US Treasuries, AI, and Crypto: Reviewing August Markets and Key Logic for Year-End Positioning

Taipei time, August 31, 2026, Monday, 7 p.m. Last Friday, Warsh made his hawkish stance on inflation explicit at Jackson Hole. The 10-year Treasury yield is at 4.7%, the 30-year at 5.2%, and oil prices are oscillating around $90. Meanwhile, Bitcoin broke through $70,000 after August 19 and briefly topped $80,000, while the semiconductor SOX valuation has fallen back to near last year's tariff war lows. Fundamentals are strong, and the macro environment is also strong—two forces pushing against each other. Where exactly is the market?

This episode of 168X War Room welcomes back qinbafrank (@qinbafrank) for the third time: on June 11, he called Bitcoin as entering an allocation zone during a livestream; on July 2, he reduced positions after seeing reports of Meta selling computing power; on July 8, he wrote about CSP value revaluation. In this episode, he lays out the map for the rest of the year: the three legs of long-term bond yields are oil prices, deficits, and AI bond issuance; before the 10-year Treasury yield falls below 4.6%, the market is stuck in a valuation friction zone; CSPs are rewriting their cost structures through open-source models and mid-tier models, and Q3 earnings will be stunning; the question "where is the next coding" is itself wrong—coding is becoming the paving technology for all non-coding workflows; Anthropic's IPO could drain $300 billion in liquidity during the roadshow phase alone; Bitcoin may pull back from mid-to-late September to October, and you should finish deploying your bullets before November. One piece of advice for investors: buy low as much as possible and don't chase highs; when the bull returns quickly, trimming a bit is the right move!

1. The July 2 Warning: Meta Selling Computing Power, US Equities Too Leveraged—Reduce Positions First
Mr. Z: It's an honor to have 168X's old friend qinbafrank, Brother Frank, back with us today. Today is Monday, August 31, at 7 p.m. Brother Frank, how have you been lately? How are your wallet and positions?

qinbafrank: I'm doing fine, pretty good. Our second chat was in mid-June, and at that time I mentioned that there could be a deleveraging in the second half of the year, one reason being that US equities are highly leveraged and positions are crowded. Then on July 2, Bloomberg reported that Meta was selling computing power, which made me a bit wary. I brought up what we had discussed again, felt it was a signal, and reduced some positions. After that, starting in early July, a large portion of what I talked about on X was CSPs: cloud providers starting to deploy open-source models, and the value revaluation of hyperscale cloud providers. Overall, in July and August, these CSPs did relatively well—they're large caps, so you can't say they rose a lot, but they barely fell during the July turmoil, and when the market rebounded in late July and early August, they were relatively stable, especially Microsoft and Amazon.

Mr. Z: The crypto market also saw an unexpected surge.

qinbafrank: I wasn't surprised by the rise itself, but I was very surprised by the magnitude. Since June, I've been saying on X that June to November is a relatively good window. If you're a long-cycle player, these months are suitable for continuous accumulation—buy at the end of each month and complete your desired allocation within five to six months. I mentioned this during the OKX Planet livestream in mid-June, so I've been gradually buying myself. In July, someone asked me how high this wave could go, and I said the rebound could reach over $70,000. I didn't expect that after the 19th, on the 21st and 22nd, it would not only break through $70,000 but also briefly top $80,000. The upward move was expected, but the strength exceeded expectations.

2. AI Demand Is Fine—It's the Macro Environment That's Pressuring: The Three Legs of Long-Term Bond Yields
Mr. Z: I've noticed you've been saying recently that for the market to improve, several conditions are needed: long-term yields need to come down, inflation needs to fall, and Warsh at least needs to be less hawkish. Could you start with his Jackson Hole speech and talk about how you see the macro side?

qinbafrank: Let's start with fundamentals. Q2 earnings proved that AI demand is okay: whether hyperscale cloud providers or Neoclouds, their earnings all said the same thing—as soon as data centers are set up, racks are installed, and deployment is done, they immediately run at full capacity. This shows that downstream demand is very strong, just that supply is insufficient. That's a very important signal. On July 31, I wrote a tweet saying August is the best summer, bullish on August's market, with leaders being optical, then CSPs and software; storage has room, but the main upward wave has passed—not that you can't expect it to rise, but it's hard to be as crazy as the past year. These all performed in August, but the strength was weaker than in April and May, and the core reason is that the overall macro environment in August wasn't that good.

The first point is that in early July, the US and Iran clashed again in the Strait of Hormuz, and oil prices rebounded from the $60s-$70s to a high of over $90 in August, recently oscillating around $90. Oil prices transmit to energy and then to inflation, pushing inflation expectations higher, which in turn pushes bond yields higher. The second is long-term bonds themselves: inflation expectations are one aspect, deficits are another, and then in Q2 and Q3 of this year, we've seen tech companies issuing more and more debt, with larger and larger scales. Hyperscale tech companies have very high ratings in the bond market—Microsoft is almost like a sovereign nation—and essentially they're competing with Treasuries for the same pool of funds. It may not be 20-30 year bonds; many are 5-year, 10-year, 20-year bonds. When you issue more debt, whether it's big tech or the government, the market demands a higher risk premium and higher interest rate compensation, pushing yields up overall. The 10-year Treasury yield suppresses the market's risk-free rate; when it rises, risk asset valuations come under pressure, especially for high-multiple, high-debt companies, as people expect their future borrowing costs to rise. That's roughly what happened in mid-to-late August.

3. Warsh's Jackson Hole: Hawkish on Inflation, Market Enters Stock-Bond Friction Zone
qinbafrank: Regarding Warsh's speech, I wrote a preview before the 28th, and it turned out to be half right. Half of my prediction was that he would definitely express a hawkish, very tough stance on inflation. Because in June he was dovish, and when the June nonfarm payrolls released in early-to-mid July came in below expectations and June CPI also came in significantly below expectations, long-term yields were about to rise, and his attitude softened—the market thought he caved. But by the time of his speech last Friday, with the 10-year Treasury at 4.7%, the 30-year at 5.2%, and US oil still at $88-$89, he had to be tough on inflation at that point. At the time, I expected that the previous week, Bessent had already stepped in—first helping Japan stabilize its currency to avoid the Bank of Japan selling US Treasuries, then announcing an expansion of long-end buyback scale—but the effect lasted only one day, and the next day long-end yields rebounded. So I expected that while he would express toughness, his overall attitude would be a bit more moderate. In fact, he was not only very tough on inflation but also proposed a strong reaction function: he wants to see inflation move down, and quickly approach the target.

How does the market see it? In the long term, this is correct: if you're tough enough and actually bring inflation down, the market will have confidence, and at least one of the three drivers of high yields will be eliminated—that's a good thing. But in the short term, the market will still be under pressure. I call the current state a stock-bond friction zone: with the 10-year Treasury at 4.7% and the 30-year at 5.2%, this level won't cause the market to crash immediately, but it's also hard to see a large-scale, trend-upward move. Everyone remains vigilant, funds don't dare to charge in boldly, and they're afraid of surprises that could push yields higher—it's very uncomfortable. To be honest, if the 10-year Treasury goes from 4.7% to 4.9% and the 30-year above 5.3%, the market might not hold; at this level it can still hold, but pushing it further up is very difficult. So it needs time—watch this week's August nonfarm payrolls and next week's August CPI, and ultimately it depends on whether the US and Iran can reconcile.

4. September: Standing Pat—Consumption Mediocre, Nonfarm Payrolls Weak, Five Working Groups' Results Not Out
Mr. Z: So what's your judgment on the September FOMC meeting?

qinbafrank: After seeing these two August data points, the market will price in September, but from my personal reasoning—which cannot be the basis for everyone—I lean toward him standing pat. Three reasons. First, consumption isn't that strong; August consumption data was mediocre, Walmart's Q2 earnings weren't great, and same-store sales growth is slowing sharply. Second, the labor market is also just average; this week's nonfarm payrolls market expectation is around 50,000, and I think it could be below 50,000, in the low single digits, and since July nonfarm payrolls were negative, August could be too. These two factors make a rate hike unlikely. Third, Warsh set up five working groups in June to study monetary policy, interest rates, inflation frameworks, and productivity, with results due by year-end—not necessarily December. I lean toward him standing pat until the results of these five working groups are out.

Where's the surprise? The surprise is that inflation isn't not coming down—it's just coming down slowly. July inflation was only slightly lower than June, and August oil and gasoline prices have risen a bit. I look at a real-time inflation data called Truflation; the absolute value isn't referential, but I've tracked it for about a year, and the trend is very referential: its latest real-time trend has returned to near the June high, just slightly lower, so August inflation could be choppy, and the market might slip a bit more.

5. US-Iran Game Becoming Protracted: Inflection Point Is Oil Prices Falling, 10-Year Treasury Below 4.6%
qinbafrank: The US-Iran issue now has a tendency to become protracted. By protracted, I mean the intensity isn't very high—it won't be a major conflict like in March—but both sides have weaknesses. Iran knows Trump cares about elections and public opinion, so it uses control of the strait to keep oil prices elevated and force concessions. The US has found a path: blockading Iran's coastline and increasing sanctions to cause internal economic problems in Iran, forcing Iran to back down. Trump now believes this is effective, so he won't commit to lifting sanctions, unfreezing assets, or recognizing Iran's position in the Strait of Hormuz. The two sides are in a standoff, and the sticking point is oil prices at that level.

The inflection point depends on two things. First, when both sides can't hold on and one side concedes, oil prices really come down—that's a huge positive. Second, when oil prices fall, one leg of long-term yield support is weakened, and at the same time, Bessent is doing buybacks; if they also adjust the issuance ratio of long-term and short-term bonds, that will also push yields down. If the 10-year Treasury falls below 4.6% to 4.5%, and the 30-year falls to around 5% or below, then AI and semiconductor upstream and downstream assets can truly break out. Fundamentals are okay and demand is okay—that's not a problem; it's just that the macro environment is still in a state of relatively high uncertainty.

Mr. Z: It sounds like the macro environment is a bit like noise right now, but actually demand never bottoms out. Jensen Huang's earnings call two weeks ago also said the problem is supply, not demand.

qinbafrank: Right, this is what we've been saying across our three chats. Penetration is rising rapidly, more and more users are coming into contact with AI, and whether C-end or B-end, each user's token consumption is also growing. In the short term, demand is surging. It's just two things: one is short-term macro suppression, and the other is short-term market FOMO to the extreme and divergence, like the state in June—demand was good, but many prices had already risen parabolically.

6. CSP Value Revaluation: From Middlemen to the Most Comprehensive Indicator of AI Implementation
Mr. Z: Speaking of token demand, we have to talk about CSPs. You previously wrote that article on CSP value revaluation, and I posted it in the Space. The market has always treated AWS and Google Cloud as middlemen selling GPUs and reselling tokens. Why do you think this perception must be reversed?

qinbafrank: It started in early July. At the end of June, everyone saw many reports that many US companies' token consumption was very large and costs remained high, to the point where they couldn't bear it and had to limit each employee's quota or even cut spending. The term "token economics" was being discussed at that time. Then I saw Coinbase CEO Brian Armstrong publish a long article about what Coinbase was doing throughout June: they differentiated their business scenarios—some ordinary scenarios that aren't high-value but are high-frequency don't necessarily need frontier large models, don't necessarily need GPT or Claude; using open-source models can still achieve compliant results. The result was that throughout June, Coinbase's token consumption grew rapidly, but overall cost expenditure declined. Many people at first glance thought this case was alarming: tokens rose but spending fell. But I saw a different signal: for an individual enterprise, building a model router, differentiating scenarios, and combining flagship large models with very cost-effective open-source models and mid-tier models makes costs controllable. This is Jevons paradox: the lower the cost, the more people use it. I worked in the internet industry for many years; the cost of going online and traffic kept falling, but more and more people used the internet, each person's consumption grew, and total spending actually increased.

So in July, I did a lot of research and asked friends domestically and abroad, and I saw two points. First, enterprise AI adoption has entered an era of scenario-based, multi-model engineering: high-value scenarios continue to use frontier flagship models; high-frequency, low-complexity, low-value scenarios use open-source or mid-tier models. Second, Coinbase is special—it's a tech company itself and has the ability to build its own model router, do context compression, set up caching, and build an entire AI operating system. But most enterprises don't have this capability; they already use cloud services, so cloud providers have the incentive to do this. In mid-to-late June, we also saw Microsoft start deploying DeepSeek and Qwen, so I wrote several articles in early July and summarized them into one on July 8.

The conclusion is: frontier large models are only one important component of an enterprise's overall AI adoption process. Most consumption may still be on frontier models, but enterprises also have many scenarios starting to use mid-tier and open-source models. Before June, the market's most important indicator for AI commercialization was the ARR of Anthropic and OpenAI, because at that time you could just calculate by connecting to their APIs. But as scenarios become more complex, the value of cloud providers grows. The performance of hyperscale CSPs is a better and more comprehensive indicator of AI commercialization than model ARR: it's essentially the entire operating cost of an enterprise putting AI into production environments, including closed-source model APIs, open-source models, inference, databases, object storage, networking, security monitoring, and even agent operations—all reflected on the cloud platform. Large model ARR remains important; it's the purest indicator of AI demand. CSP performance is the most comprehensive indicator of AI commercialization implementation.

7. Open-Source Models Rewriting CSP Cost Structures: Microsoft's Qwen, DeepSeek, and Its Own Mid-Tier Models
Mr. Z: Last Thursday, I invited Herman to chat, and the biggest takeaway was also this point: CSPs are at the forefront seeing demand before deciding on CapEx. This year they're spending a lot, so free cash flow is low, but when they use open-source models to refine and actually earn money back, and free cash flow turns significantly positive, the stock price will surge faster than we imagine.

qinbafrank: What it changes is the CSP's cost structure. Why were CSPs previously called middlemen? Capital expenditure is your spending—buying machines, building data centers—and then you resell tokens and APIs of frontier models, but the bulk goes to the model companies. OpenAI and Anthropic leave you a portion; you bear the expenditure but don't enjoy the largest gross margin. The benefit of open-source models is that they have no cost, and most of that revenue falls to the CSP itself. And it's not just open-source models—it should be called open-source models and mid-tier models: in June to July, Microsoft deployed open-source models like Qwen and DeepSeek, and Nvidia also pushed its own open-source models. On the other hand, Microsoft also launched its own mid-tier model, MAI Model, which is closed-source and focused on cost-effectiveness. Its performance may not be as good as GPT 5.6 or Fable 5, but it might be comparable to GPT 5.5 or Opus 4.7, 4.8, and in certain scenarios it's already more than sufficient. The benefit is that it's more controllable, and the cost improvement is very significant.

Mr. Z: I also noticed something. OpenAI's latest generation model, Astra, is no longer open to the general public, only to researchers. Last Friday, Anthropic tweeted asking for 10,000 scientists to use its latest model. Intelligence has started to become stratified: ordinary people don't use, and don't need to use, such frontier models. And open-source models, like GLM 5.3, have input and output costs around $0.15-$0.5 per million tokens—that's terrifying, and US large model companies simply can't compete.

qinbafrank: Right, that's the forcing mechanism. Look, OpenAI has also been cutting prices recently, pushing prices down. If you don't, Chinese models will force you down.

Mr. Z: That's what introductory economics calls price discrimination—tiered pricing for different consumer groups.

8. Q3 Earnings Will Be Stunning: Revenue, Gross Margin, and the Second Scenario
Mr. Z: So if CSPs are playing this way, by year-end or early next year, what else can we expect from CSPs? What's the second growth engine?

qinbafrank: I'm bullish on Q3 earnings. When they're disclosed from mid-October to November, the earnings of hyperscale cloud providers like Microsoft, Amazon, and Google will be very, very stunning. First, their Q2 guidance was already strong: Microsoft said Q2 growth was 43% and Q3 could be 45%; AWS Q2 was 37% and Q3 could be 40%; Google Q2 was over 80% year-over-year and Q3 could accelerate. As long as demand doesn't change, that's the first layer. Second, Q2 earnings only went through the end of June, so we couldn't yet see the impact of cloud providers' large-scale deployment of open-source models and their own mid-tier models on cost structures. But July, August, and September are exactly the months when they're actively deploying, so in Q3 we might truly see not only revenue growth but also a significant change in gross margin—I think it could be quite significant. Third is the second scenario.

Previously, the market wrestled with whether such large CapEx could earn back money and how to account for it. Q2 earnings gave everyone a preliminary ROI framework that could be initially validated, not final validation. Recently, people have started discussing again: your commercialization is so strong because coding has run particularly fast this year, but coding is a special scenario. Why? Code, logs, APIs, and documentation are all digitized, and there's a global GitHub code repository. Coding output can be quickly verified, and the process can be broken down into clear steps. Programmers are expensive—in the US and China—so if you save 20%, 30%, 40% of development time, or cut some people, ROI is easy to calculate. And it doesn't require company-wide adoption from top to bottom; programmers can just use a coding agent or API themselves, so penetration is very fast and commercialization growth is very fast. Then people ask, what's the next coding?

9. Where Is the Next Coding? The Question Itself Is Wrong
qinbafrank: On the 23rd, I wrote a long article, and the conclusion is that this question may be wrong from the start. The coding scenario is very unique, and it's hard to see another product or scenario with such extremely fast marginal growth, huge revenue scale, and quick recognition by capital markets. The non-coding world is completely different from coding, and it's very fragmented: banks have credit, KYC, anti-money laundering; insurance companies have underwriting and claims; healthcare has diagnosis and long-term post-diagnosis services; manufacturing is a big mixed bag—automobiles, steel—with different processes, different equipment, different data. Each scenario requires many processes and data that are enterprise-specific, not as concentrated as coding.

But if you look closely, there are still distinctions. Within enterprises, there's a horizontal knowledge work system—finance, recruiting, HR, documents, contracts—these functional scenarios can be generalized. Then there are many service and back-office operations management systems, which are also fast. Further out, supply chain and manufacturing production will be relatively slower. So the future isn't a coding-like scenario, but many different knowledge workflows, plus business processes in many industries, starting to be agentified, ultimately exceeding coding in total volume. IT operations, finance and accounting, administrative legal, compliance are fast; supply chain, manufacturing, healthcare are slower—different time lags. Microsoft's Q2 earnings show this: Foundry has over 100,000 customers—that's a platform for enterprises to build and deploy agents, meaning 100,000 enterprises are using it to build their own agents. Microsoft 365 Copilot has 30 million paid seats, meaning many non-coding employees are using it.

Another point everyone should understand: coding is becoming the paving technology for non-coding workflows. In the future, many non-technical employees won't think they're writing code, but by expressing needs in natural language, they'll generate financial reconciliation tools, sales dashboards, internal approval pages, contract review processes—these small applications and small agents serve themselves or their departments. Traditional enterprises have a huge number of such processes—small scale, special needs—and enterprises won't specifically purchase a SaaS for them; they're all operated with Excel, email, and manual work. Coding agents lower the barrier to custom software; processes that previously weren't worth software-izing can now be software-ized, accumulating small amounts into a strong long tail. Coding will become the bottom layer of all non-coding business processes, and together that's the second scenario. Last week, a16z's blog also shared data: since February, the fastest-growing areas for Codex adoption are legal with 100x growth, and sales, recruiting, marketing, healthcare all at 30-40x—the sprouts are already showing. Of course, we must admit that these business processes aren't that standardized, and the deeper you go, the slower the progress: enterprise-specific scenarios require permission management, internal data, retraining, redeployment—that's Palantir's FDE model. So in June and July, Anthropic and OpenAI were both setting up dedicated FDE teams with many companies to deploy and implement.

10. What If the Boss Can't See the Return? 15% of Non-Tech Listed Companies Already Mention AI in Earnings
Mr. Z: I have a few takeaways here. First, AI will help us do many things we couldn't do before, and it will also replace many mechanical tasks. Second, at the current frontier level of models, if you have a $100 Claude subscription or Fable 5 and can't max out your usage in a week, it likely means your job is easily replaceable. That sounds harsh, but it seems quite true. But I'm more curious: enterprises introduce AI to their employees to compress costs, but many business owners see the investment but not the return—they don't know where the return is and can't calculate the numbers. How do you solve that?

qinbafrank: This also takes time. First judgment: Codex has significantly improved in legal, sales, recruiting, marketing, healthcare since February. If you use it for one or two months or a quarter without effect, you won't pay next quarter. Its continued growth shows it's effective—the effect is either cost reduction and efficiency improvement or revenue generation. Last weekend, I met with a partner at a domestic USD VC who has invested in many internet and gaming companies. One of them is very well-known in iOS mobile gaming—I won't name which one. He talked with the founder, and they've already applied AI to the entire game production process, from creativity to art to planning to data systems. Previously, a major game took two to three years; now one game might be out in six months. That's worth noting.

Second, something everyone can do: have GPT or Claude go through all US non-tech traditional industry listed companies' earnings reports and conference calls from July 15, when the US Q2 earnings season started, to now—about a month and a half—and see which companies mentioned using AI and what the effect is, whether cost reduction and efficiency improvement or real business performance. According to my tally, about 15% of US non-tech listed companies have started showing in their earnings that they feel the effects of AI. That's a very clear signal. As long as the boss feels costs have been cut, or costs are the same but efficiency is higher, that's an effect, and then it gradually moves to generating real revenue.

11. When Will CapEx Break Even? Next Year—Watch Two Crossovers First
Mr. Z: So when do you think we can say with certainty that AI CapEx—this real money—can be justified? When can this account be balanced? We also saw two weeks ago that Anthropic's ARR has started to decelerate, reaching $65 billion.

qinbafrank: Possibly next year or the year after. In Q2 this year, we saw a preliminary ROI framework, but truly breaking even—we're still in the spending phase now, which is the time lag I mentioned earlier. US hyperscale CSPs' large-scale capital expenditure started in the second half of last year, from Q3 to Q4. Money spent last year only starts generating cash flow in Q2, Q3, and year-end of this year, and from Q3-Q4 last year to Q1-Q2-Q3 this year, spending has been increasing. Amazon's CEO mentioned another time lag in Q2 earnings: capital expenditure has a rhythm—first invest a sum to buy land, build data center buildings, roads, and power facilities, which may take 8 to 12 months; then bring in equipment, install racks, buy optical modules and fiber to connect racks, which takes another one or two quarters. From investing a sum to the data center being built, powered on, and generating cash flow, it takes at least a year to a year and a half. Money invested in Q3 last year only starts showing effect in Q3 this year, or even year-end.

Looking forward, this year is the peak: last year, big tech capital expenditure was about $400 billion, and this year it should reach $800-900 billion—100% growth. Next year, the market expects $1.1-1.2 trillion, which is 30-40% growth. Possibly by mid-to-late next year or the year after, we can see how much each batch of capital expenditure invested and how much output it's generating, and whether it can break even. Full accounting break-even takes longer, but you don't have to wait until it actually makes money for it to be a big deal. Watch two crossovers: first, cloud business revenue growth exceeding capital expenditure growth—this hasn't happened yet; even Google, the fastest-growing, hasn't reached it: GCP Q2 growth was 82%, but capital expenditure went from under $90 billion in 2025 to a guidance of over $200 billion this year—over 100% growth. Second, after coding's ultra-high growth comes down, non-coding business scenarios start growing rapidly—I call it the baton handoff.

12. Three Phases Before Year-End: Autumn Chill, and the Real Show After Midterm Elections
Mr. Z: It sounds like CapEx will slow down, and downstream beneficiaries will also be fewer. Looking back, how will the market from now to year-end differ from March, April, and May this year?

qinbafrank: March, April, and May were undoubtedly very strong. In March, the Iranian market was still falling; in April and May, there was a computing power shortage and Anthropic's ARR was growing rapidly—reaching $30 billion shocked everyone. Q1 cloud provider earnings were very good, commercialization had initial sprouts, and capital expenditure hadn't peaked. More critically, the macro environment was very good then—in April there was still fighting, but everyone knew it would ease over time, oil prices were falling, long-term yields weren't that high, and the market wasn't worried about a bond issuance wave. Once easing was confirmed, long-term yields fell rapidly after May, with the 10-year Treasury dropping to 3.9%, so that was a very explosive growth.

Going forward to year-end, we need to look at three phases. Early to mid-September is one phase, mid-September to mid-October is another, and after the midterm elections is another. In the last week or two, the market needs to digest Warsh's speech—yields are high and risk assets are under pressure. If this week's August nonfarm payrolls continue to be weak, the market will breathe a sigh of relief, but we need to watch next week's August CPI. In September and October, I think there will still be turbulence, and the sources are several. First is the midterm elections. When the race is tight, whether Republicans keep both chambers, lose both, or split—this political uncertainty brings uncertainty in industrial policy. Historically, before midterm elections, the market is either flat or adjusts for risk aversion. This year's special points: one, Democrats will attack Republicans on the Iran issue; two, AI data center construction has greatly affected people's livelihoods: data centers push up electricity, water, and land prices. In places like Florida, Virginia, and Texas where many are built, after the rush to build, a data center only needs a dozen or twenty people once completed. Public sentiment in many places is boiling—New York has suspended data center approvals; the Texas governor is Republican and originally supported data centers, but when the race is tight and Democrats attack this point, he becomes cautious. Data centers have become a very controversial social focal point during the midterm elections, at least affecting market sentiment.

So my July 31 tweet said August is the best summer, and on August 23, I wrote another: autumn chill is setting in. Not a deep winter, but macro suppression pushing the market into a valuation friction zone—an uncomfortable phase. After the midterm elections, there are two points: regardless of whether Republicans win big, lose big, or split, political uncertainty decreases, and the market knows how to extrapolate policy. Mid-to-late October to November is also Q3 earnings season, which may be better. Add one more point: the market oscillating for a while is to build strength—March, April, May, June rose sharply, and July's damage was severe. Momentum trading needs time to slowly recover after such a big adjustment; it can't immediately charge ahead. Historical patterns also show this—I've seen Citi's statistics and pulled my own: in 2010, 2014, 2018, 2022, before midterm elections, markets were relatively choppy with low returns, and after elections, 30, 50, 100 trading days were all good, with tech growth assets benefiting more.

Mr. Z: We've been chatting for an hour, and it sounds like in September and October, it's better to be in cash or hold more cash?

qinbafrank: Being cautious is good, but I personally lean toward not being that pessimistic, and I don't think the adjustment will necessarily be that large. Watch a few points: when the US and Iran reconcile, oil prices, deficits, Treasury oversupply plus AI bond issuance—these are the three legs of long-term bond yields. If US-Iran tensions ease and oil prices fall, one leg is knocked out, and then Bessent's buybacks and adjustment of issuance ratios will have a compounding effect. When the 10-year Treasury yield is pushed below 4.6% and the 30-year below 5.1% or even below 5%, and it really can't go back up and stays down there, only then will the suppression on risk assets ease. From mid-to-late September to October, I personally think being cautious doesn't hurt—keep some bullets, no need to overcomplicate.

13. Anthropic IPO: Raising $100 Billion, Draining $300 Billion During Roadshow
qinbafrank: The second source of turbulence—today I saw a report from The Information that Anthropic will announce IPO documents on the 7th or 8th. If it lists in October, it will essentially affect the market, similar to how SpaceX's IPO felt in June this year. Previously, people looking at large model companies didn't really know their true operations; they could only see occasional ARR numbers from official announcements or third-party monitoring. After the IPO documents come out, we'll see Anthropic's core operating situation, which might be better than we think, boosting market sentiment. But then it will do a roadshow—it wants to raise $80-100 billion, higher than SpaceX. SpaceX was $75 billion at the time, with 3-4x oversubscription. Institutions need to prepare money to subscribe, which means at least $200-300 billion in funds will be drained—that's the drain during the pre-IPO roadshow phase. Anthropic might be even hotter; if operating data is better than expected, raising $80-100 billion could end up with $400-500 billion in subscriptions, meaning the market needs to prepare at least $300 billion in funds. After listing, with a 5-6% float, it might spike up a bit, which would be a suppression on other AI assets, and then recover. That's the source of turbulence in mid-to-late September and October. By after the midterm elections, it may have been listed for over a month, and the shock will have passed.

14. The Trap of Korean Leveraged ETFs: Money Flowing In, Underlying Stocks Not Rising
Mr. Z: Some people rushed into the market in late July to pick up bargains. Taiwan stocks and US stocks have different logics—Taiwan stocks don't care as much about free cash flow because US CapEx is nourishment for Taiwan stocks; what matters is who's in the supply chain and who's placing orders. Taiwan stocks should be the best-performing market in August, with foreign capital returning to buy more strongly than in Korea.

qinbafrank: Korean stocks mainly went into 2x and 3x leverage. In July, the whole market adjusted, and Korea was the main driver. Their own regulation also had problems: they massively liberalized in May, then massively restricted starting late June, and even raised rates once in July—essentially shooting themselves down.

Mr. Z: These leveraged ETFs are deadly because they're essentially shorting volatility. You can't buy these products on a rebound—during rebounds, there's often friction. It only makes sense to buy when the market gives it a valuation and it's being pulled up in a straight line, but ordinary investors don't have that awareness. Samsung and SK Hynix leveraged ETFs have seen massive inflows, but it hasn't been reflected in the underlying stocks themselves. This shows that institutions must be distributing—there's selling pressure above, otherwise the underlying stocks should have risen.

qinbafrank: So in late July, I thought those positions had value, but as for a main upward wave soaring into the sky, that's hard—grinding upward is still effective. Hynix and Micron have still risen a bit from late July and early August prices, just not that strongly.

15. SOX Valuation at 19.2: Near Last Year's Tariff War Lows, Q4 and Next Year Look Good
Mr. Z: If we look further ahead with more imagination, from after the November midterm elections to next January, what's the picture?

qinbafrank: I'm bullish on the show after mid-to-late November. First, Q3 earnings should be quite good, especially for cloud providers, continuing to validate rapid cloud business growth, and some may even exceed CapEx growth. We'll also see non-coding scenarios starting to commercialize at scale, with more and more evidence. Second, the Anthropic IPO shock has passed, and political uncertainty has decreased. Third, the market is slowly adjusting and getting cheaper. In the past couple of days, I saw a data point: the average valuation of semiconductor SOX is now about 19.2, last year's tariff war low was about 18, and August 2024 was 16-17. This level already has value. As it oscillates further, with earnings moving up, valuations will move down further, and value will emerge. So Q4 this year or next year could still be good.

16. Bitcoin: Start Accumulating at $63,000, Finish Deploying Bullets Before November, Possible Pullback in Mid-to-Late September
Mr. Z: Let's talk about the crypto market outlook. What's the new normal for crypto? You previously mentioned compliant ICO 2.0.

qinbafrank: Split into two parts: Bitcoin and the crypto market. For Bitcoin, I've been talking about it since June. On June 11, during the OKX Planet livestream, Bitcoin was around $63,000, and I said Bitcoin had entered a high-value zone—you could start accumulating and it was worth buying. I'm a trend and cycle investor, and my suggestion is that if you look at it on an annual basis, long-cycle, Bitcoin from June to November is in an allocation zone. Complete your desired allocation over several months, then hold for one to two years, two to three years. In July, I said there would be a rebound to over $70,000, but I didn't expect it to go directly above $80,000 recently. But from a trend perspective, it won't go up in a straight line. On Saturday, August 22, I summarized in a series of tweets that there might be a pullback in mid-to-late September or October: looking at the Treasury's refinancing plan and TGA account balance targets, liquidity will shrink slightly from mid-to-late September to October. My framework for Bitcoin is that it's very closely related to dollar liquidity centered on the dollar and bank reserves, so I expect a pullback in mid-to-late September to October. Whether it can reach a new low or a secondary low is hard to say—I think a secondary low is more likely, but this is easy to be wrong about. For me, if there's a pullback, I'll keep buying, and finish deploying bullets before November.

For Bitcoin, cycles are not a flaw—they're a feature. The four-year cycle will weaken as output decreases, and in fact, we can see that from the cycles of 2010-2011, 2014, 2018, 2022, 2026, each wave's decline has been getting smaller.

17. The New Normal: Extreme Divergence in Crypto, Quality Asset Supply Is the Core
qinbafrank: For non-Bitcoin crypto assets, the new normal is something I proposed in mid-2024. For many years before, from ICO to DeFi, the problem was that your innovation was at the legal level, and the business had no value creation, leading to massive asset issuance but all junk assets. According to the laws of financial markets, the larger the crypto market, the more mature it becomes, and all financial markets differentiate as they go from barbaric to mature. In 2024, I sorted through US stocks: 5,000 companies, the bottom 3,000 by market cap account for only 5% of total US market cap, with average market caps of $800-900 million and daily trading volumes of maybe $1-2 million, while the US stock market already had a market cap of $50-60 trillion. The new normal I proposed is that the crypto market will be extremely divergent in the coming years—only a very few assets will still be okay, most assets will get worse and worse, and their ultimate value is just being accumulated and pumped by market makers and main players on exchanges. Looking back now, it's indeed validated: altcoins were obscure before July and August, and recently when Bitcoin rose a lot, they followed, but only a small handful.

The problem is that for any capital market to sustainably strengthen, the core is continuous supply of quality assets. US stocks constantly eliminate and iterate the index while absorbing quality listed companies globally, giving confidence to long-term investors, trend investors, and short-term traders alike. In the crypto space over the years, there have been too few quality assets and too many junk assets. The biggest change from last year to this year is tokenization: real business assets and US stock assets on-chain, bringing many quality assets to the chain. Last July, I wrote about what new asset types would emerge after US stock tokenization: US stock token contracts will develop quickly; equity tokenization of unlisted companies, i.e., Pre-IPO; companies on US crowdfunding platforms can directly issue tokens; US small and mid-cap stocks directly issuing tokens with stock-token linkage. Looking back now, the first and second points are already mainstream, and exchanges are still exploring making US stock tokens into unified margin—it's starting to become a Lego block. In July, Robinhood launched an L2 chain based on Ethereum, and what's interesting is that previously meme token pools were paired with Ethereum, USDT, USDC; now they're trying to pair meme coins with SPY tokens and Nvidia tokens. On-chain asset liquidity doesn't have to be bound to Ethereum, Solana, or stablecoins—it can be bound to quality assets with real value in traditional finance: gold, Nvidia, SPY, QQQ tokens all work. DeFi is a big financial Lego, and the core of lending is how many quality assets can serve as underlying assets. Since DeFi emerged in 2020, the only truly quality on-chain assets were Bitcoin and Ethereum, with Solana counting as half. Now that quality US stock tokens are coming on-chain, there are more on-chain assets, and the Lego will get bigger and bigger.

18. Crypto Is Not Productivity—It's a Capital Paradigm: The Application Is Perp DEXs
qinbafrank: Last July and August, I wrote a provocative piece, and I mentioned it in my first visit: everyone thinks there must be Web3 native applications, but I found this might be wrong. Crypto is a capital paradigm, and a capital paradigm doesn't solve productivity problems—it solves production relation problems: the efficiency of asset issuance, circulation, and trading. Asset issuance, circulation, and trading in crypto are naturally 10x more efficient. The SEC pushing the entire US traditional financial system on-chain is precisely because it sees the characteristics of this new capital paradigm. Productivity technologies can be productized and application-ized—software, internet, mobile internet, AI are all like that. But capital paradigm things need to attach to business scenarios; standalone is hard. If you really ask what crypto's application is, it's Perp DEXs and DeFi—that's the real crypto application, its most powerful part, not decentralized social or those things. So first, the capital paradigm must combine with traditional things that have real business scenarios; second is the combination of AI and crypto, including agent payments; third is compliant ICO 2.0.

19. Compliant ICO 2.0: SEC Regulation Crypto Assets (Reg CA), $75 Million Direct Public Offering
qinbafrank: Why was the rebound after August 19 so strong, reaching over $80,000 in a week? People talk about buybacks, the White House tech leaders meeting, and short squeezes—there were a lot of accumulated shorts. But I saw a very important factor: on August 19, the SEC proposed something called Regulation CA—Regulation Crypto Assets. This is a very big change in US regulation. Previously, you issued tokens first, and then I reviewed whether they were securities; now they've given you a framework. Originally, US companies had two paths to raise funds from the public: IPO, or various exemption clauses. Regulation A is a mini IPO, but after listing you can only go to OTC pink sheets; Regulation CF is crowdfunding; Regulation D is private equity financing. All three are aimed at accredited investors, with either poor liquidity or only pink sheet listing. What Regulation CA means is that startups have a very low threshold—they can raise $5 million a year; after meeting transparency and filing standards, they can raise $75 million a year, and they can directly publicly offer to the public without intermediaries. Once SEC review passes, they can raise on their own website, and after raising, they can directly list on Coinbase, Kraken, Binance—liquidity much better than pink sheets. In the past few years, US projects were all offshore structures with non-US public offerings; this is a very important change. More importantly, it separates investment contracts from asset attributes: when you raise funds, you promise to build the network well and decentralize; if you fulfill the promise within a few years, the investment contract terminates, and the token is no longer a security.

This is hugely significant. First, you can directly do compliant ICOs and public offerings in the US. Second, the SEC's scope isn't limited to native crypto or blockchain companies—as long as it's an internet-type business, computing power networks, WiFi networks, gaming, as long as it's a multi-sided platform, it's possible. This means many traditional companies with real business support but not yet meeting listing standards or not large enough can issue tokens through this framework. Such tokens have real business support and are better than traditional air ecosystem coins. It's not that compliance and transparency guarantee value—there are plenty of junk companies in US stocks too—but it increases the probability of good projects, and with SEC filing, the cost of wrongdoing is higher; cutting leeks and fraud can be subject to US long-arm jurisdiction. So on the 20th and 21st, I wrote quite a bit—this is compliant ICO 2.0. The three major trends for the crypto market in the next cycle: tokenization plus RWA, the combination of AI and crypto, and compliant ICO 2.0. It's still in the comment period now, likely to officially land in October, and we'll see the specific terms then—looser regulation also has two sides. Robinhood's move is also interesting: a compliant traditional financial platform building a public chain, using US stock tokens and stock-token pairs as underlying asset liquidity anchors—to some extent, an innovation.

20. Three Chains and the Coexistence of CEX and DEX: US Stocks Becoming Crypto-like, Crypto Becoming US Stock-like
Mr. Z: The main theme sounds like the bar for crypto projects is higher—everyone looks at whether you have real applications, revenue, and profit. The best at this are a few: first, Tether; second, Hyperliquid—a16z, Paradigm, Multicoin, these still-active big VCs all hold HYPE positions; third, Prediction Market—Kalshi and Polymarket are fighting for market share, fighting for sports, and raising more funds. Looking ahead, it's actually a showdown between Peter Thiel's Founders Fund and Sequoia. At this point in time, I'm only looking at three chains. First is Solana—it has meme and retail-driven stuff, but it also worries me: since last summer, it's been shouting about building an Internet Capital Market, an internet-native capital market, but I haven't seen a real use case land yet. Second is BSC—it serves the Binance empire, Binance points where to fight, fostering Perp DEX like Aster and Prediction Market, but whether it can truly complement Binance as a centralized exchange on-chain and non-KYC, stablecoins are still mainly on Tron and Ethereum, Perp DEX and Prediction Market haven't broken out yet—what exactly is on BSC needs more observation. Third is Robinhood—a traditional broker holding so many retail users wanting to do RWA. The founder goes on CNBC in a suit during the day to talk serious business, while on-chain there's a team issuing meme coins to attract attention, but what they're actually doing is RWA. This echoes what you said: what's happened in the past year or two can be summed up in one sentence—US stocks are becoming crypto-like, with increased volatility and fierce valuation growth; crypto is becoming US stock-like, with everyone increasingly looking at cash flow and fundamentals.

qinbafrank: This will definitely be integrated in the future—no distinction between on-chain and off-chain, no distinction between US stocks and crypto. Many things played on-chain, except meme which is native, are related to stocks or things with real business. Beneficiaries, I'd roughly divide as follows. US stock-related crypto: HOOD, Circle, Coinbase all benefit—Robinhood has its own chain, Coinbase has Base. Over the past year or two, I've been comparing Coinbase and Robinhood: Robinhood's team I call bad kids—act first, talk later, very good at grabbing topics; Coinbase's team is compliance, steady and stable. In the future, Base may also benefit from tokenization and compliant ICO 2.0 dividends. US-concept public chains are just a few: Ethereum, Solana, Base, Robinhood's chain. Token assets may be on a few public chains: Ethereum, Solana, BNB. BSC is backed by Binance—on one hand doing meme, on the other hand Binance has also done US stock tokens, and BSC has them too. The second area I'm bullish on is institutional brand assets related to RWA and DeFi: lending, oracles—those that have gone through multiple cycles and proven security, application value, and business volume should still perform well in the new cycle. Overall, the biggest significance of this late August rally isn't that we're completely optimistic and about to charge, but that after extreme pessimism, we can now be structurally optimistic—we can't be pessimistic anymore, and a new cycle is slowly opening.

Mr. Z: One more question—the rise of Perp DEXs and the ebb and flow with CEXs, and the ebb and flow among top players like Binance, OKX, Bitget, Bybit—if convenient, say a few words.

qinbafrank: Last September, I wrote about several challenges US stock tokenization poses to the crypto market. First, more and more US stock tokens will greatly squeeze the altcoin market, because US stocks have more wealth effect and are more sustainable. Crypto altcoins in the past two years are like meme coins, or even worse—suddenly pumped one day, you chase, and two days later it falls back. US small-cap stocks with real fundamentals and exploding business can hold from tens of billions to hundreds of billions in market cap; a 30% pullback won't go up and then come all the way back down. Most people aren't suited to be day traders; they're more suited to be holders who can withstand drawdowns but can't tolerate holding something that goes to zero in a year. This wealth effect will make US stock tokens squeeze altcoins. Second, it's a big challenge for exchanges—fiat assets and user operations have different logics. CEX and DEX—during Hyperliquid's pre-meme era, users were already moving on-chain, and the proportion is getting higher. But I think the future is coexistence: several major exchanges are pursuing compliance through various paths, which is more convenient and trustworthy for users. DEX has been rising since 2024, but DEX and CEX are different user scenarios—playing meme you're on-chain; trading Bitcoin, US stocks, or US stock token contracts, some people are suited to Hyperliquid, some to CEX—depends on user habits.

Mr. Z: I suddenly thought—Gate launched Gate AI and an OpenRouter, so crypto is also getting into this business.

qinbafrank: Everyone is exploring how exchanges should adapt in the AI era. Gate did Gate AI, Binance launched similar trading agents, Bitget also did, OKX's Web3 wallet focuses on on-chain automated trading. Exchanges are also exploring tokenization: Binance built its own public chain and has its own US stock tokenization; Bitget has its own token; OKX didn't build a broker chain but did aggregation—aggregating US stock tokens issued on different chains. There might be five Nvidia tokens on-chain, all aggregated together—that's also innovative. OKX also has an equity investment relationship with the NYSE. Now on Binance, US stock token contracts—I estimate that among the top ten by contract trading volume, more than half are US stock tokens or gold. These are all trends, and completely ignoring them is also wrong.

21. Mindset: Buy Low, Don't Chase Highs; When the Bull Returns Quickly, Trim a Bit
Mr. Z: We're at the end of the interview. Frank, do you have any investment mindset tips for AI or crypto, or any final words on the market for the coming months?

qinbafrank: My approach has always been to try not to chase highs, but when bottom-fishing, you must dare to buy—buy low as much as possible. In US stocks and crypto, the only thing I might be confident about is this: if I have a good asset and I'm bullish on it, I can buy when it falls; when it rises again, some people keep adding, but I stop at a certain point and just hold. Because the market is changeable, you can't be sure something won't go wrong. In July this year, I also had drawdowns—although on July 2, I saw signs and reduced a bit, and later switched to Microsoft and other CSPs, I still had drawdowns, but relatively okay. Because I was selling storage before March and early April, and after that, although I remained bullish, I didn't buy more—all positions were from before, with cost advantages, which is important. Second, when the bull returns quickly and the market is too frenzied and everyone is excited, you still need to be cautious. After a big rise, trimming appropriately is right—give your positions some flexibility, don't be too rigid, don't go all in.