As long as we don't enter an interest rate hike cycle, is the current pullback a buying opportunity?

Author: Mr. Z, 168X

Macro is just noise, sentiment is the signal: blindly stacking hardware in June was the peak of euphoria, no one wanted semiconductors at the end of August was the point of neglect

Taipei time, September 9, 2026, Wednesday at 10:00 AM. Last Friday's non-farm payrolls came in at 162,000, far exceeding the expected 50,000; this Friday's CPI takes the stage, the US and Iran are not reconciling, oil prices are hovering between 90 and 100, and the market is pricing in 1.5 rate hikes this year. Semiconductors experienced a decline in July without any rebound, a small rebound in the first half of August, and a second test in the second half. Today, hardware is rebounding from lows, software is falling, and the market is still running the "software down, hardware up" narrative. Is rate hiking the main theme? Should we buy hardware or software?

In this episode of the 168X War Room, we welcome back our old friend Investment Talk Jun (@TJ_Research). In this episode, he puts macro back in its place as noise: as long as we don't enter a rate hiking cycle, the market can digest two or three rate hikes; CPI itself is not important, the Fed's reaction function is what matters; the market has been wrong about Nvidia's expectations since 2023; depreciation mismatch is a false proposition, chips from three years ago now pay for themselves in one year; open-source models return pricing power to CSPs, the era of Amazon making only $4 on a $100 API is changing; token price reductions are the prerequisite for the agent explosion. A parting message for investors: Buy when there is neglect, sell when there is euphoria. This year's market has tested us several times!

1. Rate Hikes Are Not the Main Theme: As Long as We Don't Enter a Rate Hiking Cycle, Two or Three Hikes Are Just Speed Bumps

Mr. Z: Welcome everyone to 168X. Today is Wednesday, September 9th, Taipei time 10:00 AM. It's a great honor to have our old friend Investment Talk Jun back again. Let's first talk about the overall macro situation: last week's non-farm payrolls, Friday night's CPI, whether or not to raise rates, and Kevin Warsh's statements.

Investment Talk Jun: Looking at last Friday's non-farm payrolls and unemployment rate from the results, no matter how you slice it, it was a set of data that exceeded expectations. At the very least, it represents that the labor market is good, with no signals pointing to it deteriorating. Of course, judging from the trading results in the bond market, the probability of a September rate hike didn't increase dramatically after the data was released. This is mainly because, from the Fed's perspective, they have two mandates, and the labor market is usually not the focus relative to CPI: for now, it's stable, not getting significantly worse or significantly better, and they won't treat one month's data as a trend. So, the main factor determining whether to hike in September is still this Friday's CPI.

Someone asked me today whether they should hedge or adjust positions for Friday's CPI. Personally, I won't. Even if there is a rate hike this time, one in September and another in December, we still won't enter a rate hiking cycle; as long as we don't enter a rate hiking cycle, whether it's two or three times, the eventual rate hikes won't have a very significant impact on asset prices. We are not in a rate hiking cycle like 2022, at least not based on the current inflation trend. So, rate hikes themselves are not a very big main theme; they are just noise, a speed bump on the path of fundamentals and AI. Another point: the index fell at today's close, and overall market participation and sentiment, looking at the S&P, are not high. Entering Friday's CPI with low sentiment, there are only two scenarios: if the data is bad, the market falls, which is reasonable, but the downside is limited because sentiment is already low; if the data is good or meets expectations, a rebound under conditions of low participation and low sentiment is very reasonable. If it's bad, it will fall a bit until the rate hike is confirmed next week, after which the market might catch its breath and rebound. So, under the current sentiment and participation levels, there's not much need to hedge against CPI.

2. CPI Itself Is Not Important: In 2021, CPI Rose from 2% to 5%, and the Market Kept Rising Until November

Mr. Z: The US and Iran are not reconciling, oil prices are hovering between 90 and 100. How big is the impact of this on CPI? It looks like the data released on Friday will probably not be very good.

Investment Talk Jun: I've shared a viewpoint before: CPI is not important; what's important is how the Fed interprets CPI. What influences market trends is not the CPI data itself, but the Fed's reaction function. If CPI is bad, but the Fed says there's no need to raise rates, it has no impact on asset prices. The best proof is 2021. CPI rose from 2% to 5%, but the Fed remained dovish, and asset prices kept rising until November 2021, when the Fed pivoted. So, whether CPI is good or bad is not the direct trigger for asset price movements. The market will react immediately when the data comes out, but the reaction function is more important. Whether we get two or three hikes this time, whether CPI is good or bad – if it's bad, we might end up with 2.5 hikes. Looking at it from a 3 to 6-month perspective, will two or three hikes really have a huge impact on asset prices? Not that much.

Building on this: oil prices are a component of CPI. If CPI is relatively less important compared to the Fed's reaction function, then oil prices are an even less important factor for CPI. Moreover, the Fed focuses more on core CPI when formulating policy, which at least removes oil prices. Recently, I've encountered a viewpoint different from mine: if the US-Iran conflict shows signs of slowing down, ending, or pausing, oil prices would fall, CPI would fall, and the Fed might even cut rates this year. I don't quite agree with that. Oil prices have no direct impact on core CPI, and from a transmission perspective, the short-term impact won't be huge either. So, the end of the US-Iran war wouldn't have a massive impact on core CPI, core PCE, or Fed policy. Of course, if it does end, directionally it would certainly help make their need to hike less urgent, but it's not a very important factor.

3. The 2-Year Treasury Yield: Betting with the Market, 1.5 Hikes This Year

Mr. Z: I found one of your posts quite interesting. Data from past years shows that a rising 2-year Treasury yield is a leading indicator, with a high probability that the Fed's rate range will be adjusted upwards. There are signs of that this time too.

Investment Talk Jun: Yes. If I absolutely had to guess whether the Fed will hike rates this year, I would agree with the market's view: I follow what the market does, and I believe the Fed will ultimately do the same. Right now, the market is pricing in 1.5 rate hikes this year, so I think there will be 1.5 hikes this year. It's not that every time the 2-year yield rises, the Fed will definitely hike. The market could be wrong, but to be proven wrong, the data must contradict market expectations. For example, if inflation data is very good or employment data is very bad, the market will naturally push the 2-year yield down. The bottom line is: when the market is pricing in a rising 2-year yield and potential rate hikes, at the very least, the Fed cannot cut rates. The final outcome will be either maintaining rates or hiking. So, I'm on the market's side. If the market thinks there will be a hike, I make investment decisions assuming there will be one. I won't bet against the market, thinking there will be no hike this year when the market expects one.

Because, at the very least, I don't see any signs of core inflation or core PCE slowing down. In the medium to long term, at least 6 months to a year or even longer, I don't see signs of CPI slowing down: with AI CapEx being poured in, overall wage growth hasn't returned to pre-pandemic levels and is still higher than before the pandemic, much consumption remains strong, making it difficult for inflation to return to 2%. So, the best case is no hike, the worst case is two to three hikes. It's just moving between these two options. In my investment framework, for the next 3 to 6 months, macro itself is still more like noise: when the market is performing very well, it needs some negative news to sound a warning bell; but stepping back, as long as we don't enter a rate hiking cycle, the market can digest two or three hikes. Two or three hikes are a result. Whether it's caused by inflation or something else in between is actually not that important.

4. Extreme Sentiment: Debating Whether Micron is a Cyclical Stock in June, No One Wanted Semiconductors at the End of August

Mr. Z: I saw you recently said that deciding to buy semiconductors and sell software in June seemed like a stupid move; now it's somewhat reversed, and people should gradually increase their hardware/semiconductor positions and slowly sell software. Is my interpretation correct?

Investment Talk Jun: Half right. I don't think software needs to be reduced. That tweet was just a description of sentiment. At the end of June, the market was intensely debating whether Micron was a cyclical stock, whether it deserved a 5x PE, or more than 10x or 20x PE. If you had to pinpoint a time for this round of correction, it might be after Micron's earnings report, when hardware and semiconductors started their continuous decline. But back then, no one was bullish on software; they thought software was a sector being disrupted by AI. Sentiment was extremely extreme. By the end of August, sentiment had become very pessimistic towards semiconductors: semiconductors experienced a decline in July without any rebound, a small rebound in the first half of August, and another decline in the second half. That second test is the most painful. Some stocks were approaching their late July lows, and many holding chips and semiconductors were very discouraged. But this is precisely when market sentiment is very low: SOX's forward P/E was close to late July levels. If I thought late July was cheap, then valuations at the end of August, being close to late July levels, are still cheap. In mid-to-late June, semiconductor valuations were 28x, 29x, 30x, yet everyone was discussing whether Micron was a cyclical stock and were very enthusiastic about semiconductors. Now, AI is still advancing without any signs of slowing down, valuations have dropped 40% from their highs, yet everyone has no confidence in semiconductors. You can feel it from the sentiment: you shouldn't follow the sentiment. That was the main point I wanted to make.

5. Software's Right Side Has Arrived: The First Quarter of Disconfirmation

Investment Talk Jun: As for software, I don’t think it’s time to sell. In the past couple of days, semiconductors have rebounded from the bottom, while software has declined again. The market is still playing out the narrative of “when software falls, hardware rises; when hardware falls, software rises.” I, on the other hand, believe that software will present a relatively good entry point on the right side, and the left side should already be behind us. Because judging from this quarter’s earnings reports—this is the first quarter—many software targets have shown in their reports that the integration of their business with AI is beginning to generate revenue. The implementation of AI in the software sector is just getting started. So the narrative of “AI destroying software,” which has been running for eight quarters—over two years—will not only be disproven, but the market may gradually come to realize: software valuations should not be at 10x or 11x; they should at least return to around 20x, about 70% to 80% of their previous valuations. I’m not saying it’s time to sell software now, but rather that at least now we shouldn’t sell hardware. On the contrary, hardware—and semiconductors—are still cheap at this point.

6. Just Buy SOX Without Thinking: Qualcomm Is a Good Target Left Behind

Mr. Z: In the semiconductor space, among foundry, GPU, optical communications, and memory, which sector should everyone look at first?

Investment Talk Jun: Any of them. Look at which targets have rebounded and which have corrected. Optics are still relatively strong, like Lumentum nearing new highs and surging yesterday. But not all optics are strong; there's divergence among individual stocks. The simplest thing, if you don't want to think, is to just add the index. SOX is cheap. If I have to pick, I was looking at the Qualcomm and Amazon partnership before coming in today. I think Qualcomm is a very good target: its layout in data centers and its future layout in consumer-side AI are things the market hasn't seen yet, but both are excellent positions for Qualcomm's future. Qualcomm has been left behind in this wave, a bit of a latecomer. Consumer-side phones are also affected by hardware price increases, and delivery isn't good. Looking at the market setup and future potential, the risk/reward is pretty good, and the valuation isn't expensive.

Mr. Z: But the market is quite strict on hardware earnings reports right now. Qualcomm had a very good earnings report a week or two ago, and it still dropped.

Investment Talk Jun: The strictness over the past two months is itself a matter of positioning. The so-called "sell the news" – great earnings but a falling stock price – means the market was over-positioned, or it's in a distribution process where the original positions chose to exit. But because of this situation, the valuation of the entire semiconductor sector has been brought back to a very low level. Investors should focus more on: Where is this industry heading now? Is it getting better, stagnating, or getting worse? Is the valuation at a reasonable level? If you have answers to these two questions, then one quarter's news or one earnings report's news is actually an opportunity.

7. Intel Is the Largest Position: Imagine It as AMD Plus TSMC, Don't Look at It with PE

Mr. Z: Let's talk about Intel. It seems the best case still requires the market to reprice it: the breakthrough of 18A, attracting foundry customers, Apple being very unhappy with TSMC this year. When can we see more violent growth?

Investment Talk Jun: I am very bullish on Intel. After today's rise, it has become my largest position again, although I didn't add back the portion I reduced at $120. I bought it relatively early: at $20, its market cap was under $100 billion; now it's $500-600 billion. From an investment perspective, going from $500 billion to $1 trillion is a double. Is this target worth investing in? My answer is still yes. Can it double five times to $2.5 trillion? It's not entirely impossible, because you can think of its imagination space as AMD plus TSMC: it has its own design and also does wafer foundry. The foundry business is losing money now, but as yields improve and customers come in, I believe its foundry business will turn positive by 2027.

The market, especially retail investors, has a major misunderstanding about Intel's valuation: you can't just look at PE. Its PE of 60x or 70x looks expensive, mainly because the foundry business is losing money. You absolutely cannot use PE to look at a loss-making business; its PE is infinite or zero? The foundry business should be looked at with PS, giving a multiple on revenue; only the design and product part should be looked at with PE. If you break it down, Intel is not expensive. In fact, I don't think many targets in semiconductors are expensive: Intel isn't expensive, Nvidia isn't expensive, and Broadcom isn't expensive.

I'm not worried about Intel at all because Lip-Bu Tan (Intel's CEO) has already proven his network in the entire industry very well, and that network has helped him solve internal yield issues. Looking at the earnings reports over the past two quarters, in the first quarter he said yields met the company's expectations but not his personal expectations, and in the next quarter he said yields had already exceeded his own expectations. From PR self-promotion, attracting customers, and hiring to yields, it's a complete 180-degree turn compared to other companies. Now it's also riding the CPU wave and the future AI agent wave. When will it explode? First, go from $500-600 billion to $1 trillion. After that, see what multiple the market gives and what the next step's outlook is. Fundamentals are constantly improving, and valuation will definitely be constantly updated. With this management, and the push from the entire industry behind it, this company has high certainty in my eyes, and its valuation isn't expensive. It's very reasonable to put it as the largest position, and it deserves it.

8. The Market Always Underestimates Nvidia: Depreciation Mismatch Is a False Proposition, Three-Year-Old Chips Recoup Costs in One Year

Mr. Z: Nvidia has been very cheap recently. What exactly is the market underestimating?

Investment Talk Jun: The market can find many reasons to sell off, but we never know what the market is trading on. Whether you say it's trading on dissatisfaction with circular financing or other financing. A while ago, it announced it would provide residual value guarantees for $500 billion worth of GPUs over the future with 6 institutions. Ultimately, it comes down to implementation and demand: based on current demand, A100 and H100 chips simply have no residual value risk. Chips from 3 or 5 years ago, based on current leasing prices, recoup their cost in one year. This shows that 3 or 5 years ago, the entire market was under-investing in AI infrastructure. Current demand is far higher than the investment back then, which is why demand far exceeds supply.

Now, anyone talking about whether Nvidia will have risks in 3 or 5 years – nobody knows. We can only take it step by step. Over a year ago, people were talking about a depreciation mismatch: GPUs should have a 3-year cycle, but Neocloud and the big clouds assumed 5-year depreciation, thinking the big clouds underestimated depreciation risk. Now, facts have proven the opposite: the big clouds were too conservative. Based on current on-demand leasing demand, computing power leased out by companies like SpaceX recoups costs in one year. The concern from over a year ago about a mismatch between depreciation and the payback period was a completely wrong conclusion. For the B series in hand and the Vera Rubin (Nvidia's next-generation GPU architecture) rolling out soon, no one knows what the leasing price will be in two or three years. But the leasing prices from past deployments to the present have completely disproven that view.

Since 2023, the market's expectations for Nvidia have always been wrong, consistently wrong: the market never knows what Nvidia's revenue growth rate will be next year. Whether buy-side or sell-side, they just write down whatever guidance Nvidia gives, never knowing the underlying demand. If we don't even know about next year, who knows about two or three years out? AI is the same; you can only have vague evidence and then continuously verify it. Jensen Huang giving 70% revenue growth for the next fiscal year, while the market expected over 40%, precisely proves the market is underestimating it again. Now it's at 17-18x forward. Even if the valuation doesn't recover, with 70% revenue growth, even without any operating leverage, EPS growth is also 70%. Even if the valuation contracts further, a 40% annual increase, EPS times PE gives a 40% return, which is also very good. I see memory the same way: don't expect PE to get very high. Assuming PE stays flat, relying on EPS and fundamental growth to drive the stock price up 40% annually is already a very good result.

Many people ask if the market is running out of money. I never think the market is short on money. The market is in a process of continuous verification, which also means the market is healthy. The mindless stacking of chips and hardware in June, discussing whether memory is a cyclical stock – that was an unhealthy market. The constant questioning now is actually very healthy. As long as these companies keep executing, their fundamentals keep materializing, valuations are so low, and AI is constantly being implemented and monetized, what a great market this is.

9. The Three Major Clouds' Q3 Is an Open Card: When Fundamentals Accelerate, It's Definitely Not a Selling Point

Mr. Z: On the CSP side, what are your thoughts on Meta, Microsoft, and Amazon? It feels like Q3 earnings reports should be good, and the investments made last year should be realized.

Investment Talk Jun: Google, Amazon, and Microsoft should be looked at separately from Meta. The Q3 earnings reports for the three major clouds are almost an open card: they will be very good, and they will accelerate. Microsoft's Q2 earnings report only rose 4-5% after hours. The market reaction was quite slow. I said "buy Microsoft" at the time, which I rarely do so directly on X. The market completely underestimates Microsoft. The most underestimated point is its monetization in software and AI. Copilot's implementation saw a 100% quarter-over-quarter growth rate, reaching 30 million seats. At the time, the market suspected Copilot was doing very poorly and no one was using it, but that expectation was proven wrong. Regarding the cloud, Microsoft has clearly discussed this in past earnings reports: it's not that Azure's growth rate can't be raised; it's that internal consumption consumes part of the computing power. The cloud's growth rate can be manipulated artificially. As long as you allocate some computing power to customers instead of internal use, Azure's growth rate naturally goes up. As a result, the cloud growth rate increased this quarter, and the Q3 guidance will accelerate further. Regardless of Microsoft, Google, or Amazon, cloud growth rates will continue to accelerate.

Microsoft's Q2 was 43%, and the next quarter's guidance is 45%. It's very possible to achieve 46% or 47%. When a company's fundamentals are accelerating, it is definitely not the time for us to sell, because the market will do linear extrapolation. The market starts to worry when the growth rate slows down, for example, from 43 to 40 or 38. Although it's still increasing, the magnitude is smaller. Now it's going from 43 to 45, and Q4 could be even higher. Margins are still expanding. There is absolutely no reason to be bearish on these CSPs.

10. Anthropic Saved SpaceX and Meta's Computing Power

Investment Talk: More importantly, CSPs have not yet fully benefited from the rise in short-term rental prices. The ones that can immediately benefit are Neoclouds: for example, SpaceX selling computing power to Anthropic, claiming a one-year payback period, or even nine months. The contracts signed by CSPs are relatively long; long-term contracts are stable, but the on-demand pricing for short-term rentals will only be reflected in long-term rentals when the original contracts expire and are renewed at on-demand pricing. Currently, on-demand prices are 2.5 to 3 times the price of a three-year long-term contract from Neocloud. So Neocloud also benefits greatly: CoreWeave and Nebius’s financial reports both mention that the margins on newly signed contracts are higher than before, because the CapEx is already behind them, and the pricing of new contracts is current. As long as on-demand prices continue to rise, the prices of newly signed one-year, two-year, three-year, and five-year contracts will all rise accordingly, shortening the payback period. CapEx is a rearview mirror; the money you invested is already behind you. Your initial expectation of a two-and-a-half to three-year payback period does not assume continuous price increases, but if on-demand prices keep rising, it will significantly shorten the original payback period. This is why CoreWeave says the operating margin on new contracts for the coming year has increased by 5 percentage points. For CSPs, the margin is also positive, just not as pronounced as for Neocloud. With accelerating revenue growth and accelerating margins, there is nothing to worry about with such targets.

Meta is different. It is not a cloud company yet, and it may rent out computing power in the future. However, as a company that monetizes AI, renting out computing power is not necessarily a good thing: renting out computing power itself disproves your original roadmap. If your own implementation is very successful, why would you rent out computing power? Of course, AI implementation is going well now, and there is a short-term offer for this, which they also took, selling some computing power to Anthropic. Anthropic's contribution to the entire market is really significant: it raised prices, which is beneficial for Neocloud, CSPs, Meta, and SpaceX. SpaceX originally didn't know how to handle its excess computing power, and its internal monetization ability was weak; the same for Meta. Anthropic has a very strong monetization ability, saving the excess computing power of these two companies and giving them an offer to sell at a very good price. Meta also released its personal AI agent today. The future test will still be its implementation in AI.

11. For a $100 API, Amazon Only Makes $4: Open-Source Models Return Pricing Power to CSPs

Mr. Z: CSPs are now starting to use open-source models, distilling them and then selling them. This also seems to be one of the reasons for improving margins.

Investment Talk Jun: When I first invested in Amazon, I saw it as a service provider. It never intended to build its own model from the start. The expectation was that whichever model came out, it would ultimately run on Bedrock. I mentioned earlier this year that Amazon's best monetization path is Bedrock, and now Bedrock has become its monetization machine. The future trend is that models will become commodities. Now you can say closed-source models depend on CSPs, or you can say CSPs depend on closed-source models. But from a unit economics perspective: for a final API sold for $100, the lion's share goes to the closed-source model; otherwise, its gross margin couldn't be 80% or 90%. Out of $100, $90 goes to Anthropic, and the remaining $10 is Anthropic's cost paid to the CSP. Amazon's gross margin is 40%, $6 is its original investment and cost, and Amazon only makes $4. For a $100 API implementation, it reflects only $4 in profit for Amazon, while $90 in revenue is entirely taken by the closed-source model. The pricing power lies with the closed-source model.

With open-source models, it's different. Most open-source models come from China. The monetization model for Chinese open-source models still relies on US CSPs, and the pricing power instead falls into the hands of the CSPs. Moreover, Amazon allowed Chinese open-source models to run on its platform very early on. For CSPs, a hundred flowers blooming is the best outcome. As the intelligence of open-source models improves, it is a huge positive for both CSPs and software. Earlier, I mentioned that the intelligence corresponding to a $100 API is assumed to be 130. This $100 is the cost absorbed by software companies. Software companies need to find ways to transfer this $100 cost to consumers, combine it with products, and sell at a higher price. Now with open-source models, in 3 to 6 months, for the same intelligence level of 130, the cost of running an open-source model on a CSP might only be $10 or $20. The software company's cost drops from $100 to $20 or $30. The cost reduction is very fast, and they don't necessarily have to lower prices for customers because customers don't know your cost; customers only care about how much you sell for and whether it brings better value.

Gavin Baker from Silicon Valley recently said that open-source models benefit the application layer and the infrastructure layer. I completely agree. I held the same view before he spoke: software companies' costs drop significantly, allowing them to experiment with different products, with low trial-and-error costs and short development cycles. If a product doesn't work, they can eliminate it and move to the next. Extending this logic: what used to cost $100 can now be done for $20. Software companies are still willing to spend $100, but the volume of tokens called will increase fivefold. For CSPs, this is equivalent to doing 5 businesses instead of just one before. Pricing power returns, and volume increases; this is definitely positive for CSPs.

Mr. Z: Previously, the so-called lion's share was taken by closed-source models. Now it seems CSPs can fight back. How will the competition between open-source and closed-source evolve? Currently, an efficient usage pattern is using closed-source models as the brain for command, while repetitive and mechanical tasks are handled by open-source models.

Investment Talk Jun: This is already happening, not a prediction. Two or three months ago, the CEO of Coinbase stated that their spending remained the same or grew very slowly, but the number of tokens increased exponentially because they optimized the models behind the scenes. This is bound to happen: the most difficult tasks are given to closed-source models, while repetitive and relatively simple tasks are given to open-source models. Ultimately, the demand for tokens will be higher than before, but spending won't necessarily increase. From the CSP perspective, Bedrock's position in the future will also be to help enterprises optimize model usage. All models can be called on the platform, and suggestions are given based on needs, achieving the best optimization without significantly increasing total spending. Optimization is also happening on the hardware side: HBM supply is constrained, and the HBM ratio for the new Vera Rubin is being reduced. Qualcomm's roadmap in the data center aims to address future inference and agent inference needs and the HBM constraint by separating inference decoding and prefill. Jensen Huang is also pursuing this path, having acquired Groq. Recent collaborations between Cerebras and Amazon, and between Amazon and Qualcomm, are examples. After spending too much money, everyone becomes more cost-conscious and will inevitably choose a more reasonable path: model optimization on the consumption side, and separation of usage scenarios on the hardware side, with work allocation between GPU, XPU, and Groq. This path will become increasingly clear in the future.

12. Agent Implementation: Look at Six Months to a Year: Token Price Reduction is a Prerequisite for the Agent Explosion

Mr. Z: So where is the next explosion point for AI implementation? AI agents don't seem very general-purpose yet.

Investment Talk Jun: There have been two examples in the past couple of days. Today, Meta released Muse, which does similar things to Gemini Spark, helping you order food, search, check calendars, and arrange things. The two companies have different advantages. Meta controls social media and social habits, so its implementation direction will be somewhat different. I also look forward to Apple Siri moving towards the consumer-side agent direction. Another piece is OpenAI's Astra (GPT-6), which leans more towards the enterprise-side agent. If the "Crayfish" earlier this year was the first generation of agents, I expect that in the next six months to a year, agent implementation will truly see some concrete progress, because the price of tokens has dropped. The drop in token price is the prerequisite for the potential explosion of agents. In the first half of this year, there was basically no discussion about agent-related targets and stories, mainly because coding implementation was too good, and all the computing power in the market was attracted to coding. Now that open-source models are catching up in intelligence and coding is gradually being commoditized, a lot of computing power will be released, and prices will come down. The number of agent calls increases exponentially compared to human calls. If prices don't come down, agents cannot be implemented. As intelligence increases, problems that could be solved for $1 can now solve more and harder problems. Therefore, the implementation of agents must be accompanied by a drop in API prices and an increase in the intelligence of open-source models. Today, Meta's Muse offers a limit of 1 million tokens per month, and you can subscribe to burn more. It dares to give you tokens to try, which to some extent indicates that token economics already makes sense for agent implementation. Coupled with a lot of computing power coming online in 2027, I am quite looking forward to the real implementation of agents in the next six months to a year.

13. Common Ailments of Nike and Lululemon: Too Comfortable, Not Admitting Mistakes, Turnaround Takes a Year

Mr. Z: You've also been looking at consumer goods like Nike and Lululemon recently. Any comments?

Investment Talk Jun: These two are traditional companies, not much related to AI. To get alpha in the overall market, you have to rely on AI, but I won't bet everything on AI. It is necessary to pay attention to and allocate some traditional companies to diversify risks related to AI. Lululemon, Nike, and Starbucks from a year and a half ago – the biggest common ailment of these three companies: being too comfortable, with management completely unable to see changes in consumer demand and no product innovation whatsoever. Another point is not admitting mistakes. It wasn't until prices dropped and it reflected in revenue that they realized they couldn't maintain their position without doing anything. After realizing this, they started to struggle. Product innovation and internal company culture cannot be solved in a day or two. The turnaround for this type of company takes at least a year. Starbucks is considered to have started its turnaround relatively successfully, and consumers are now accepting recent product price increases.

Lululemon has attempted to transform, but it is clearly insufficient: management also mentioned that the market acceptance of newly launched products has been poor. Its biggest problem is that it has yet to face the impact of its biggest competitor, Alo, in the Chinese market: Alo only entered China in August, and China is Lululemon's best-performing market. In its best-performing market, a competitor that has already challenged its position in other markets has just entered, which only means the challenges in the Chinese market are just beginning. The new products have not been accepted by the market, indicating that at least the first wave of innovation has failed, and a fundamental shift still needs time to be verified. Nike will fare somewhat better: it has been confirmed from other companies' financial reports that its new products have good market acceptance, but the problem lies with legacy old products, which are very problematic. The footwear market, like overall U.S. consumption, has become very picky; sales in the high-end market are not bad, with high-income groups and the top 10% of earners accounting for 50% of U.S. consumption. Good restaurants and Costco still have lines, and mid-to-high-end consumption has not been affected at all. Brands like On Running sell running shoes at higher prices than Nike, yet consumers still buy them. So Nike cannot continue selling based on the original Jordan brand; old products must be cleared at deeper discounts. New products have already been accepted by the market, and it is in a transition phase, a mix. Looking ahead, the focus is not on what discount old products sell at, but on whether new products can be accepted and their growth rate. There is still a difference in fundamentals between Nike and Lululemon.

14. Circle and Robinhood: One Step Away from Agent-to-Agent

Mr. Z: After August 19, Bitcoin rose, driving up crypto-native assets and crypto stocks. You mentioned that Circle and Robinhood have made quite obvious contributions to the portfolio. Let's talk about these two.

Investment Talk Jun: I'm still at a floating loss on Circle because I bought it relatively early and at a relatively high price. But Circle has indeed contributed quite a bit to me this year. It was already significant early in the year. Later, it dropped from 130, 140 to 60, and now it's back up. Since August, it has risen from 5% of my portfolio to a maximum of 8%, all from its own price increase. The rebound of these two companies during this period was definitely influenced by Bitcoin, no doubt about it. But another part of Circle's rise might have been overshadowed by Bitcoin's increase: earlier this year when Circle was rising, the market was hyping AI agents. It just turned out that when the "Crayfish" (Crayfish) finally came out, no one used it. People found that burning four or five hundred dollars a month on API calls, the Crayfish gave me no return. The Crayfish was the first product to ride the AI agent wave. But just like we spent so much time earlier discussing the implementation of agents, after API prices dropped and intelligence improved, part of Circle's rise should also reflect the market's speculation on whether USDC could be used for agent-to-agent payments, human-to-agent communication, and agent-to-merchant payments. This isn't really about the company itself; it's the industry moving forward, and Circle was already positioned in this industry.

But Circle is positioned for agent-to-agent. True agent-to-agent still needs some time. We've only just seen the first step make sense: humans give instructions to agents, and agents repeat what humans used to do. Today, Meta's Muse announced a partnership with Stripe. You use Muse to shop, Muse uses Stripe to pay – it's still a traditional payment method, not yet at the stage where USDC payment is needed. Right now, it's just agent commerce, maybe the second or third step. It hasn't reached agent-to-agent yet; it's one step away, but it's one step closer to agent-to-agent. The direction is correct. It's also an infrastructure company. When it was around 60-something, I shared with many people to put a portion of their portfolio in it, positioning for the future infrastructure layer. However, the management's execution capability lags behind Robinhood. Robinhood is arguably the highest execution company among all crypto-related stocks. Many very successful memes have emerged on-chain. They are really good at product development, including prediction markets. So the market gives it a premium, with a valuation of over 30 times. Whether viewed as a traditional brokerage or a crypto target, it's not cheap. But from the CEO's execution perspective, Robinhood deserves this premium. One has very strong management, no need to worry about missing the trend; the other requires waiting, hasn't reached agent-to-agent yet. It's simply about buying some when it's cheap, and reducing when market sentiment is very high but there's no clear implementation. Right now, Circle is nowhere near the stage of very high sentiment.

Mr. Z: Regarding the crypto space itself, Bitcoin's outlook for the next three to six months – how will it go?

Investment Talk Jun: My personal view is that this wave of crypto was driven by gold, and gold was driven by the real yield of US Treasuries. That's the sequence. I think gold still has room to run, so Bitcoin also has room to run, and the whole market still has room to run.

15. Tesla and SpaceX: Don't Bet Against Musk, But Don't Buy Ideals; Price Matters

Mr. Z: We forgot to talk about the most important ones, Tesla and SpaceX. What do you think of Tesla's Cybercab day?

Investment Talk Jun: I think that Robotaxi event was quite successful. The success was in the publicity. It made many people aware that the biggest difference between the new Cybercab and Waymo is that it has no steering wheel, and its design is very eye-catching, gold-colored. For a company's product to be known, it needs to create hype and do marketing. Tesla itself doesn't do much marketing. From the perspective of the event itself, it was successful – it had a gimmick and created buzz. But fundamentally, it didn't change anything: the autonomous driving that Cybercab can achieve, the Model Y can also achieve. In Austin, they have both two-seater and four-seater versions. It didn't disprove or prove autonomous driving. I hold some Tesla, not a large position. I reduced it some time ago and have been waiting for a good price opportunity to add some back. Because I still believe that the widespread implementation of autonomous driving will land on Tesla. It's just a matter of whether the price is right.

Mr. Z: What about our number one space stock, SpaceX?

Investment Talk Jun: I think SpaceX is overvalued. Looking at the results, the acquisition of Cursor was quite good. Cursor has become a very good monetization channel for xAI. Originally, xAI clearly lacked a monetization channel. But most of Musk's promises come from the business of building data centers and selling computing power. Selling computing power itself isn't worth much. The market doesn't give very high valuations to companies that just sell computing power, unless you are a hyperscaler with added services. Otherwise, you can only reference Neocloud. Its last quarter's earnings report said the return on selling computing power is 9 months, less than a year. But a model with a payback period of less than a year is unsustainable: you build a data center, computing power must be signed on long-term orders. If you place 8 GW of capacity yourself and sign short-term orders, you are essentially undermining your own business. If future AI revenue all relies on selling computing power, it's not a good business. It's also asset-heavy, meaning it will need further financing in the future. Look at how hard Oracle has had to raise funds. If Cursor takes off in the coding space, it could be a better growth point for SpaceX. It was a very good acquisition. But I don't fully agree with the promises made about the AI side. SpaceX is still expensive right now.

Mr. Z: Musk talks about a 30 trillion economy. Musk has always loved making grand promises. But there's a saying in Silicon Valley: don't bet against Musk.

Investment Talk Jun: "Don't bet against Musk" should be interpreted as don't bet against his aspirations: things you think he can't achieve, he eventually does. But when investing, you have to consider price. You can't just buy an ideal. You need to distinguish between what is possible, what is basically impossible, or what is very far off. Investing requires a timeframe and a thesis. You need to know under what circumstances and after how long it can be proven false or true. You need a yardstick. I agree with "Don't bet against Elon Musk," but that doesn't mean investing blindly. It doesn't mean you just invest over 200 when SpaceX goes public. Price still matters.

Mr. Z: Give an overview: from now until Anthropic's IPO is one phase, then the midterm elections, and after the elections. How will things evolve in these three to six months? Reuters said the roadshow is postponed to the end of September, so the IPO should fall around the end of October or beginning of November. They are expanding the Revolving Credit Facility (RCF) to $15 billion, getting more US investment banks involved.

Investment Talk Jun: For these companies going public – SpaceX, the two major model companies – you still need to be careful when investing. Because they have already priced in the next two to three years or even further. And there is a huge sell-side behind them: primary market investors need to cash out. They have returns of tens or even hundreds of times. They will definitely sell. This is a massive wealth transfer: when you buy shares of these companies, money flows from secondary market participants to primary investors, finally realizing returns on their original investments. From a liquidity perspective, these targets are too large in size and will impact the market. It's like the capital drain on Tesla when SpaceX first listed. When the market looks for reasons to decline, these things can become reasons. There will indeed be a capital absorption effect. But the main narrative still depends on the implementation of AI. Including the midterm elections, I won't mechanically apply historical trends, like the market being very good after midterms. I focus more on short-term market sentiment and participation levels. If we enter the midterm elections with very high sentiment, it's very possible that the market will decline in November. Before August, many people thought September would see a decline. But recently, software fell and hardware rebounded. Applying historical trends mechanically isn't very meaningful. I follow sentiment and valuation: valuation is a higher-level judgment logic. If valuation is high and sentiment is high, I reduce positions a bit. Now, with low valuation and low sentiment, at the moment, I don't have much to worry about.

Mr. Z: At the end of the interview, is there anything you'd like to share or advise?

Investment Talk Jun: The last question in your list is quite meaningful: if there's one sentence to share with everyone. Buffett said, "Be fearful when others are greedy, and greedy when others are fearful." This is very suitable for the market from this year until now: sentiment was very high at the beginning of the year; it dropped in March, everyone worried about war, sentiment was very low; in June, everyone was very excited about semiconductors and hardware; at the end of July and end of August, everyone was very pessimistic about semiconductors again. The market has tested us many times this year. Can you sell when others are greedy? Can you buy when others are fearful? Can you truly implement this sentence in your own investment and trading? Right now, people might think it's a cliché, but the experience left by old Buffett is actually quite useful.