Vol.236 Macro Talk 111 | China-U.S. Summit Talks, AI's Limits, and the Global “Reservoir” Logic of Capital
Summary
- The core outcome of the China-U.S. summit talks was not a package of detailed agreements, but a top-level calibration of the framework for sensitive issues. 李丰 says the only clear quantitative result was a further 2-month extension, to Jan. 10, 2027, of the suspension arrangement for China-U.S. trade measures that had already been extended to November 2026—conveniently covering the shopping season from Black Friday through New Year’s Day. The rest was closer to confirming “what you broadly want and what I broadly want,” with implementation left to subsequent trade and AI dialogue mechanisms.
- The AI narrative is shifting from continually raising the capability ceiling to defining capability limits, safety, and controllability. 李丰 compares the shift to autonomous driving in 2014-2017: during the bubble’s upswing, the focus was on technology, competing approaches, and demos; only after deployment, accidents, and a downturn in capital did the focus turn to rules. The digital world can tolerate “six fingers,” but the physical world “can’t let my hand pass straight through the table to the underside,” making physical and mathematical constraints and white-box capabilities increasingly necessary as AI moves into robotics, industrial controls, and autonomous driving.
- Three to 4 years after ChatGPT’s launch, there is still no second mass-market AI application, suggesting the real constraint may be cost rather than model capability. Every conversation, inference, and long context window adds cost, while the value of ordinary search, summarization, and document generation is insufficient to support expensive advertising. 李丰 therefore expects cloud models to first dominate high-value use cases such as coding and gaming, with everyday needs eventually handled by edge models that solve “99%, 89%, or 79%” of tasks.
- Today’s large language models may not evolve along the same extension path into a unified foundation for all AI. 李丰 argues that world models and physical models need to “cross at least half the road,” incorporating different architectures, physics, mathematics, and interpretable computation. Since the second quarter, the embodied-AI investment consensus has also begun shifting from VLA, VLM, and VLTA approaches that “pass through L” toward VM or no-L routes that reduce the language intermediary, because language is inherently a lower-dimensional representation and necessarily loses information.
- The fact that large-model companies are attractive investments in the short term does not mean they will deliver ideal returns over a 5-year horizon. 李丰 says that, in the short run and ignoring other constraints, the obvious answer is to invest; after missing the first listed names, financial investors have rushed into Kimi, effectively settling the short-term disagreement through their actions. But if foundation models eventually become infrastructure like the cloud did in 2015, competition will shift toward vertical integration across AIDC, chips, models, MaaS, SaaS, and end-user services—a low-margin, oligopolistic business won on scale and cost. 李翔 also notes that the valuations of 2 unlisted large-model companies already exceed the market capitalization of 智谱, the largest among listed companies.
- China’s capital migration has begun, but so far it is mainly a shift from deposits into bonds and fixed income, not a broad move into equities, funds, or property. Deposits rose sharply from 2024 through the first half of 2025, while July and August showed smaller increases alongside rising non-bank deposits. More importantly, central-bank data show direct and indirect financing at roughly a 5:5 ratio in the first 8 months, faster than 李丰’s original expectation of moving from 5:3 toward 5:5 over time.
- Whether the latest U.S. rate-hike cycle can drain global liquidity on a sustained basis depends on whether it is a short-term adjustment or creates continuous expectations lasting more than 1 year. The world has not gone through another classic broad-based liquidity wave, China has capital controls, and Japan is being forced to manage pressure on the yen around 155-158, so 李丰 leans toward a short cycle this time. Sustained hikes could instead draw in global capital through a stronger dollar, sending U.S. stocks lower first and higher later, while sustained cuts could trigger outflows from dollar assets and produce a market that “rises first, then declines persistently.”
- The most dangerous intersection is the U.S. trying to protect Treasuries, equities, and the dollar at the same time while AI expansion remains heavily dependent on debt issuance. Delays at Oracle data centers and related bonds with less than $20B in total value falling below 90 cents on the dollar have already linked construction, power, hardware, financing costs, and delayed revenue. Once financing costs for mid-sized and large technology companies rise to 8% or even the low-to-mid teens, infrastructure projects may need annual returns above 15% to make economic sense. Asked whether this means “a financial crisis is about to happen,” 李丰 answered: “It’s possible,” while stressing that the real question remains the timing of the trigger.
Deep dive
1. The summit set the framework first; specific deals were left to follow-on mechanisms
李丰 infers from the timing of the announcements that the itinerary may have been broadly settled long ago, but Beijing did not formally confirm it until shortly before departure. That suggests the preceding trade negotiations and AI dialogue were mainly setting the framework for “what would be discussed and in what direction.”
On 李翔’s observation that the delegations were asymmetric—the Chinese side brought no business executives, while the U.S. side included major figures from chips, AI, and Elon Musk—李丰’s explanation is that the meeting was not primarily about brokering deals between companies. The focus was an exchange of top-level views on key issues.
He had expected the two sides to discuss AI chips such as H100 and H200. The U.S. might have wanted to cap export performance around Huawei Ascend levels, expanding Nvidia’s commercial footprint while constraining China’s GPU development; but the issue may have been “too specific” and did not appear in the public readout.
2. Chinese exporters were the real beneficiaries of the 2-month trade extension
The only clear quantitative outcome was a further 2-month extension, to Jan. 10, 2027, of the arrangement agreed in Kuala Lumpur to suspend additional measures targeting various forms of competition, following an earlier extension through November 2026.
The seemingly limited extension fully covers the U.S. shopping season from Black Friday and Christmas through New Year’s Day. 李丰’s investor-style read is straightforward: exporters can complete their peak-season orders and shipments, which is “definitely a good thing for foreign-trade companies.”
Longer-term details may still be implemented through China-U.S. trade talks, AI consultations, and other mechanisms. With the composition of Congress still subject to change ahead of the U.S. midterm elections, neither side may view it as reliable to lock in too many specific commitments now.
3. The most sensitive exchanges likely took place in meetings of 4 to 6 people
In the public schedule, the most important events to watch were not the large banquets but the small meetings and tea sessions attended by only about 4 or 6 people. 李丰 suspects these were where issues unsuitable for a public statement—regional conflicts and China’s surrounding security environment, for example—were discussed.
One supporting clue came from Japan’s prime minister making an urgent visit to the U.S. 1 or 2 days before the Chinese delegation arrived and seeking a brief meeting with Trump. The public statement ultimately contained nothing corresponding to that request, leading 李丰 to conclude that sensitive matters discussed by the sides cannot be inferred from the communiqué alone.
The meeting also gave Trump a chance to showcase diplomatic capability ahead of the midterm elections. 李丰 sees it, together with the agreement reached with Denmark on Greenland during the same period, as serving the political narrative of “U.S. influence.”
4. Anthropic’s distillation allegations mix competition, security, and growth anxiety
李翔 noted that before the visit to the U.S., Anthropic accused Chinese large-model companies of distilling, and even “wrapping,” its models; Trump also responded in a post. He therefore asked whether the episode was a government pressure tactic or normal competition in the AI industry.
李丰 did not reduce it to a matter of state rivalry. He finds a commercial explanation more plausible: with growth under pressure or slowing at some companies, the issues of model distillation and control over AI development are being amplified again, and “at least part of the reason” may lie there.
李翔’s reservation is worth preserving. After former Anthropic employees made disclosures, mainstream media followed quickly, while another group of investors and entrepreneurs argued that the episode itself contained a large element of, in quotation marks, marketing strategy. Jensen Huang and others also reject slowing model progress in the name of safety. The shared conclusion is not that the noise is unreal, but that noise, genuine risk, and commercial motives are all present at once.
5. AI’s next phase will be a competition over boundaries, not just capabilities
李丰 strongly agreed with the Chinese leader’s remarks before the banquet that AI should remain “under human control” and contribute to human progress. He sees this as a signal that China’s AI policy may be concerned with both expanding capabilities and defining safety boundaries.
Over the past year, the industry has continuously generated terms such as Agent, “lobster,” Harness, RSI, GEV, world models, and physical models. At root, they are all exploring the capability ceiling from different directions. 李丰 cautions that “in addition to exploring the capability ceiling and extending the capability frontier, it now also needs to define and explore the capability boundary.”
His cycle view is that markets discuss “who is stronger” during an upswing and bubble phase; only after applications become more widespread, accidents multiply, and capital turns down does the market start asking about safety and rules. The evolution of autonomous driving after 2014-2017 is the closest precedent.
6. Once black-box models enter the physical world, white-box constraints become essential
Today’s large models are primarily data-driven and trained end to end: feed in data and solve the capability problem directly. But black boxes are difficult to explain, and tuning them creates a “seesaw effect,” where fixing one capability can trigger another problem.
李丰 therefore proposes that “the more physical the application, the more white-box it should be”: concrete applications need more physical and mathematical laws built in so that at least part of the behavior can be understood, explained, and tuned. The emergence of world and physical models reflects the need for robots to be constrained by real-world laws.
In the digital world, visual artifacts, six fingers, or other image errors can be tolerated; the ability to write code has already made attacks more dangerous. In autonomous driving, robotics, and industrial controls, errors can be a matter of life and death. His analogy is: “You can’t let my hand pass straight through the table to the underside.”
Once accidents involving Tesla FSD and some Chinese autonomous-driving brands were exposed or amplified, related constraints and rules for L2 and L3 systems followed. The sequence shows that corner cases are not peripheral noise; they are core evidence for defining capability boundaries and driving regulation.
7. The lack of a mass-market application after ChatGPT may start with the cost structure
李翔 frames the contradiction clearly: from ChatGPT’s launch in late 2022 through 2026, models have changed rapidly and new concepts have proliferated, but the applications ordinary users can feel directly remain concentrated in chat, office work, and making PPTs. No second product has broken through beyond the early-adopter class.
李丰 says the most fundamental difference between AI and the internet is that the internet’s marginal cost is close to zero. When a search engine scales from 10,000 users to 100M, the cost of each additional user is low apart from incremental server concurrency; AI still has to understand, compute, and generate a response for every user.
Long conversations also require the context to be loaded repeatedly. “If you’re a chatterbox,” the cost rises further. Meta’s Muse personal assistant briefly went viral before suffering outages and an inability to sustain service, exposing the same scaling problem.
8. Cloud models will take high-value use cases first; everyday demand may return to the edge
For search, ordinary summaries, or low-value documents, the incremental value generated by a model may not be high enough. If advertising has to cover inference resources, the advertising value generated by an individual user may need to be an order of magnitude higher than in internet search, or even 10x to 20x higher, making the business model unnatural.
李丰 therefore expects the first applications to scale more easily in coding, some games, and other use cases with high willingness to pay or medium-to-high value. Charging these users more is not a subsidy for low-value demand; it is because they generate greater economic value.
Once cloud models generally charge for usage, many everyday needs may migrate to small models on phones, computers, and other edge devices, initially solving “99%, 89%, or 79%” of users’ tasks. This avoids recurring cloud costs and addresses some privacy concerns, though whether edge data is uploaded will depend on the implementation.
9. Large language models may not be the common ancestor of world and physical models
The larger technical question is whether today’s large language foundation models can naturally evolve into world and physical models by adding modules, or whether different architectures and disciplinary approaches must be explored. 李丰 clearly favors a middle position rather than a single path.
His judgment is that the next phase must “cross at least half the road”: at least half of the problems will need to be solved through new technical routes, architectures, physics, and mathematics. It cannot be assumed that today’s LLMs are the foundation models for all future AI.
This does not make language capability useless. Language massively improves knowledge transmission, abstract understanding, and coordination among people. But animals lack human language and can still evade predators, move, grasp objects, and engage in limited group coordination. Physical decision-making and language understanding “are related, but they do not necessarily follow exactly the same development path.”
10. Embodied AI is shifting from “passing through L” to reducing the language intermediary
李丰 uses VLA, VLM, and VLTA to describe early embodied-AI approaches: whenever an architecture contains L, visual information still has to pass through language understanding or language labeling before entering planning and execution.
In the first quarter, the industry was still clearly biased toward L for the robot “brain.” By the second quarter, investors had begun to develop some consensus around reducing L and moving directly toward VM, or adding more physical and mathematical computation so the system becomes “more tunable, or a little more white-box.”
World models have become an important term, but the definition itself has changed several times and remains unclear. 李翔 adds a key constraint: language is inherently a form of dimensionality reduction, so it cannot fully express the information in the mind and will always leave gaps and distortions.
11. The 2023 large-model investment split has already been settled by price in the short term
Looking back to late 2022 and 2023, venture firms were clearly divided into 2 camps: firms such as GSR Ventures were unwilling to build large positions in large-language-model companies, while some younger VCs were making aggressive bets. 李丰’s answer depends on the time horizon: “If you’re looking at the short term, there’s no question you should invest.”
The reason is concrete. The first 2 large-model companies to list mainly allowed state-owned capital to capture the biggest gains, leaving many financial investors regretting that they had missed out. Kimi, which had not yet listed, then attracted financial investors from across the market in a “closing the barn door after the horse has bolted” pursuit. Alibaba is more like a strategic investor and cannot be treated as equivalent to pure financial capital.
The short-term disagreement has largely disappeared, but the medium-term conclusion remains unresolved. 李翔 then points out that the private-market valuations of 2 unlisted large-model companies already exceed the market capitalization of 智谱, the largest listed company by market value. Even if latecomers have the right direction, their entry price may already have priced away the return.
12. Once foundation models move to the cloud, winners will be determined by full-stack cost, not a single model
Google, Amazon, and Microsoft, along with Alibaba, ByteDance, and Tencent in China, are not merely building compute centers. They are also developing foundation models, applications, or Copilot-type products internally. 李丰 believes they have probably exceeded market expectations in capability acquisition, utilization, and cloud growth.
李丰’s medium-term framework is that foundation models may become “cloudified,” as the cloud did around 2015. Model companies may not become cloud providers, but as foundation-model iteration slows, cloud companies can connect front-end users, back-end data, and compute-center capabilities, filling in the model layer.
A decade ago, vertical IaaS was briefly a hot area, but the market ultimately showed that cloud requires horizontal scale and vertical full-stack capabilities. The stack runs from IaaS to MaaS and SaaS, and now extends downward into the chip hardware inside AIDC. The more complete the chain, the more effectively it can spread the cost of serving each user.
For investors, foundation models becoming infrastructure means both a huge market and low margins, oligopolization, and efficiency-driven competition. “Holding for 2 years” may still make money; over 5 years, vertical model companies may end up like the vertical cloud companies of the past, far below initial expectations.
13. The “reservoir” framework puts money supply, risk appetite, and asset prices on one map
李丰 divides the global financial system into several pools. Household deposits are the calmest and may also become a “barrier lake”; bonds and fixed income are the second layer; property, equities, and other risk assets are the more turbulent third layer. Digital currencies are a smaller, more aggressive pool, while gold forms a separate reservoir.
The first variable is whether it is “raining” from above. When central banks print money together, the amount of water in the entire system increases; the larger the scale, the more it resembles a downpour. In the age of fiat currency, repeated and increasingly forceful use of liquidity injections and quantitative easing has not only amplified financial cycles but also driven a sharp rise in global money and debt.
When the real economy grows more slowly than the money supply, new funds enter financial assets and push their prices higher in a self-reinforcing cycle. People who already own assets and capital therefore earn more. 李丰 sees widening inequality, social tensions, and even political conflict as the accumulated result of this mechanism.
14. Chinese deposits have started to migrate, but risk appetite has not truly opened up
From 2024 through the first half of 2025, Chinese deposits rose significantly, suggesting that funds were moving from more volatile pools back into the calmest one. In July and August, household deposits still increased but by less year over year, while deposits at non-bank financial institutions grew, indicating that the migration had begun.
Because of differences in time-deposit rates and concentrated maturities, the largest shift so far has been into bonds, fixed income, and related wealth-management products, not into equities or other risk assets. 李丰 says the next destination—funds, the stock market, or even property—depends on whether risk appetite can recover.
A more structural change is occurring on the financing side: China is shifting from indirect to direct financing, which also includes special-purpose bonds and local-government financing vehicle bonds. 李丰 originally expected the ratio to move gradually from roughly 5:3 toward 5:4 and 5:5, but central-bank data show it had already reached roughly 5:5 in the first 8 months of this year.
15. The dollar’s unique rate-hike power is its ability to drain 3 pools in other countries at once
As the reserve currency and the primary currency for trade, the dollar’s liquidity and reach make the U.S. almost the only economy able to absorb capital from other countries on a large scale during a rate-hike cycle. Higher rates raise the risk-free return on dollars and dollar bonds, while also usually signaling that the U.S. economy is running hot.
There are 2 accelerants to the drain: regional conflicts and financial crises raise demand for safe havens, while continued Fed hikes provide a yield differential. 李丰 places the Asian financial crisis, the euro’s decline during the Kosovo War, the European debt crisis, and pressure on emerging-market currencies within this framework.
Funds flow “in a steady stream” toward the U.S. only if the market expects rate hikes to continue for 1 year to 1.5 years. If the move is merely a 2-month fluctuation triggered by a change in Fed chair, a regional conflict, or oil-price concerns, it produces a brief migration but does not naturally become a one-way trend.
Whether capital can be drained also depends on how much water is in other countries’ pools and how high their dams are. The effect is strongest when the world has just gone through broad-based easing; if countries have already been drained and strengthened capital controls, even a wider U.S. rate differential may not reproduce the scale of the previous cycle.
16. The amount of global liquidity available to drain is limited; Japan and China are the 2 key exceptions
The world has not gone through a new, classic easing cycle, so 李丰 believes there is not much incremental capital for the U.S. to drain. The major economies that can still maintain clearly low rates are mainly China and Japan, but both have their own “dams.”
Japan hiked by 25 basis points the day after the U.S. hike, yet the yen remained weak around 155-158. If the market does not believe Japan will continue raising rates, the currency may remain under pressure. At the same time, the Bank of Japan holds large amounts of bonds and risk assets; if Japan hikes, bonds issued by the finance ministry may need to be absorbed by the central bank. Either side of that bubble would be difficult to withstand if it burst.
China has enormous savings, but capital controls, offshore and onshore management, and rules governing outbound flows make it harder to drain. 李丰 therefore sees the latest hiking cycle as creating short-term pressure on Hong Kong stocks and emerging markets, but leans toward a short cycle of 1 or 2 hikes—or perhaps 3—rather than a new draining wave lasting more than 1 year.
17. U.S. consumption, housing, and employment have placed a ceiling on sustained rate hikes
U.S. blue chips are also facing the predicament of “old-guard stocks.” Nike was removed from the S&P 100 and its share price fell sharply; McDonald’s dropped about 30%, grew only 0.8% in the second quarter, and is expected to decline year over year in the third quarter. These stocks have been abandoned by capital, but their fundamentals are also genuinely weak.
Taco Bell, Burger King, and other competitors can still grow, in part because they offer meal deals below $10. The issue is therefore not a synchronized decline across all restaurants, but inflation and rising costs squeezing consumers’ ability to pay for necessities.
The pillars of job growth are narrowing toward government, healthcare, and other defensive sectors, while construction has dropped out of the lead. Before the rate hikes, U.S. mortgage rates were already above 7% annually; if home-price appreciation and rental yields cannot reach 8%-9%, the economics of buying a home become difficult to justify.
Remaining U.S. resilience comes more from technology investment, gains in household asset values on paper, and services growth driven by finance. A capital market worth roughly $75T—about 2.5x the roughly $30T GDP—also explains why “protecting U.S. equities” has come to resemble protecting the economy.
18. The Treasury-equity-dollar triangle is pushing AI debt to the risk frontier
U.S. policymakers face 3 simultaneous tasks: issue Treasuries at affordable rates, maintain the technology-stock boom to attract global capital, and prevent lasting damage to confidence in the dollar and government debt. Any imbalance at one end increases pressure on the other 2.
AI makes the balance harder. Many data centers are being built with debt, and corporate bonds compete with Treasuries for the same pool of capital. Force majeure at an Oracle project prevented delivery on schedule, while related debt totaling less than $20B fell below 90 cents on the dollar in the secondary market—an adverse signal.
Compute centers are simultaneously facing higher hardware, power, and storage costs; rising financing expenses; and delays to delivery and revenue. If a mid-sized or large company such as Oracle faces financing costs of 8% to the low-to-mid teens, a project may need annual returns of at least 15% to make sense, while an infrastructure business is unlikely to sustain 15%-20% returns over the long term.
19. Rate hikes do not necessarily depress U.S. stocks, and cuts do not necessarily create a lasting bull market
李丰 revises the standard intuition: rate hikes do shift some U.S. domestic capital from equities into bonds, but if a persistently strong dollar attracts more capital globally, external inflows can offset the domestic outflow and U.S. stocks may fall first and rise later. He uses the period from 2022 through the end of 2024 as an example.
Conversely, if the U.S. risk-asset pool is already clogged with global capital, sustained rate cuts can signal both a weakening economy and weaker dollar-denominated assets. Capital flowing overseas from U.S. stocks may then exceed the amount moving from deposits and bonds into equities inside the U.S., producing a market that “rises first, then declines persistently.”
This counterintuitive conclusion has 2 strict preconditions: the market must develop expectations for hikes or cuts lasting more than 1 year, and the global and U.S. pools must already be at sufficiently extreme levels. A single short-term adjustment or 2 creates volatility, not an automatic long-term direction.
20. Oracle’s crack resembles the internet cycle, but the crisis clock has not given an answer
After hearing the full negative feedback loop, 李翔 asked directly: “Isn’t this just a financial crisis about to happen?” 李丰 answered, “It’s possible,” then returned the focus to the timing of the trigger rather than treating the similarity as a definite forecast.
The internet cycle went through 2 rate-hike episodes. The longer 1994-1996 cycle drained Asian capital and was followed by the Asian financial crisis; the shorter 1999-2000 cycle eventually burst the internet bubble, while the Kosovo War and the euro’s decline redirected European capital into dollar assets and supported the final bout of speculation.
The aggressive hikes of 2022 drained liquidity most severely because the world had just gone through historic easing during the pandemic. By 2025, much of that liquidity had already been drained. The difference this time is that Treasury yields cannot be pushed down, technology companies also depend on high-cost debt, and the world lacks a large economy that has clearly collapsed and from which capital can freely flee on a massive scale.
The final variable left by the program is the renminbi. While the dollar strengthened and most emerging-market currencies depreciated, the renminbi remained relatively stable. 李丰 believes capital and foreign-exchange controls explain only part of it; the larger reasons may lie in trade structure, settlement methods, and global capital circulation, to be explored next time.