168X in Conversation with QinbaFrank: AI Commercialization, Anthropic’s IPO, and the Q4 2026 Market Window
Summary
- QinbaFrank believes AI infrastructure can still support U.S. nominal growth, but energy inflation will determine how many more rate hikes the market can absorb. He estimates big tech capex will rise from $400B–$500B in 2025 to about $900B in 2026, allowing a strong macro backdrop to coexist with weak spending among lower- and middle-income households. The pause in rate hikes reflects “a hawkish consensus to be patient”; if the U.S.-Iran standoff drags on and high oil prices spread to industrial goods, the risk shifts from the one or two hikes markets can still absorb to a longer tightening cycle.
- He still sees AI commercialization accelerating, citing actual revenue and cloud growth rather than eye-catching annualized run rates. Based on Anthropic revenue of $16.2B in the first half and $16.8B in Q3, Frank projects roughly $50B–$60B for the full year and sees a strong chance it reaches $200B in 2028. Microsoft cloud growth of 48% and Google Cloud growth of 103% support that view. He stresses that full-year actual revenue and an annualized run rate “are completely different in quality”; large-scale commercialization has lasted only 3 quarters, and he sees no sign of saturation yet.
- The debate over AI investment returns comes down to when money spent a year ago turns into capacity, revenue, and cash flow today. Frank’s “time lag” framework runs from buying land and building power and data centers, to bringing compute online and fulfilling orders, and finally to cash returns; current-quarter capex should not be compared directly with current-quarter revenue. The next key inflection point is cloud revenue growth outpacing capex growth. But he acknowledges that his belief the numbers work does not mean the market has enough evidence yet.
- An Anthropic IPO could first create a rebound window, then trigger institutional repositioning and divert capital after the listing. A roadshow starting November 9 and a listing in the week of November 16 are assumptions discussed in the conversation; Frank stressed, “Assume it’s November 9.” If the company raises $100B and demand reaches about 4x the offering, investors may need to line up roughly $400B. He uses SpaceX’s listing to think through the “siphoning” effect, but treats neither the timeline nor the subscription scale as confirmed.
- The AI trade is shifting, and AI, semiconductors, and memory can no longer be treated as interchangeable. Frank says memory, the strongest area in the first half, had given way by August and September to optical modules and some software. Model commoditization and falling token costs could also benefit cloud platforms, enterprise deployment, agents, and high-value applications. “AI isn’t just semiconductors, and semiconductors aren’t just memory.” The old leaders could rebound, but their early straight-line gains should not be extrapolated.
- High long-term bond yields could make Q4 a stock-picking market even as the indexes set new highs. Energy inflation, Treasury supply, and AI companies issuing debt are all competing for capital, making long-term yields “more likely to rise than fall” and investors less tolerant of cash-flow and earnings blemishes. Frank believes gains in mega-cap tech can rotate from one stock to another and support the indexes. But if an energy shock forces more hikes, or cloud growth slows for 2 consecutive quarters, both the broader market and the AI trade need to be reassessed.
- Warsh’s institutional reforms deserve closer attention than any single rate decision, but the details will have to wait for the working-group reports. With less focus on rate communications, Frank is watching 3 questions: whether the 2% inflation framework changes, how AI productivity will be measured, and whether the Fed changes its ample-reserves regime and large balance sheet. He says the working groups “aren’t just for show,” without claiming the target will definitely change or the balance sheet will shrink sharply.
- For crypto, the short-term trade depends on liquidity returning; the long-term themes are compliant fundraising and traditional assets moving on-chain. Frank puts the Treasury General Account at about $1.02T and says the crypto market could recover if funds start flowing out in November. He describes October’s pullback as “a normal retracement in a larger uptrend.” On regulation, he distinguishes fundraising proposals from innovation exemptions: even compliant ICOs may not be possible until the first half of next year, while tokenized securities, Treasuries, and stocks could provide an ongoing tailwind for on-chain and DeFi infrastructure.
Deep dive
1. AI infrastructure supports nominal growth as consumer spending diverges
Frank breaks the U.S. economy into two scorecards: nominal GDP and consumers. He estimates big tech capex will rise from $400B–$500B in 2025 to about $900B in 2026, with the incremental spending alone potentially adding roughly 1.5–2 percentage points to GDP growth. As the “AI infrastructure buildout” accelerates, it supports aggregate growth; weakness in some areas of consumer spending is not enough to call a broad recession.
His analogy is China a decade ago. China’s growth rested on consumption, exports, and infrastructure, with property playing a central role; in the U.S. today, AI infrastructure is the key driver. He also cites Warsh’s remarks that U.S. GDP growth in the first half was revised up from 1.8% to 2.4%, with further acceleration in Q3. That is one reason Frank believes the macro backdrop is “fine.”
Strong macro growth does not mean everyone is better off. Frank says spending remains robust among middle- and high-income households but weak among lower- and middle-income ones, while AI-driven labor displacement could deepen the “K-shaped” divide. He cites Microsoft’s voluntary buyout offer to about 60,000 employees, distinguishing it from forced layoffs. Power and electrical contractors are among the infrastructure beneficiaries; other industries may not share the same growth.
2. Middle Eastern oil flows have recovered, but that does not mean prices can fall in step
Frank divides this year’s energy shock into 2 phases. Before July and August, the main issue was supply disruption from the closure of the Strait of Hormuz; since September, it has shifted more toward transportation costs. Citing third-party data, he says daily oil shipments out of the Persian Gulf in late September were close to their level before the war began in February. But “oil prices haven’t come down because transportation costs are still too high.”
Alternative routes include a Saudi pipeline to the Red Sea carrying about 6M–7M barrels a day, and a UAE pipeline to the port of Fujairah carrying about 3M barrels a day. He also describes oil-producing countries chartering very large crude carriers, each able to load about 2M barrels, to move oil through the strait. These workarounds restore volume but do not eliminate escort, risk, and transshipment costs.
The key comparison is the cost of shipping Persian Gulf crude to Asia: about $3–$4 a barrel at the start of the year, versus roughly $35–$40 now, while Brent remains above $110 a barrel. Tankers gather near the strait and sail out at night along the Omani side under U.S. military escort, before transferring their cargo to ocean-going carriers; that leg is the most expensive. Shipowners face attack risk, and producers accept high freight costs to keep customers. So oil moving out of the region is not the same as low-cost supply returning.
3. Refinery bottlenecks and AI infrastructure add a second lag to disinflation
The host adds that refining capacity has been damaged, noting that Ukraine struck many refineries during its war with Russia. Frank had also mentioned that many Middle Eastern refineries were destroyed, and that Ukraine and Russia hit numerous refineries on both sides in August and September. Even if crude shipments recover and freight costs fall, refining capacity will take time to come back, so finished-fuel prices may not decline right away.
In Frank’s account, energy remains the main driver of this year’s inflation rebound: gasoline and diesel feed through to chemicals and industrial inputs, while core inflation has not yet seen the same degree of spillover. AI infrastructure is a separate, longer-term force pushing up prices for electricity-related products, upstream materials, and some goods. He does not conflate the two: energy is the main issue now; AI-related pressure has emerged but is not yet especially visible.
He also attributes part of September’s CPI decline to a change in statistical weights: several categories linked to AI infrastructure were given lower weights, with a potentially larger effect on PCE. That is Frank’s interpretation of the data. His point is not that price pressures have disappeared, but that headline readings may diverge from the cost pressures the market sees in the infrastructure buildout.
4. The rate pause is a hawkish camp willing to wait a little longer
Frank rejects reading the FOMC’s decision to hold rates as a dovish turn. He calls it “a hawkish consensus to be patient.” In his framework, policymakers still give inflation significant weight; they are simply willing to wait for more data. If high oil prices stem mainly from a one-off geopolitical shock, easing tensions could mean further hikes are unnecessary.
The optimistic path is for the U.S. and Iran to de-escalate around the midterm elections, oil prices to fall materially, and inflation to improve in November and December. Even then, Frank says the Fed may simply hold, because it will need to confirm that lower energy prices are feeding through to inflation. He does not present de-escalation, lower oil, and an easier policy stance as an automatic chain of events.
The pessimistic path is for the conflict to continue through November and December, with energy costs gradually feeding into chemicals, industrial goods, and broader prices. A shock that initially looked temporary would then become a rebound lasting several months. Frank thinks the market can absorb 1 or 2 hikes, but if another comes in December and more follow next year, concern would rise sharply. That is one of the bigger macro risks in his view.
5. A U.S.-Iran breakthrough is more likely after the midterms
Asked when the war might end, the host says the U.S. campaign has been “incoherent” and questions whether Israeli intelligence and Benjamin Netanyahu’s political motives have influenced decisions. Frank gives no ceasefire date. Instead, he points to the week or two after the midterm elections, or November as a whole, as the key window.
Frank says he needs to revise his earlier scenario. He had thought Iran had far more staying power than the U.S., and that Trump might back down first before the two sides reached an accommodation. But backing down before the election would damage Trump’s strongman image and leave him “boxed in.” Frank believes Iran’s centralized system has been underestimated: sanctions cause public hardship but may not quickly force the leadership to concede.
After the election, he sees 2 possible paths: a direct concession, or another heavy strike followed by a concession. Frank does not think additional bombing would necessarily change the outcome; the U.S. already carried out a strike in March and April, and it was more intense then. Without ground troops, it would be difficult to inflict serious damage on Iran; a ground intervention, meanwhile, risks becoming a Vietnam- or Afghanistan-style quagmire. He hopes tensions ease but leaves open the possibility of renewed fighting.
6. Military threats and diplomatic efforts are moving forward at the same time
On reports that the U.S. military is preparing a major post-election strike, Frank says the preparations could be real. He points to at least 2 carrier strike groups in the Middle East, another amphibious assault group, and reports that a third carrier could head there. At the same time, he sees these signals as a form of “maximum pressure.” The targets are not just Iran, but also any countries unable to absorb another oil-price surge.
His reasoning is that Iran is far weaker than the U.S. overall but could still hit specific targets such as tankers, refineries, and pipelines. A new U.S. strike would raise energy risks and prompt Gulf states and other powers to step up mediation. Frank mentions senior officials from Pakistan, Saudi Arabia, and Turkey taking part in contacts, along with signs that China might mediate. He speculates that regional countries are also discussing the postwar security order.
Frank also cites reports that Vance had communicated with Iran’s national security council secretary about a negotiating framework. Direct communication and third-party mediation happening at the same time are reasons to think de-escalation is possible, he says. In his view, the sides may already agree on some terms; the sticking point is sequencing. Iran wants the U.S. to lift its blockade, sanctions, and asset freezes, while the U.S. wants Iran to lift its blockade of the strait first. Both sides fear making the first concession and then being left without a deal.
Looking back at how the war began, Frank says Trump extrapolated from the success in Venezuela and overlooked the differences between the 2 countries. He also revisits his view after Khamenei was attacked: removing a leader might not make Iran more conciliatory; it could instead bring a harder-line faction to power. That is why he doubts another show of military strength alone would resolve the problems of trust and political terms.
7. Warsh’s pullback from rate communications is part of broader institutional reform
The host turns from the war to Warsh’s management style. Frank describes a plan to reduce the number of policy meetings from 8 a year to 7 next year, with one meeting devoted to no discussion of rate policy. In 2028, 2 more meetings would be cut, leaving 6 rate meetings. He sees this as a sign that Warsh does not want markets to revolve so heavily around the next rate decision.
Frank puts particular weight on the 5 working groups established after Warsh took office in mid-June. Their remit includes the inflation framework, AI productivity, the relationship between the balance sheet and inflation, and external communications. He says the groups include economists and people from other central banks. They were initially expected to report in December; he believes they “aren’t just for show,” but are studying real policy questions.
On communications, Frank sees Warsh favoring less guidance, opposing forward guidance, and not publishing the dot plot. More officials may choose to skip the dot plot or avoid extensive comment on the next meeting, he says, instead responding to the data at each meeting. But fewer promises do not mean the Fed has already chosen an easier or tighter direction.
8. The inflation framework, AI productivity, and reserve system are the 3 reform priorities
The first question is the 2% inflation target and how it is measured. Frank recalls that under Powell, the Fed moved from a symmetric 2% target to an average-inflation framework, under which above- and below-target readings could offset each other over time. He is asking whether the current framework could change again; he does not claim Warsh will definitely raise or abandon the target.
The second question is AI productivity. Frank sees technological progress as a long-term deflationary force: the same inputs produce more goods, lowering the price per unit. But the near-term AI investment boom strains power, materials, and equipment supply, pushing prices up. The policy challenge is to determine when AI is genuinely raising productivity across the economy and how to balance near-term supply constraints against long-term efficiency gains.
The third question is the large balance sheet and ample-reserves regime. Frank thinks Warsh is unlikely to make a drastic reduction in the balance sheet before the working groups report. The real questions are whether the Fed will shrink it further next year and change the ample-reserves system in place since 2008. In Frank’s view, that could make the banking system more active. The plan’s specifics, scale, and direction will depend on the reports and Warsh’s comments.
9. Anthropic’s actual revenue is a better test of commercialization than its annualized run rate
The host turns to how much revenue the investment can generate. Frank first defines what to watch: Microsoft, Amazon, and Google Cloud, along with Anthropic and OpenAI revenue, offer a more direct test of commercialization than Meta, which mainly uses compute internally. Meta has set up an enterprise business platform, but its compute is still used mostly for its own models and advertising, so it cannot be judged by exactly the same standard.
Frank cites Anthropic’s disclosed figures: about $4.5B in revenue in 2025, roughly $16.2B in the first half of this year, and about $16.8B in Q3, for approximately $33B over the first 3 quarters. Even if Q4 revenue comes in at only about $17B, full-year revenue would be around $50B. That is why he projects a $50B–$60B range, without treating it as company guidance.
He rejects comparing that $50B–$60B in full-year actual revenue with OpenAI’s reported $100B ARR: “Those 2 metrics are completely different in quality.” At current growth rates, he sees a strong chance Anthropic crosses $100B next year and reaches $200B the year after. He does not expect the company to sustain the thousandfold percentage growth it saw at the outset.
10. Microsoft’s results offer evidence of both broader enterprise demand and cash flow
Frank says Microsoft’s results were “very good”: cloud revenue grew 48% year over year, and free cash flow came in at about $18.6B, standing out while other big tech companies face cash-flow pressure. He also mentions more than 6,700 orders, about 30% of which are expected to convert to revenue in the next 12 months. He does not specify the unit for the total value of the orders.
The shift in enterprise customers is more persuasive than a single model customer. Frank recalls Microsoft’s Q2 disclosure that roughly 90% of the increase in its remaining performance obligations came from customers outside the frontier-model companies. That suggests incremental demand for compute is not supported only by OpenAI and Anthropic; traditional enterprise adoption is also producing visible orders and helping broaden demand.
This quarter, he also points to Foundry revenue reaching 3x its year-ago level, about 44,000 paying customers, and a 75% sequential increase in enterprise consumption revenue from Q2 to Q3. Frank says cloud acceleration and enterprise adoption are happening at the same time. Microsoft cannot be viewed simply as a channel reselling compute to a handful of major AI labs.
11. Cloud platforms are becoming the enterprise AI orchestration layer
Frank walks through a Q&A from Microsoft’s earnings call. An analyst asked how the company would divide work among its OpenAI partnership, its own models, and open models such as DeepSeek and Qwen. He paraphrases Nadella’s answer: select models based on the task, sometimes call multiple models for the same job, then shift mature tasks to open or in-house models to optimize cost.
Frank mentions Microsoft’s proprietary closed model, Microsoft AI, launched in June, but his focus is not on which model ranks highest. Enterprises do not have to commit to one model and may not care which one is used. They care that the work is done reliably and well, and how much each task costs. Models can rotate, but task outcomes, costs, data connections, and permissions need to be managed centrally.
That makes cloud providers what Frank calls the “AI operating system”: they handle task scheduling, model routing, data connections, and access controls, rather than just selling an API for one model. He also cites the claim that about half of Anthropic’s revenue this year comes through cloud platforms. In his view, the platforms can serve demand for closed, in-house, and open models, making them an important place to track AI commercialization.
12. Google’s results fell short mainly on ad monetization and headline cash flow
Frank does not dismiss the concerns around Google’s results: revenue and earnings missed expectations, and the stock fell sharply. But he says the results need to be unpacked. Cloud revenue grew 103% year over year, topping $30B for the quarter, with operating profit above $10B and margins still rising. Even including TPU sales, he cites the company’s view that most cloud revenue comes from GCP, so the growth cannot all be attributed to chip sales.
The problem is that AI answers and conversational features in Search are drawing a lot of queries but have not monetized as well as expected, weighing on the ad business. Of Google’s more than $120B in quarterly revenue, about $80B–$90B still comes from advertising. A fast-growing but smaller cloud business may not offset underwhelming monetization in the core business. That is why Frank acknowledges the results had a blemish.
Free cash flow was about negative $9B, worse than the market’s expected negative $3.5B. Frank breaks out 2 factors: a roughly $3.2B increase in TPU inventory, and about $5.2B in settlement costs for a 2018 Europe-related lawsuit. Add those back and Q3 free cash flow would be roughly flat. His view is that the result “wasn’t that bad,” not that cash flow has returned to strength.
He also acknowledges that the market may not make those adjustments. When the macro backdrop is poor and rates are high, investors see pressure on the core business and a cash-flow miss, then sell first. The contrast between Microsoft’s strong results and Google’s shortfall reflects a market demanding better-quality earnings delivery.
13. Large-scale commercialization has lasted only 3 quarters, with penetration and spend per user still rising
Frank says the market can mistake the duration of the AI investment boom for the maturity of commercialization. In his timeline, big tech began ramping capex sharply in Q3 last year, Anthropic revenue started accelerating in Q1 this year, and cloud results showed token demand more clearly in Q2. “Large-scale commercialization has only been going on for 3 quarters,” so he sees no evidence yet that growth is saturating.
He uses the internet’s 2 revenue drivers: more people using the product, reflected in penetration, and more spending by each person and company, reflected in ARPU and token consumption. He has previously suggested penetration could rise rapidly from about 10% toward 40%. Once that trend takes hold, he believes adoption and deeper use can reinforce each other, with token consumption rising for both individuals and businesses.
Supply is also limiting how much revenue can be recognized. Frank says big tech has consistently reported over the past few quarters that “demand exceeds supply” and compute is scarce. If demand has not been met, new capacity could turn into revenue as soon as it comes online. That is why he sees commercialization still accelerating and far from a slowdown, though the view depends on demand remaining strong.
14. Capex and revenue cannot be reconciled quarter by quarter
Frank repeatedly comes back to the “time lag”: incremental cloud revenue today often depends on spending a year ago to buy land, build data centers and power infrastructure, and purchase chips and servers—not on capex added today. Construction ramped up in the second half of last year; capacity may come online in volume from the second half of this year through next year. When demand exceeds supply, released capacity can itself accelerate revenue.
His sequence for evaluating returns is capex, capacity coming online, orders converting to revenue, then depreciation and cash returns. On offense, ask whether revenue is accelerating; on defense, ask when cash flow will improve. That does not give companies permission to spend indefinitely. It avoids using construction-phase cash flow to judge the economics of a mature operating business. Frank acknowledges the market needs more data before it will accept the numbers he believes add up.
The next inflection point is cloud revenue growth exceeding capex growth. Frank uses Google as an example: capex was about $80B–$90B in 2025, with this year’s guidance at $195B–$205B, more than 100% growth. Cloud growth rose from a lower rate to 82%, then 103%. If Q4 growth reaches 110%–120%, cloud growth could outpace capex growth. That is his projection, not something that has happened yet.
15. Returns also depend on how long data centers and GPUs remain productive
Frank cites Amazon management’s framework: an investment takes about 2 years to build and enter service, starts paying back in 3 years, then earns for another 3. The buildings and power infrastructure can keep operating; only the chips need to be replaced. The composition of capex changes as a result.
He revisits the debate over extending GPU depreciation from 4 years to 6. Critics worried GPUs might be too slow after about 3 years, making the accounting useful lives too optimistic. Frank points to the H100: roughly 6 years after its launch, it remains attractive on a performance-per-dollar basis, and rental prices are still rising. His focus is that older chips can still earn money on tasks they are suited to.
He also cites H100 orders CoreWeave signed in Q2 that run through 2029, suggesting the chips could have a potential useful life of 9 years rather than just 6. But he does not treat one order as definitive proof of industry-wide returns. He wants to see the next round of results from Amazon and other neoclouds to better understand the link between capital spending, utilization, and cash recovery.
16. Demand after Codex does not need to come from one new blockbuster use case
On concerns about whether there is a second use case after Codex, Frank says it does not have to be a single use case. Enterprise demand can come from finance, legal, tax, marketing, human resources, and vertical workflows across industries. Each application may be smaller than Codex, but together they could be just as significant.
He also thinks that as Codex and Web Codex improve, they will support many applications beyond Codex itself. Personal AI agents are another opportunity. Frank notes that Meta, Google, Grok, Anthropic, and OpenAI have all launched their own agents since September, and sees strong potential in personal AI agents too.
At the higher-value end are scientific discovery applications such as AI drug development. Frank points to increased investment by several major companies and acquisitions of AI healthcare companies by OpenAI and Anthropic. These tasks could consume a lot of compute and tokens. He presents them as possible paths for demand to deepen, not as proof that every application is already commercially successful or can support all capex on its own.
17. Neoclouds risk borrowing to build ahead of demand
Frank describes the neocloud business model as taking on substantial debt to position for future demand: borrowing to build compute capacity before the revenue has fully arrived. Big tech has cash flow from other businesses as a buffer; neoclouds have weaker credit profiles and less ability to absorb pressure. Market concern about default risk is therefore reasonable. Strong AI demand does not erase financing risk.
But Frank still says to start with demand: are penetration and spend per user rising, and can new compute be absorbed? Then look at whether revenue converts quickly enough to cover construction and financing costs. He believes both demand indicators are currently rising, but acknowledges that there is no conclusive evidence yet to convince the market every company’s cash flow will eventually improve smoothly.
18. Anthropic’s IPO timeline is still hypothetical, and the offering size could change the trading rhythm
The host mentions a possible listing in the week of November 16 and the market around a November 9 roadshow. Frank first acknowledges that he had previously expected Anthropic to list in mid-November. He then explains that the November 9 roadshow date is an estimate based on the filing date and U.S. SEC rules, not a final date. He reminds the host: “Assume it’s November 9.”
Neither the valuation nor the size of the raise is settled. Frank is discussing a scenario in which Anthropic raises $100B. He calls it a “pure-play” AI model company, with attractive actual revenue and growth, which could draw institutions willing to make a bet. He also mentions market concerns about its roughly $600B–$700B in compute commitments between 2031 and 2034.
Frank says those commitments need to be assessed alongside the revenue trajectory: the view of its ability to meet them changes if it does $50B–$60B this year and keeps growing. He thinks revenue performance is “pretty good,” but does not claim all long-term commitments are already backed by revenue, or treat the possible IPO size, listing date, or subscription multiple as settled facts.
19. The IPO’s “siphoning” effect could start with institutional repositioning, then capture trading attention
Frank looks to SpaceX for comparison: it raised about $75B, with investor subscriptions of roughly $300B–$400B, or around 4x the offering. If Anthropic raises $100B with similar demand, investors may need to line up roughly $400B. Institutions may not have that much idle cash, so they could sell other positions to free up money for the IPO around the start of the roadshow or a day or 2 later.
Once the stock lists, subscription capital will be locked up, and some of it may later be released. But the “siphoning” effect might not end there. Frank says a hot new stock could continue attracting retail money once it trades. He compares this with pressure on space-related stocks after SpaceX went public. The release of subscription money and a return of capital to old holdings are not necessarily the same thing.
Before then, Frank sees room for a rebound. Bitcoin began falling around September 27, 28, and 29; U.S. stocks started falling around October 6 and 7. The market has been down for 3 straight weeks, and many stocks have fallen sharply. Microsoft and Google cloud results suggest AI commercialization is broadly healthy, though Google’s core business still has issues and Amazon’s results are still due. That could support a short-term rebound, but the macro backdrop may not improve after Anthropic lists, leaving room for risk aversion or a “siphoning” effect.
20. The AI trade has shifted from memory toward optical modules and software
Frank has a clear warning for investors: “AI isn’t just semiconductors, and semiconductors aren’t just memory.” Looking back at his view in late July, he thought optical components should be the strongest area. Memory had already fallen too far for shorting to offer good risk-reward; there was still upside, but the explosive phase might have passed. Optical modules and some software were stronger in August and September.
He thinks big tech stocks could trade in patterns like 3 days up, 1 day down, or 3 days up, 2 down. Momentum stocks were still making new highs in September. His point is not that the original leaders are finished, but that the straight-line gains of the first half could give way to choppier trading. Investors need to distinguish between stocks with room to recover and those still in the strongest trend.
The constraint on semiconductors is that the market is already asking whether capex will slow in 2028 and where demand will come from after that. Frank says he also cannot see how long AI demand can last, so he will not claim upstream valuations can expand forever just because activity is accelerating now. At the same time, model commoditization and falling token costs could shift more value toward cloud platforms, software, and applications, with the leaders continuing to evolve.
21. Enterprise deployment and proprietary data could push value further downstream
Frank sees CSPs as the foundation AI runs on, with non-generic software and SaaS companies that control permissions and unique data flows acting as the execution layer. Some software companies that were heavily sold off in the first half have since strengthened. He sees that as support for his earlier view that cloud providers would be revalued and that downstream software should not be written off across the board.
Enterprise AI adoption is not as simple as installing standard software. Companies need to map information, business, and cash flows, then embed AI in management workflows. Businesses with proprietary data may need to train their own small models; technology companies can do this themselves, but many traditional businesses cannot. Companies that help with deployment, data connections, and workflow redesign could be among the next beneficiaries.
He cites what Coinbase CEO Brian said in early October: Coinbase used internal data in September to train an anti-fraud model that could already match GPT-5.5 or Opus. Frank says he cannot remember which one. The example illustrates the value of proprietary data and task-specific optimization; it does not claim that one general model beats others at every task.
22. Long-term yields are more likely to rise than fall, making investors pickier about earnings delivery
Frank lists several forces behind high yields: oil prices raising inflation expectations, a rising U.S. fiscal deficit and Treasury supply, and a wave of AI infrastructure borrowing. Tech companies are issuing mostly 5- and 10-year bonds, competing for capital in the bond market. More supply means investors demand a higher risk premium, leaving long-term yields “more likely to rise than fall” and weighing on equity valuations.
That does not mean stocks cannot rise; it means they need stronger fundamentals. Frank says to separate narrative from actual business performance: look at revenue growth, whether there is a basis for growth over the next 2–3 years, and whether cash flow can improve. He repeatedly cites the difference between Microsoft and Google after their results. The market wants “no blemishes”; even a modest shortfall can sour sentiment, making investors more selective.
He also sees opportunities when good companies face headwinds. When comparing cloud providers in July, he favored Microsoft and Amazon. Microsoft had fallen from above $500 to around $370–$380 in the first half, while Amazon had not gained much; Google had already performed strongly. His reasoning is that core assets and competitive strengths remain. Temporary headwinds, lower valuations, and a falling share price can create a better entry point, but a stock is not automatically a buy just because it has fallen.
23. Q4 could be highly differentiated, while the indexes still have a path to new highs
The host mentions fund managers potentially closing positions before Thanksgiving in late November and the trading around Christmas in late December. Frank agrees that individual stocks could diverge sharply. Falling oil would provide an additional lift; if the problem remains unresolved, a repeat of Q2’s strength would be hard. Even if energy improves, Treasury supply and other factors remain, so long-term yields may not fall quickly from the 5%-plus range to 4.5%.
Frank still thinks new highs for the S&P and Nasdaq are “quite possible,” because high-weight stocks can recover in turn and indexes regularly adjust their constituents. He points to Google: fears that AI would replace Search had been intense, then expectations reversed after the release of Gemini 3. Google rose from about $160–$170 to a high near $380. One company facing headwinds does not mean all the biggest stocks have lost momentum at the same time.
But Frank says AI is entering “the second half,” when investors can no longer expect every participant to benefit. In the early phase, everyone talks about “the vast potential”; in the later phase, some players will be eliminated. The internet and mobile internet both went through that process. Investors can wait 2 or 3 quarters, but they cannot wait forever without seeing cash flow improve.
24. The real threats to the broader market are more hikes or a sustained slowdown in commercialization
Frank divides the AI debate into 3 rounds. In 2023, the question was whether it was a real trend. From late 2024 into early 2025, DeepSeek prompted a debate about whether so much compute was still needed. In the second half of 2025, demand was confirmed and attention shifted to capex returns. The question now is whether commercialization, investment returns, and future cash flow will improve; every pullback should not send the debate back to whether AI exists at all.
The first systemic risk is still energy. If high oil prices persist long enough for inflation to spread beyond a one-off shock, the Fed may have to hike not just twice but 3 times or more. “Even the strongest fundamentals will be suppressed,” and the broader market could see a moderate correction. The earlier view that mega-cap stocks can take turns supporting the market does not cover a broader tightening shock like that.
The second risk, and the one tied to the industry’s own economics, is a marked slowdown in AI commercialization. Frank does not view Anthropic’s revenue growth slowing after rising 1,000% from last year to this year as a failure; he asks, “What company can sustain 1,000% growth for many years?” The more meaningful signal would be cloud growth falling for 2 consecutive quarters. That could make the market seriously question whether demand is saturating, whether users can spend more, or whether all the available work has already been targeted.
On Q1 next year, Frank offers only a limited view: AI commercialization should still be accelerating, or at least should not be slowing, and cloud growth rates can be used to check that. The macro picture is too difficult; he hopes it will become clearer in November and December. Asked to forecast, he says he cannot see a clear framework now. He gives no explicit index target and does not turn his industry optimism into an unconditional bullish call.
25. The near-term crypto inflection still depends on fiscal liquidity returning
The host says crypto has lacked a breakout since late August. Frank returns to his earlier timeline: the rally could continue from late August into mid-to-late September, then retrace as liquidity tightened. He had expected liquidity to bottom in late October and the crypto pullback to be close to over. He thinks the move has broadly followed that view, but does not say a rebound will automatically happen just because the date has arrived.
He cites an estimate of the Treasury General Account at about $1.02T as of yesterday, versus about $980B in August, and says the buildup in the Treasury account is still draining liquidity. The next question is whether funds start flowing out in November. If so, “the crypto market could recover.” That is a conditional call: the key variable is actual money flowing back into the system, not the calendar alone or an analogy with the last market cycle.
On the larger trend, Frank calls October’s pullback “a normal retracement in a larger uptrend” and does not think the market has entered a new bear market. But he does not promise a straight line up, noting that previous cycles also had patterns of 3 days up and 2 days down. His long-term direction and the drawdowns investors have to bear in the near term can both be true.
26. Compliant fundraising and tokenization are separate regulatory themes with different timelines
On the “innovation exemption” announced on August 19, the host mentions, Frank draws a distinction: he is referring to a crypto fundraising proposal called Crypto Assets. Its significance would be to let U.S. crypto projects raise money from retail investors, rather than relying on private placements to accredited investors or excluding U.S. investors from public offerings. He compares it with existing routes such as Regulation D, crowdfunding, and Mini IPOs, and sees it as a possible path to compliant ICOs.
Legislative setbacks do not mean regulators have stopped acting. Frank recalls that when crypto market structure legislation stalled in the Senate in mid-September, he thought the SEC and CFTC could still move within their respective jurisdictions on tokenization, custody, DeFi, and derivatives oversight. He then cites the tokenization innovation exemption and permissioned AMM arrangements announced on September 17, which he believes helped drive a rally from that month into late September.
But compliant ICOs have not started. Frank says the Regulation CF proposal only opened for public comment on October 20, with a 2-month comment period. The earliest it might take effect is the first half of next year. He distinguishes positive signals about the direction of regulation from a fundraising regime that is ready to use, and does not treat a proposal as an effective rule.
Frank is more confident in tokenization as a long-term theme: traditional finance and crypto will converge, with more assets such as Treasuries and stocks moving on-chain to form financial markets. He thinks the first beneficiaries could be projects providing foundational on-chain and DeFi infrastructure. The opportunity comes from supporting new assets and trading activity, not from every token necessarily rising with them.