QinbaFrank on AI's Second Growth Engine, CSP Value and Crypto Reform

Summary

• QinbaFrank sees US equities in a “valuation-friction zone”: with the 10-year Treasury at 4.7% and the 30-year at 5.2%, “the market is unlikely to collapse immediately, but a broad, sustained upside trend is also hard to see.” The pressure comes from oil prices and inflation expectations, the deficit, and a wave of tech-company bond issuance competing with the Treasury for funding; if US-Iran tensions ease, Bessent continues buybacks and adjusts the mix of long- and short-dated issuance, pushing the 10-year below 4.6% and the 30-year to around 5% or lower, AI, semiconductors and related assets could finally break out.
• There is no problem with AI demand: the common signal from second-quarter earnings was that “once data centers are built, racks installed and deployment complete, they can run at full load immediately”—the bottleneck is supply. Mr. Z made the same point using Nvidia’s earnings: the issue is supply, not demand. Cloud providers’ deployment of open-source models and their own mid-tier models such as Microsoft’s MAI are changing the cost structure of CSPs as “middlemen”; most revenue tied to open-source models can accrue to the CSPs themselves. Frank expects third-quarter earnings from mid-October through November to be impressive, with gross margins potentially improving materially. “The performance of hyperscale CSPs” is the most comprehensive gauge of AI commercialization.
• “Where is the next Coding?” may be the wrong question: the second growth engine will not be a single major application, but the long tail created when non-Coding workflows become agentic. Coding is “the enabling technology for non-Coding workflows,” allowing nontechnical employees to generate in natural language processes that previously were not worth turning into software. An official OpenAI Codex blog showed adoption in legal applications up roughly 100x since February, and adoption in sales and recruiting up about 30x to 40x. Microsoft Foundry has more than 100,000 customers, Microsoft 365 Copilot has 30 million paid seats, and Agent 365 has more than 40 million agents.
• The financial returns on CapEx may only be validated progressively from the second half of next year, year-end next year and into the following year. The cross-check to watch is cloud-revenue growth overtaking CapEx growth, which even Google has not achieved: cloud revenue grew 82% year over year while CapEx growth exceeded 100%. This year marks the peak in Big Tech CapEx growth: roughly $400B last year, $800B-$900B this year, and a market expectation of $1.1T-$1.2T next year, with growth slowing to 30%-40%; the other inflection point is non-Coding business volume taking over as Coding growth slows.
• Mr. Z leans toward caution in September and October and keeping more cash; Frank describes this stretch of the rally as “the chill of autumn setting in”（“秋意渐凉”）—“not winter, just every easy question turning into a hard one”（“不是寒冬，就是简单题都变成了复杂题”）. Sources of volatility include the midterm elections, with AI data centers already a political flashpoint in some regions, and the Anthropic IPO. The Information says Anthropic may file on the 7th or 8th; if it raises $80B-$100B, the roadshow could draw $300B-$500B in capital. After the midterms, policy uncertainty should fall; with third-quarter earnings in hand and SOX valued at roughly 19.2x, the fourth quarter and next year could still be favorable.
• On Bitcoin: when it was around $63,000 in June, Frank believed it had entered a “high risk-reward zone,” with June through November serving as the accumulation window for a long-cycle allocation that he planned to build out gradually. He expected a short-term rebound above $70,000 in July, but after the 19th, the price briefly broke above $80,000, exceeding his expectations in magnitude. Based on the Treasury’s refinancing plan and its TGA target, he expects liquidity to tighten slightly from late September into October, giving Bitcoin a good chance of retesting lower levels; if using his own capital, he would continue buying according to plan. The cycle is a characteristic, not a flaw, and the drawdown in each historical cycle has become smaller.
• Crypto’s supply-side shift: Frank views the SEC’s August 19 proposal for Regulation Crypto Assist (Reg CA) as “compliant ICO 2.0.” A startup could potentially raise $5M in a year, or $75M after meeting standards on transparency, filings and other requirements; it could conduct a public offering directly to retail investors without necessarily using an intermediary, then seek listings on Coinbase, Binance, Kraken and other exchanges. The proposal also seeks to separate the investment contract from the token’s asset attributes, a framework that could eventually apply to network businesses, although it remains only a proposal in the public-comment stage. Together with tokenization, RWA and AI×Crypto, Frank believes the crypto market has shifted from extreme pessimism to structural optimism.
• His closing rule: avoid FOMO where possible, but have the courage to buy the dip—buy low, do not chase strength, trim after large gains, keep a buffer and never go all in. He built his spot holdings in March and early April, completing them by early April at the latest; even though he remained bullish afterward, he did not chase, relying on a lower cost basis to absorb the June-July pullback.

Deep dive

1. Review: The July 2 report that Meta was selling compute signaled Mr. Z’s trimming; in crypto, gains were expected but the magnitude exceeded expectations

• Mr. Z revisited his earlier view that US equities were carrying high leverage and crowded positioning, with a potential deleveraging event in the second half. A July 2 Bloomberg report that Meta was selling compute raised his alert level, and he trimmed some positions; from early July, he focused his posts on X on CSPs. In the volatility of July and August, hyperscale CSPs barely declined and also held up relatively well on the rebound.
• On crypto, Frank said in an OKX livestream in mid-June that Bitcoin, at around $63,000, had entered a “high risk-reward zone”; long-term investors should build positions in stages from June through November.
• In July, he saw a short-term rebound reaching above $70,000; after the 19th, the price briefly broke above $80,000 on the 21st and 22nd. “I wasn’t surprised by the rally, but I was very surprised by its magnitude.”

2. Sources of pressure on long-term yields: oil, deficits and the tech-company bond wave

• Three conditions would improve the market: lower long-end yields, falling inflation and a Powell who is no more hawkish than he is now. After US-Iran clashes around the Strait of Hormuz, oil rose from the $70s in July to above $90 at its August peak, transmitting pressure through the chain of “energy—inflation—bond yields.”
• Deficits and tech-company bond issuance are additional sources of pressure. Large tech companies have extremely high credit ratings—Microsoft’s rating can almost be viewed as sovereign—and they are competing with the government for funds in the bond market, requiring a higher risk premium.
• The 10-year Treasury yield represents the market’s risk-free rate. Rising yields pressure high-PE and highly levered assets because investors expect their future financing costs to continue rising.

3. Powell’s Jackson Hole speech: hawkish on inflation, pushing the market into a “valuation-friction zone”

• Frank’s self-assessment of his pre-speech call was “half right, half wrong”: he expected Powell to remain hawkish on inflation, but did not expect the overall tone to be so forceful.
• Powell set out a relatively strong reaction function, saying he wanted to see inflation continue to fall and move rapidly toward target. Over the long term, a hawkish stance helps the market believe inflation can be brought under control; in the short term, it weighs on risk assets.
• The “valuation-friction zone” refers to a 10-year yield around 4.7% and a 30-year yield around 5.2%: the market can still hold, but investors are not willing to charge ahead. If the 10-year rises to 4.8% or 4.9%, and the 30-year moves above 5.3%, the market could come under real pressure.

4. September scenario: inclined to stay put, but inflation could reaccelerate

• Frank personally leans toward Powell holding rates steady in September. His three reasons are soft consumption, with Walmart’s second-quarter same-store sales growth slowing sharply; a potentially weak labor market, with payrolls consensus around 50,000 and August possibly below expectations or even in the single digits; and the five working groups Powell established in June, which are still studying monetary policy, interest rates, the inflation framework and productivity, with results due around year-end.
• The risk is not that inflation stops falling, but that it falls slowly. Frank has tracked Truflation for about a year and believes its absolute level may not be reliable, though its trend is informative; the latest real-time data have moved back close to June levels, raising the possibility of a rebound in August CPI.

5. US-Iran tensions becoming a long game: oil stays elevated, and the turning point depends on who buckles first

• Frank believes the US-Iran conflict has become protracted: the intensity will not reach the level of large-scale exchanges in March, but the conflict will not end quickly either. The two sides signed a ceasefire in June and tore it up again in July.
• Iran knows Trump cares about the midterms and public opinion. It can control the Strait, keep oil prices elevated and pressure him to make concessions; the US is trying to strain Iran’s domestic economy by blockading its coastline and tightening sanctions. Trump believes this approach works and is therefore reluctant to lift sanctions, unfreeze assets or recognize Iran’s rights in the Strait of Hormuz.
• Bessent previously helped stabilize Japan’s currency and prevent the Bank of Japan from selling Treasuries; after he expanded Treasury buybacks, yields fell for only one day. If US-Iran tensions ease and oil prices decline, support for yields would weaken; adjusting the mix of long- and short-term issuance could push yields down further. Only if the 10-year falls below 4.6% and the 30-year falls to around 5% or lower can AI and semiconductor-related assets truly break out.

6. Demand is still growing rapidly in the near term: the earnings signal is “build it and it runs at full load”

• Frank believes the most important signal in second-quarter earnings was that large cloud providers, hyperscalers and emerging cloud providers all said that once data centers are built, racks installed and deployment completed, they can immediately run at full load. That points to strong downstream demand, with supply as the bottleneck.
• Mr. Z reinforced the point with Nvidia’s earnings: the issue is supply, not demand. Frank believes near-term demand is still growing rapidly as AI penetration rises, the consumer and enterprise user bases expand, and each user consumes more Tokens.
• The two current problems are an unsupportive macro backdrop and the parabolic rise in some asset prices, creating a divergence between prices and demand.

7. The Coinbase case: entering an engineering era of “different workloads, multiple models”

• In early July, Coinbase CEO Brian described the company’s approach on X: frontier flagship models remain reserved for high-value use cases, while ordinary, high-frequency, low-complexity workloads are shifted to open-source or standard models. The result was a sharp increase in Token consumption in June while total spending fell.
• Frank sees this as a manifestation of the Jevons paradox: lower costs bring in more users and increase per-user consumption, potentially driving total spending higher. Controlling the cost of individual enterprises is a prerequisite for expanding AI usage overall.
• Coinbase has the capabilities to build model-routing, context-pruning, caching and storage systems; most companies do not, giving cloud providers an incentive to take on the broader AI operations stack. Microsoft began deploying models such as DeepSeek and Qwen in mid-to-late June.

8. Change the metric: hyperscale CSP earnings are the most comprehensive measure of AI commercialization

• The market previously focused mainly on the ARR of foundation-model companies such as Anthropic and OpenAI. In a July 8 summary, Frank proposed a new framework: foundation-model ARR remains the “purest measure of AI demand,” but hyperscale CSP performance is better suited to measuring AI’s commercial deployment.
• A CSP bill is the comprehensive operating cost incurred once an enterprise enters production, covering closed-source model APIs, open-source model inference, databases, object storage, data warehouses, networking, security, monitoring and agent routing.

9. A changed cost structure: the “middleman” narrative is under pressure, and intelligence is becoming tiered

• The old narrative was that CSPs absorb the CapEx, buy the machines and build the data centers, then resell frontier-model Tokens and APIs while model companies capture most of the profits. Open-source models carry no model-licensing cost, allowing most of the related revenue to accrue to CSPs themselves.
• In June and July, Microsoft deployed open-source models such as Qwen and DeepSeek while also launching its own closed-source mid-tier model, MAI (Microsoft AI), positioned around value for money. It may not match GPT-5.6, but Frank believes it could approach GPT-5.5 or Claude Opus 4.8 and 4.7; it is already sufficient for some workloads, while offering more control and lower cost.
• Mr. Z observed that OpenAI’s latest-generation model Astra, along with a possible Astra 2 and Astra 3, may no longer be offered to ordinary users and could instead be made more available to researchers. He also saw Claude recruiting 10,000 scientists to use its latest model and believes intelligence is beginning to stratify in some way. Frank said that trend is real.
• Frank cited China’s open-source model GLM 5.3, whose input and output prices are both $0.15 per million Tokens, as a level at which Western model providers will struggle to compete. OpenAI, Llama and others have continued cutting prices under pressure from this competition. He likened the structure to price discrimination, with different pricing for different consumer groups.

10. Third-quarter earnings preview: beyond revenue guidance, watch gross margin

• Frank is bullish on third-quarter earnings reported from mid-October through November. Microsoft’s guidance at the time implied cloud growth of 43% in the second quarter and potentially 45% in the third; AWS was at 37% in the second quarter and could reach 40% in the third; Google Cloud grew more than 80% year over year in the second quarter and could accelerate again in the third.
• The impact of open-source-model deployment and mid-tier models on the cost structure was not yet visible in second-quarter earnings. July through September is when cloud providers have been deploying aggressively, so the third quarter could bring not only revenue growth but also a meaningful improvement in gross margin.

11. “Where is the next Coding?” is the wrong question

• Coding is distinctive because code, logs, APIs and documentation are already digitized, with GitHub serving as a global code repository; outputs are easy to verify, processes can be broken into clear steps, and developer costs are high enough that saving 20%, 30% or even 40% of development time, improving productivity or reducing headcount makes ROI relatively easy to calculate.
• Coding also does not require company-wide, top-down adoption. Developers can use Coding Agents or APIs on their own, allowing penetration and commercialization to move quickly.
• Frank believes it will be difficult to find another single application with the same combination of clear boundaries, extremely rapid growth, massive revenue potential and immediate recognition by capital markets. The non-Coding world is highly fragmented: banks, insurers, hospitals, manufacturers and retailers all have their own workflows, equipment and internal data.

12. The second growth engine: agentic knowledge workflows could exceed Coding in aggregate

• Frank expects enterprise horizontal functions and knowledge-work use cases to spread first, including finance, recruiting, human resources, documents, contracts, IT operations, legal and compliance; services and back-office operations should follow, while supply chains, manufacturing and healthcare may move more slowly.
• The second phase therefore will not produce a single application analogous to Coding. Instead, large numbers of knowledge workflows and industry processes will gradually become agentic, with the aggregate eventually potentially exceeding Coding.
• Microsoft said in its second-quarter earnings that Foundry customers had surpassed 100,000, Microsoft 365 Copilot had reached 30 million paid seats and Agent 365 had more than 40 million agents. Foundry is a platform for enterprises to build, deploy and run agent applications, while Copilot reflects non-Coding employees’ adoption of the tools.

13. Coding is the “enabling technology” for non-Coding workflows

• The core point in Frank’s article is that nontechnical employees will not feel they are writing code in the future; they will use natural language to generate financial reconciliation tools, sales dashboards, internal approval pages and contract-review workflows.
• Many companies are small and have highly specific needs. In the past, they had to rely on Excel, email and manual work because dedicated SaaS or IT systems were not worth purchasing. By lowering the barrier to custom software, Coding Agents can gradually turn these long-tail workflows into software and automate them.
• The official OpenAI Codex blog said the fastest-growing adoption areas since February included legal, sales, recruiting, marketing and healthcare; legal grew by roughly 100x, while sales and recruiting grew by about 30x to 40x.
• These workflows are less standardized than Coding, and progress slows as deployment moves deeper into company-specific processes because permissions, internal data, retraining and deployment become issues. Frank understands Palantir-style FDEs as forward-deployed engineers; in June and July, OpenAI, Anthropic and other companies also began building dedicated FDE teams for enterprise customers.

14. How enterprises confirm ROI: roughly 15% of US-listed companies have already shown results

• Frank believes that if a company sees no results after one month, two months or one quarter of use, it may not renew in the following quarter; continued usage and rising consumption usually indicate cost savings, efficiency gains or revenue creation.
• He cited a well-known iOS mobile-game company backed by a China-based USD VC partner that has already applied AI across the full workflow, including creative work, art, game design and numerical systems. A large game that previously took 2 to 3 years to develop might now be completed in six months.
• From the start of the US second-quarter earnings season on July 15 through the end of August, Frank had GPT compile disclosures from traditional industries and non-tech public companies about AI use in earnings reports and calls. He estimated that roughly 15% of US-listed companies had already shown an AI-driven effect, or at least clearly felt a change. The impact need not appear immediately as revenue; it could be lower costs, or higher efficiency at unchanged cost.

15. When does CapEx break even? Batch timing and two inflection points

• Amazon’s CEO said CapEx first goes toward land purchases, data-center construction, land grading, roadwork and power infrastructure, which can take 8 to 12 months. Installing racks, optical modules and network connections afterward requires another one or two quarters.
• As a result, a dollar of CapEx takes at least one to 1.5 years to turn into an operating data center and generate cash flow. Money invested in last year’s third quarter may only begin to show results in this year’s third quarter, or even at year-end.
• Large tech companies spent roughly $400B on CapEx last year; that figure rises to $800B-$900B this year, up more than 100%. The market expects $1.1T-$1.2T next year, with growth slowing to 30%-40%. The timing of cash generation from each investment batch, and whether it breaks even financially and in accounting terms, may only become clear progressively from the second half of next year through year-end and into the following year.
• The first inflection point is cloud-revenue growth exceeding CapEx growth. Even Google, the fastest grower, has not reached it: Google Cloud grew 82% year over year while CapEx growth remained above 100%. The Google figures Frank relayed also included full-year CapEx of roughly $80B-$90B and a second-quarter figure already raised to more than $200B.
• The second inflection point is Coding growth slowing after its period of extreme expansion while non-Coding business volume begins to accelerate. Mr. Z said Anthropic’s ARR had already slowed to around $65B; Frank did not elaborate or push back, emphasizing only that the market remains in the lag between investment and returns.

16. Year-end roadmap: three phases, with “the chill of autumn setting in,” not winter

• From March through May, especially April and May, compute was scarce, Anthropic ARR rose rapidly to $30B, cloud providers reported strong first-quarter earnings, CapEx had not yet peaked, oil prices were falling and long-end yields were low, with the 10-year briefly reaching around 3.9%. The rally was therefore more explosive.
• Frank divides the year-end outlook into three phases: early to mid-September; mid-September to mid-October; and after the midterms. The market first needs to digest Powell’s speech, August payrolls and August CPI, while September and October could remain volatile under the influence of the midterms and IPOs.
• On July 31 he wrote that “August is the best summer”; on August 23 he wrote that “the chill of autumn is setting in.” This is not winter, but a macro-constrained period in which “every easy question has become a hard one.”

17. Midterms: AI data centers become a contested local issue

• Democrats may attack Republicans over Iran and AI data-center construction. Concentrated data-center development can push up electricity, water and land prices in some areas, while a large data center may employ only 10 to 20 people once built.
• Public opposition has emerged in Florida, Virginia and Texas, among other places; New York has at times suspended approvals for hyperscale data centers. The Texas state government originally supported data-center construction but became more cautious as the race tightened.
• Frank believes the issue may not change industry fundamentals, but it could affect market sentiment. After the midterms, regardless of which party controls Congress, the market will at least know the result and can rework its policy assumptions.

18. The Anthropic IPO’s “capital-sucking effect”: it may begin during the roadshow

• The Information reported that Anthropic may file its IPO documents on the 7th or 8th. The filing would give the market its first relatively complete view of the company’s actual operations, with an effect somewhat similar to Circle’s listing; if the underlying results beat expectations, sentiment could improve.
• The roadshow, however, could drain capital from elsewhere. If Anthropic raises $80B-$100B—even above SpaceX’s $75B valuation at the time—institutions may need to prepare $300B-$500B to subscribe.
• After the listing, if the public float is only 5%-6%, the stock could continue rising and create short-term pressure on other AI assets before the broader market recovers.

19. Why be bullish on the fourth quarter: earnings, valuation and the post-midterm pattern

• The SOX index trades at an average valuation of roughly 19.2x; it was around 17x-18x during last year’s tariff war and around 16x-17x in August 2024. Frank believes the current level already offers some value, and if the market continues to churn while earnings grow, valuation could compress further.
• Third-quarter earnings may remain strong, and the shock from Anthropic’s listing could gradually fade. Historically, market performance 30, 50 and 100 trading days after the 2010, 2014, 2018 and 2022 midterms was generally solid, with technology and growth assets often benefiting more.
• Mr. Z personally leans toward caution and holding more cash in September and October; he said his maximum drawdown after entering the market in late July was around 4%-5%. He also noted that Taiwanese investors worry the “National Stabilization Fund” or government capital may exit after local elections at the end of November, leading some to expect Taiwan equities to diverge from the US market.
• Frank responded that the US market is worth roughly $70T-$80T versus around $5T for Taiwan equities, with different levels of marketization and different degrees of capital-driven price action.

20. Interlude: Mr. Z questions the volatility decay of leveraged ETFs

• Mr. Z believes Taiwan equities performed well in August, with foreign investors returning more aggressively than to Korean equities. Leveraged ETFs linked to Samsung and SK Hynix attracted roughly $45B-$50B, but the flows did not show up in the underlying stocks, suggesting institutional selling pressure overhead.
• He believes leveraged ETFs are essentially short volatility, and ordinary investors may not understand the mechanics of their decay. He was also skeptical of recent reports that Temasek had invested in SK Hynix.

21. Bitcoin: the cycle is a feature, not a flaw; a retest may come from late September into October

• Frank believes Bitcoin is closely tied to liquidity centered on the dollar and bank reserves. Based on the Treasury’s refinancing plan, funding arrangements and target TGA balance, liquidity may tighten slightly from late September into October, so he expects a pullback but cannot determine whether it will make a new low or a higher low.
• If using his own capital, he would continue buying according to schedule, aiming to complete the planned position before November, or by October at the latest, and hold for 1 to 3 years.
• Bitcoin runs on an approximately four-year cycle, and the cycle is a characteristic rather than a defect. As issuance declines, the cycle effect may weaken; from the 2010 and 2011 halving cycles through 2014, 2018, 2022 and 2026, the drawdown in each cycle has indeed become smaller.

22. The new altcoin normal: extreme dispersion, with most assets reduced to “shell value”

• Frank’s mid-2024 view was that crypto would become increasingly bifurcated over the following years: only a very small number of assets would sustain strong performance, while most would gradually lose attention and eventually be left with only “shell value” on exchanges.
• He used US equities as an analogy. At the time, the US market had around 5,000 companies; the bottom 3,000 accounted for only around 5% of total market capitalization, with average market caps of roughly $800M-$900M and daily trading volume of around $1.5M-$2M. Large markets also become highly differentiated.
• Crypto’s problem is not too many quality assets but too few quality assets and too many low-quality ones. Capital markets need a continuing supply of quality assets to sustain a strong trend; US equities achieved this through淘汰, index iteration and absorbing high-quality companies from around the world.

23. Tokenization expands DeFi’s building blocks; crypto-native applications are Perp DEXs and DeFi

• Frank sees tokenization and RWA, the combination of AI and Crypto—including payments by AI Agents—and compliant ICO 2.0 as key directions for the next phase.
• After Robinhood launched an Ethereum-based L2 in July, users on-chain began experimenting with trading pairs combining Meme coins with SPY tokens and Nvidia tokens. Gold, Nvidia, SPY and QQQ could all be tokenized in the future.
• Historically, the only genuinely high-quality assets suitable as on-chain collateral were mainly Bitcoin and Ethereum, with Solana qualifying to some extent. As more US-equity tokens enter the chain, the supply of quality collateral will expand, giving DeFi’s set of “financial building blocks” room to grow.
• Frank sees crypto more as a capital-market paradigm that improves the issuance, circulation and trading efficiency of assets; it must attach to real business use cases. Perpetual contracts, Perp DEXs and DeFi are more representative crypto-native applications than concepts such as decentralized social media.

24. Reg CA: a supply-side shift Frank views as “compliant ICO 2.0”

• On August 19, the SEC proposed Regulation Crypto Assist (Reg CA). Frank believes it could be one of the important factors behind Bitcoin’s rise, alongside Treasury buybacks, the White House technology leaders’ meeting and short squeezes.
• Past fundraising routes included IPOs, Regulation A mini-IPOs, Regulation CF crowdfunding and Regulation D private placements. Reg A offerings typically reach OTC or pink-sheet markets, while Reg D is mainly for private placements and accredited investors, with limited liquidity.
• Under the Reg CA proposal, a startup could potentially raise $5M in a year; after meeting standards for transparency, filings and other requirements, it could potentially raise $75M in a year. It could offer directly to retail investors without necessarily using an intermediary, potentially raising funds through its own website before seeking a listing on Coinbase, Binance, Kraken or other exchanges.
• The proposal seeks to separate the investment contract from the asset characteristics of the token: a project could first promise to build a public blockchain, achieve decentralization or meet another objective; if it fulfills that promise within several years, the investment contract could terminate and the token would no longer be treated as a security.
• The framework could eventually apply to compute networks, WiFi networks, gaming networks and other network-based or multisided-platform businesses, giving companies with real operations but not yet at IPO scale a new funding channel. The proposal remains in the public-comment stage and may not formally take effect until October; compliance and transparency do not mean a project is good, only that the odds of quality projects emerging may improve.

25. Who benefits: public chains, RWA and DeFi infrastructure

• Mr. Z sees Tether, Hyperliquid and the prediction-market competition between Kalshi and Polymarket as projects worth watching. His summary was that US equities are beginning to take on crypto characteristics, while crypto is beginning to take on US-equity characteristics, with increasing attention to cash flow and fundamentals.
• Frank focuses on Solana, BSC and Robinhood. Solana has a strong Meme ecosystem, but its “internet capital market” thesis has yet to produce a clearly realized application; BSC benefits from Binance’s backing, with projects such as Aster and Predict.fun supported by Binance, but stablecoins remain concentrated on Tron and Ethereum, so it remains to be seen whether BSC can develop a genuinely valuable on-chain ecosystem filling that gap. Robinhood has regulatory advantages and a retail user base, and its main direction may be RWA.
• Among US-equity-related assets, Frank watches Circle, Coinbase and Robinhood. He sees Coinbase as more compliance-focused and methodical, while Robinhood is more of a “bad boy”—better at seizing narratives and more willing to act first and explain later. Base could also benefit from tokenization and compliant ICO 2.0.
• Frank is positive on Ethereum, Solana, Base and BNB, as well as RWA- and DeFi-related infrastructure such as lending, exchanges and oracles. These projects have gone through multiple cycles, giving their security, application value and business volume some degree of validation.

26. US-equity tokens squeeze altcoins; CEXs and DEXs will coexist

• Frank believes US-equity tokens offer a more durable wealth effect. Genuine small-cap stocks with fundamentals and growing businesses can rise from a few billion dollars to several hundred billion or even $100B and hold those gains after a pullback; altcoins and Memes often rise fast and fall fast, leaving late entrants on a roller coaster.
• He expects CEXs and DEXs to coexist. Memes and similar use cases may be better suited to on-chain markets, while some users will choose on-chain trading for large coins, US equities or equity tokens and others will choose centralized exchanges with deeper resources.
• Every major exchange is exploring AI, tokenization and direct brokerage connectivity. Gate has Gate AI and OpenRouter; Binance and Bitget have launched Agent-type products; the OKX Web3 wallet is exploring automated on-chain trading; Binance and BingX have built broker links; Bitget has R Token; and OKX aggregates US-equity tokens across different chains. Binance has also launched US-equity-token products such as BSOX.
• Trading volume in US-equity-token contracts has moved into Binance’s top tier. More than half of the top 10 may be equity tokens or tokenized contracts on assets such as gold and crude oil. AI and tokenization are trends exchanges cannot ignore.

27. Closing rules: no FOMO, buy the dip, keep a buffer

• Frank’s rule is to “avoid chasing strength where possible, but have the courage to buy the dip”: buy favored assets when they fall, then stop adding after they reach a certain level, or continue holding without chasing.
• He built his spot holdings in March and early April, completing them by early April at the latest; although he remained bullish afterward, he did not keep adding, and the lower cost basis helped him maintain an advantage through the June-July pullback.
• The second rule is to keep a buffer. When the market becomes excessively euphoric, investors should be more cautious, trim after large gains, keep positions flexible, avoid building them too rigidly and never go all in.