one-sided function, I continuously improve my machines and debug my processes, and the customers develop on top of that; it's a two-way collaboration.
Mr. Z: I think it's also related to culture. To put it bluntly, we Taiwanese are quite servile and have very strong livers. On this small island, what are our advantages? Only semiconductors are left.
Herman: Yes, your advantage is that you have better livers. Our chip manufacturing system, which involves working non-stop, is also quite good, with a six-month cap, which is terrifying.
It currently has a book value of $108 billion. Investing $15 billion to $20 billion annually, the book market will grow significantly in three years. Multiply that by six to seven times the petabyte (PB), plus CPUs—and we haven't even factored in the shortages—CPU prices in Huaqiangbei have already increased three to five times and are still rising.
Moreover, Intel's business ratio is roughly 1:8, with data center computing power and PC computing. The gross margin for data center computing can reach 70%, but PC CPUs are barely selling, with a gross margin of only around 30%. If you assume they could also achieve a 70% gross margin for PC CPUs, Intel would skyrocket, a tenfold increase. But that won't happen; PC CPUs are currently unsellable. However, considering only the data center segment, plus the PB revaluation of foundries, the valuation is definitely going up over the next three years. There will absolutely be various problems along the way; it would be strange if the 18A didn't encounter any issues.
Mr. Z: Some people have been saying lately that Intel can surpass TSMC?
Herman: That's a pipe dream. We won't see it in the short term. I'm very optimistic about Intel in the long term, but surpassing TSMC is a long way off.
However, Intel has proven that yield can be improved, while Samsung has proven that it cannot. The internal yield for the 18A is around 85%, in the 80s. The external figures are somewhat unpredictable; Lip-Bu Tan might think it's 65%, while his subordinates might think it's 60%. But whatever, it's highly probable that the yield will climb. Basically, it will improve by one to two percentage points per month, two percentage points in months with faster progress, and one percentage point in months with slower progress. Looking ahead, the result will still be a valuation. Even if Intel doesn't have a 2nm wafer fab today, only 3nm and 5nm, with yields above 90%, it's still valuable. I can't get TSMC at 10x PB, but 5x PB is possible. So it's not about replacing TSMC, but about being able to substitute for some of TSMC's capacity. If it reaches the fourth level, its market capitalization will be over 1 trillion, or even more.
IX. TSMC, Advanced Packaging, and the Real Source of Computing Power Enhancement
Mr. Z: TSMC's PE ratio is only 20, why is it so low?
Herman: Wafer fabs should be valued using PB (price-to-book ratio), not PE (price-to-earnings ratio). It's like I'm in the real estate leasing industry; wafer fabs are essentially a real estate leasing industry. TSMC's most profitable wafers aren't EUV, but DUV, and specifically the older DUV production lines. Those have long been depreciated and are worthless, yet they still generate a large amount of cash flow for the company every year.
Next, I'll explain the sources of computing power improvement. Previously, computing power improvements were all driven by wafer technology: 7nm, 5nm, 3nm, 2nm. However, the improvement rate of wafer technology is decreasing, becoming a logarithmic function. Meanwhile, the improvement in packaging technology is an exponential function, meaning future computing power improvements will primarily come from packaging. This is why TSMC is focusing on CoWoS as its core technology.
You can think of it this way: my wafer is selling liquor, and my CoWoS is the bottle. You can't just sell the liquor; you have to include the bottle to make it valuable.
For example, Google is becoming increasingly shrewd with its costs. Once it enters CapEx mode, CapEx efficiency translates to cost control. Google's ideal scenario is to place orders directly with TSMC like Apple does, but TSMC can't handle that many orders. Google has also made a rather outrageous request: to have Intel handle the packaging. This involves placing wafer orders with TSMC through MediaTek, and then having Intel handle the packaging of those wafers. Essentially, Google is taking TSMC's liquor and putting it in Intel's bottle.
I don't think this issue will lead to significant volume. TSMC won't allocate that much capacity to you. There will be a whole host of problems with allocating production lines. The market is currently speculating about this, so I won't contradict it, but you need to understand the essence of the matter.
Packaging equipment is what the market will truly lack. Take ASE (Advanced Semiconductor Engineering Co., Ltd.) for example. Backed by TSMC, once TSMC's production line was running smoothly, they could directly compete with each other. Packaging requires a lot of manual operation; it can't be fully automated. It's really hand-made, so expanding production capacity will be very slow.
10. The debate over storage valuation: DDR is a commodity and should not be valued using PE ratios.
Audience member Bobo ( @bboczeng ) asked: You explained why DDR is a commodity. But what about NAND? Does your point fully apply to NAND companies?
Herman: NAND is also a commodity. But the opportunity in the market now is this: it's not that smart people like us can make all the money; it's possible that fools, or smart people who understand fools, can make the most money.
Do you know the relationship between DDR and HBM? DDR5 stacked together becomes HBM. When HBM prices rise, it also drives up the price of DDR raw materials. Therefore, three-quarters or even more of a storage company's profits come from the price increase of DDR, not HBM. This overall increase in gross profit should be valued like selling oil. But that's not how people value it now.
Why? Because the memory market is currently attracting investment from people who know absolutely nothing about semiconductors. Their first instinct is to look at the price-to-earnings ratio (P/E). Seeing Micron's 6x P/E and other high P/E ratios, they buy immediately. This group consists of traditional fund managers and even pension fund managers with only a finance background. In the first two weeks, this influx of money significantly drove the market up.
Is there still that money? Yes, there's still a lot. Many of my former colleagues, who didn't have the opportunity to learn about semiconductors, all thought Micron was cheap. But I think Micron is very expensive.
You can't use PE to value oil companies. If oil prices rise to $300 tomorrow, can you still use PE? What if they fall to $30? So you either have to value them based on reserves or on price-to-book ratios (PB). Storage companies used to be valued using price-to-sales ratios because margins fluctuated too much.
Why don't I buy storage? If you're making money from the industry itself, I'd rather buy optical modules. Optical modules might not necessarily rise less than storage. But if you're making money from information asymmetry, and you know that those on Wall Street who don't understand semiconductors are still coming in to buy, then buy storage; it's the fastest way to realize profits.
XI. China's Semiconductor Strategy: Focus on Memory, Not Advanced Manufacturing Processes
Mr. Z: You just mentioned that China should develop storage technology?
Herman: China's semiconductor industry is extremely strong, and I'm about to get criticized again. China is very advanced in advanced packaging; the advanced packaging technology in that Wuhan lab is definitely on par with TSMC. But the problem is, you're packaging, but without a chip, what are you packaging? Why don't you have a chip? Because you don't have EUV.
China's semiconductor industry has been led astray. Ren Zhengfei stated that China must have domestic alternatives, specifically for CPUs and GPUs. However, CPUs and GPUs utilize the advanced EUV process. What to do when you don't have EUV equipment? The solution is to have SMIC cooperate by repeatedly exposing 14nm chips, doubling costs, doubling time, and halving yield. It has no commercial value. But it doesn't matter, because Huawei is doing patriotic business; being blocked, China must use Huawei's products.
Mr. Z: So where is the boundary between DUV and EUV?
Herman: DUV is for 14 nanometers and above, while EUV is for below 7 nanometers. Why can't China compete with Taiwan? It's not that China isn't strong; China's semiconductor industry is very strong, but it's stuck at the EUV glass ceiling. It's like the Trisolarans have blocked your technology, preventing you from advancing to this level. The EUV machine is that blockade.
Furthermore, your repeated exposure of developing advanced manufacturing processes involves seven or eight steps, going back and forth repeatedly. The yield rate becomes an Nth power. For example, if the yield rate is over 70% once, multiplying it by two yield rates drops it to 50%. This results in longer processing times, more machines required, lower production capacity, and higher prices. It is fundamentally uneconomical.
Actually, China should pursue the storage route. China's strategy is changing; the National Development and Reform Commission's direction is to replace CPUs/GPUs with domestically produced storage, using storage to squeeze out the global market. Storage can be done with DUV, and DDR is a standard product; Changxin's DDR is no different from Samsung, SK Hynix, or Micron's.
To put it bluntly, all the semiconductor companies in China were invested in by Taiwanese investors, specifically those affiliated with TSMC. The earliest imitations were all Taiwanese investments. Even companies like AMEC (Advanced Micro Devices), which were initially invested in by Lip-Bu Tan at Walden International, were all backed by Taiwanese investors. I have an older brother who took over Walden's secondary fund; much of our semiconductor knowledge came from him. He took over Lip-Bu Tan's Walden secondary fund and ultimately made a fortune. Essentially, all these resources and money originated from Taiwan.
Changxin and Changcun fit the follower logic I mentioned earlier—companies that almost went bankrupt. Changxin only had 40 billion RMB in funding last year, but this year it's earning 10 billion RMB a month. Last year, Samsung and SK Hynix held back from raising prices because Changxin and Changcun's DDR memory was sold at a 20% to 30% discount, and their yield rates were lower. Changxin's gross profit was zero; they lost money on every unit produced, relying on government subsidies. But this year is different. Yield rates have improved, and the premium for DDR memory in China has basically leveled off. As a result, Changxin's profitability this year will be extremely high. This is the follower logic.
Changxin Memory and Yangtze Memory Technologies are two of the most worthwhile companies to invest in. When they list on the STAR Market, you can buy them at market price. Changxin has already submitted its application, so you can rest assured.
(Audience question: We previously invested in Changxin at a valuation of 150 to 170 billion yuan, and Changcun at 75 billion yuan. What was your price?)
Herman: I invested relatively late, at the IPO price. I bought some PE (private equity) myself a couple of days ago, but I think the secondary market is better because Chinese PE firms have to pay capital gains tax, which is worrying. If you buy on the secondary market, there's no problem; you can buy directly, and it won't rise to its full potential on the first day.
The power it can mobilize is the will of the entire nation: 100 billion yuan in interest-free loans, and Changxin sending a 1,000-person team to various locations to coordinate operations. The government might even require other wafer foundries to expand their production capacity beyond automotive-grade chips; they must cooperate with Changxin to produce DDR. However, the equipment cannot be purchased quickly enough; DUVs are currently in high demand worldwide and are unavailable. But at least this path is correct, and all the resulting production capacity is profitable.
One more point: among domestic companies producing outdated processes, Huahong is definitely better than SMIC. In China's semiconductor industry, you should invest in companies producing outdated processes; you can do very well and make money with them. But trying to turn outdated processes into advanced ones is like practicing evil cultivation, and evil cultivation will never succeed.
12. Cloud Computing Centers: Oracle is the most committed, while CoreWeave has been consistently delayed.
Audience Question: In the NeoCloud sector, what are your thoughts on CoreWeave and Oracle recently?
Herman: CoreWeave is always experiencing project delays. But I pulled its data, and looking only at CapEx, Oracle is consistently investing, and very quickly. In Q4 2024, Oracle and CoreWeave weren't that far apart, but now the gap is probably three times. Oracle ignores everyone's doubts about its CDS and continues investing without any bottom line.
Oracle and NVIDIA directly correspond to the availability of tokens. Semiconductor components, optical modules, and storage are dependent on the availability of tokens. We know they definitely lack tokens, so they will invest, and therefore they will definitely lack components. But the current market pricing discrepancy lies in the availability of tokens, and Wall Street is starting to question whether they are stepping on their own two feet.
If the cloud sector is to take off, Oracle's investment is the most decisive. Looking at the past nine to twelve months, Oracle's CapEx has been consistently growing. After its CapEx is completed, it takes about six months to debug, deploy, and lease it out, and the financial data will definitely reflect this within a year.
Scaling up CoreWeave isn't easy. With your current 800-volt power supply, as the models get larger, you'd need a 300-megawatt power plant, requiring significant investment in power distribution and setup. Not every location has electricity; site selection is like a billionaire buying land. Oracle has consistently been able to invest successfully because of its experience. Microsoft isn't a small company either. But look at CoreWeave, which isn't small either, and look at how exhausted it is, constantly facing delays.
13. Audience Q&A: AMD's heavy investment rationale, NVIDIA's competitive advantage and short-term market trends
Audience question: In the entire AI industry chain, from model companies to energy infrastructure, which segment adds the most value? When there is a shortage, is the logic of "the laggards profiting" entirely valid?
Herman: I don't actually know. Not all those things that are in short supply will necessarily get a PE valuation. Recently, a lot of people have been speculating on Wolfspeed, power plants, and transformers; I think this is based on the traditional shortage logic. But can the barrier to entry for building a substation be compared to that for NVIDIA? NVIDIA's moat is much higher. If you gave NVIDIA a PE of 12 back then, and now you're giving those other things a similar PE, I don't think it's worth it. But they've also been speculated on many times over; there's randomness in the market.
Under what circumstances do underdogs win? I think it's in the main storyline. The underdog in the main storyline is more likely to win. My current large position is in AMD, which I bought at 210 yuan. AMD is a loser in GPUs; its cards don't sell. But because of the shortage, people are willing to use AMD for testing. CUDA was developed and tested collaboratively; normally, AMD could never catch up, but because of the shortage, you might be able to sell your cards.
Plus, I knew that AMD's CPU market share was increasing, while Intel's market share was decreasing but its share was still increasing so much. How could AMD not be worse than Intel? So I bought quite a lot of AMD CPUs.
Audience question: Where does NVIDIA's moat manifest itself? Are there any companies with high competitive advantages that people haven't noticed or haven't been pricing in?
Herman: Actually, the moat is very high in every sub-sector of semiconductors. For example, Changxin and Changcun have DDR5, but they can't develop HBM because they can't handle the PHY process. Each small sub-sector has various moats, so once there is a shortage of semiconductors, everyone with moats will "raise prices as much as possible," which is the current market situation.
But at the very top, we're looking at who provides value, who provides more tokens, and who provides more computing power. Aside from a few model companies, NVIDIA has undoubtedly improved the most. NVIDIA doesn't just make GPUs; it builds a whole rack, encompassing procurement, ODM assembly, liquid cooling debugging—every pitfall you encounter has to be revisited with others. Therefore, NVIDIA's advantage lies in having built a complete moat, from ODM to procurement (it bought up all the HBM and optical module production capacity), somewhat like Apple in its early days. That's why it's worth five trillion.
Audience question: Can you make a rough prediction about whether the market will experience a short-term correction?
Herman: We can't film it now. I was certain there would be a surge in late March and early April because of the reliable CTA buying. But now there's no such reliable flow.
The current Gamma Skew is quite large, with many calls during upward movements, easily forming a Gamma Squeeze. Downward Put Walls are lowering, indicating selling pressure to absorb buying pressure, resulting in significant resistance during declines. However, all walls are meant to be broken; if a wall is broken, the decline will accelerate.
It's highly unlikely to break through; an upward move would likely be a squeeze. However, the stocks are very concentrated, with many calls on Micron and Intel, primarily focused on semiconductors. For the price to squeeze upwards, semiconductors need to be pushed up; the chances for other stocks are not very high.
Audience question: OpenAI is going public this December. Is everyone just speculating on the hype? Will optical modules, storage, and CPUs cool down a bit then?
Herman: I think so. Didn't I just say that? Because OpenAI is more likely to collaborate with upstream and downstream partners, if it's caught off guard and its performance plummets, it could drag everything down with it.
Audience question: How long can the CPU shortage last?
Herman: The CPU story is the main focus, unlike storage. Storage prices rose along with DDR prices when HBM prices increased, but CPUs won't have that effect. Intel assumes a 70% gross margin in the data center sector, but it's impossible for PCs to achieve the same margin—PCs aren't even selling. Therefore, the hype surrounding CPUs can't reach the level of that in storage.
Mr. Z & Victor: Okay, that's about it for our Space segment today. Thank you so much, Herman, for sharing.
