Feng Li: After Global Liquidity Peaks, Where Does This AI Cycle Go?

In mid-September, the world's three major central banks all turned in the same direction within the same week.

On September 18, the Bank of Japan announced a 25-basis-point rate hike, bringing interest rates to their highest level in 31 years — its second hike of the year. The day before (September 16 U.S. time), the Federal Reserve delivered its first rate increase in over three years, raising the target range to 3.75%–4.00%; a week earlier on September 10, the European Central Bank had also raised its three key rates by 25 basis points in unison, likewise its second hike of the year.

With this, the United States, Europe, and Japan have all pivoted to tightening. The master valve on global liquidity is being turned shut, one notch at a time.

Rate hikes sound distant, but their echoes are right beside us.

Over the past year, the topics flooding our social media feeds have quietly shifted from "what new features does this model have" to "how much did this stock move today" and "how many times did that company's valuation multiply." In Silicon Valley's big tech firms and startups, engineers are no longer heatedly debating technical specs — they're arguing about whether a company's valuation is inflated, and whether "the options in my hand should go to Company A, Company O, or Company X."

These subtle shifts in daily conversation are signals in themselves: a dollar-asset cycle driven by massive liquidity is reaching its tail end, and the AI industry cycle is entering its second half.

That's what this piece sets out to discuss. Drawing primarily from the "Macro Chat" segment of the Gao Neng Liang podcast and recent sharing by Feng Li, it unfolds around three questions: Where exactly does the global capital market stand in the cycle right now? Where is the AI industry positioned at this point? And after the inflection point, where will new opportunities grow from?

This is the fourth article in the "Macro Chat" column for 2026. We hope to share one lens for observation, and welcome exchange from all angles. You're also invited to search for Gao Neng Liang on Xiaoyuzhou or Apple Podcasts to listen to the full episodes.

To judge where a cycle has gotten to, let's first briefly recall where it came from.

The financial starting point of this AI boom wasn't ChatGPT in November 2022 — it was the global monetary easing of 2020.

In 2019, the ratio of total global stock market capitalization to global GDP was still around 100%, within the reasonable range of the "Buffett Indicator." To combat the pandemic in 2020, the United States launched unlimited quantitative easing; major central banks around the world printed over $10 trillion in base money within roughly a year. Amplified three- to four-fold through the banking system's money multiplier, this was equivalent to a more than 10% sudden expansion in global currency in circulation — almost unprecedented in financial history. Global stock markets and primary markets accordingly rallied broadly from the second half of 2020 into 2021.

This money became concentrated in dollar assets in 2022: the Russia-Ukraine conflict made Europe difficult to allocate to, while China faced pandemic-related uncertainties. With two of the three major economic regions accounting for nearly two-thirds of global GDP effectively off-limits, torrents of capital surged into dollar assets.

At the same time, Fed rate hikes reinforced expectations of dollar and dollar-denominated asset appreciation: in March 2022, the Federal Reserve delivered its first 25-basis-point hike, followed by another 200 basis points between May and July. Then in November, ChatGPT's emergence provided an ideal narrative for rising asset prices. From 2023 through 2025, global capital continued increasing allocations to dollar assets, inflating an ever-larger round of market cap gains.

Consider these figures: global stock market capitalization rose from $89 trillion in 2019 to $148 trillion in 2025, with U.S. stocks' share of global market cap climbing from 38% to 47%. (For a detailed account of this financial cycle's origins, see "Feng Li's 2025 Year-End Sharing: The Logic and Outlook of AI Investment".)

This capital cycle can be divided into three phases. Before the end of 2024 was Phase One: broad-based gains in total global market capitalization. Around mid-2025, it entered Phase Two: zero-sum games — inflation had become a global phenomenon, no one could print money aggressively anymore, so when the Nasdaq rose the Dow didn't, and vice versa; A-shares behaved similarly. Phase Three is the reallocation of capital.

The simple logic for judging where this dollar-dominant financial cycle stands and whether liquidity is peaking: for any major asset class to keep rising, it needs ever-increasing incremental inflows, because the higher the price, the more money it takes to push it up another notch. Once new money becomes insufficient, markets can no longer rise across the board — they shift to rotation, where "A rises while B falls, A falls while B rises."

Right now, global liquidity appears close to peaking; new money is starting to run short.

Meanwhile, those who have already made money and want to take profits are accumulating. Several signs in the past month all point the same direction: bond yields in major developed economies — the U.S., Europe, Japan, South Korea, and the U.K. — have simultaneously spiked (rising yields mean bonds are being sold); the U.S. has intervened in yen exchange rates; and American tech giants, no matter how strong their earnings reports, are struggling to push their stock prices higher.

Some numbers illustrate how much money this rally has consumed.

Total U.S. stock market capitalization rose from $34 trillion before the 2019 pandemic to roughly $75 trillion today. Given that not all shares are liquid at all times, that extra $40 trillion in market cap likely required net inflows on the order of $5–8 trillion to sustain. Bonds differ from stocks — behind every bond is real money being put up: over the same period, global debt grew from roughly $250 trillion to about $350 trillion, an increase of roughly $100 trillion; U.S. Treasuries expanded from about $23 trillion to about $40 trillion, implying $17 trillion in cash inflows.

When the total money supply stops growing, market participants simultaneously develop two mindsets: fear of heights, and chasing hot themes. So the same pool of money starts "hopping from one hole to another." Indices no longer rise together; the speed and magnitude of sector rotation both intensify. This rotation even occurs within the semiconductor industry itself: one moment CPUs are up, the next memory, the next GPUs. Intensifying rotation shows that zero-sum games have fully unfolded; and the incremental funding exhaustion we'll discuss next indicates this phase is approaching its end.

The root of this zero-sum dynamic is that the sources of incremental funding for dollar assets are drying up.

First, America's own books. Federal revenue is roughly $5 trillion. At current interest rate levels, annual net interest payments on the national debt come to about $1 trillion, plus roughly $0.9 trillion in defense spending — these two items alone consume nearly 40% of revenue. The deficit is projected to keep widening, so Treasuries can only be issued in ever-greater volumes: the Treasury Department expects to issue over $700 billion in bonds between June and September 2026.

But buyers are shrinking. Over the past two years, central banks of multiple countries including China have steadily increased gold holdings while reducing Treasuries' share of reserves. The remaining major buyer, Japan, is preoccupied with the yen having depreciated to a 40-year low. The U.S. claims it will help stabilize the yen's exchange rate, but the Treasury's exchange stabilization fund that can be deployed without Congressional approval amounts to only a few hundred billion dollars — it's essentially expectation management, with limited practical effect.

More supply, fewer buyers — the market naturally demands higher rates: the current 10-year Treasury yield has risen to around 5%, and the 30-year has breached 5.2%, approaching or reaching highs not seen since 2007.

This forms a "sensitive and fragile triangle": the U.S. needs to issue Treasuries at massive scale; the winning bid rates on those Treasuries can't be too high; and Japan cannot trigger a liquidity crisis by refusing to buy or by selling.

The hardest problem at the tail end of a cycle is: there's no money left. For the Federal Reserve, rate hikes raise debt service costs, while rate cuts expose weakness. The U.S. Treasury Secretary and Fed Chair are straining to maintain the thread-thin channel of "appropriately strong dollar," but their actual tools are few: one, hawkish rhetoric and expectation management; two, benefiting from safe-haven flows back to the dollar whenever global risk flares up.

A small loop from March to June 2026 perfectly illustrated this mechanism and its limits.

After the U.S.-Israel-Iran conflict erupted in late February, global capital first panicked into cash, then rotated into dollars and U.S. Treasuries. The dollar index rallied sharply from April to May, with the U.S. memory chip sector surging in tandem. The crisis objectively drove another round of global capital concentration into dollar assets. But this time, the pool available to drain was already shallow. Once that extraction failed, both Treasuries and equities entered wide-range volatility. Even if leverage in Japan and Korea were blown up, the capital that could be extracted was no longer sufficient to drive the next leg up in a $75 trillion stock market.

Compounding the problem, this limited capital is being pulled in multiple directions simultaneously: Treasuries need liquidity support, which is priority one; pushing the $75 trillion equity market higher requires sustained inflows; and AI data center financing needs are expanding rapidly, with private debt approaching the trillion-dollar scale.

Meanwhile, other major economies are similarly preoccupied: the European Central Bank has resumed rate hikes, and the Bank of Japan is tightening gradually with yen capital repatriating home — neither can inject large-scale liquidity into global capital markets as they once did.

At elevated valuations with tightening marginal liquidity, black swans rarely begin with fundamental deterioration at specific companies. They are more likely triggered by external liquidity and geopolitical shocks.

The first black swan is the yen.

For more than a year, Japan has been the primary provider of dollar liquidity. Japan itself is among the largest foreign holders of U.S. Treasuries. More critically, there is the "carry trade": with Japan maintaining near-zero rates and no capital controls, global speculators borrow yen cheaply, convert to dollars, and buy U.S. Treasuries and equities — a major source of dollar asset liquidity over the past year.

Most carry trades are leveraged. As Japan continues raising rates, these positions face mass unwinding; Japanese insurers and other institutions will also concentrate on selling dollar assets to repatriate capital home, capturing yen appreciation gains.

The second black swan is data center private debt.

Companies like Apple and Google, sitting on hundreds of billions in cash, were historically important buyers of U.S. Treasuries. But in Q2 this year, except for Microsoft with its steadiest cash flow, Google, Amazon, Oracle, and other internet and tech giants saw free cash flow turn negative due to excessive capex, transforming from Treasury buyers into issuers competing for liquidity themselves. Oracle issued the most: $43 billion in bonds last year, with another $20 billion expected this year, and its corporate debt trading yields have been climbing steadily.

If this private debt blows up, it could trigger a cascading selloff across the entire asset class.

The third black swan is the Strait of Hormuz.

The "partial reopening" window created by the June memorandum has closed; the strait has returned to near-shutdown status. Global energy and commodity markets today are far less resilient than two and a half months ago. If the blockade persists, commodity prices face greater upside pressure (on September 9, Brent crude futures breached $100 per barrel for the first time since July 24), potentially dragging down the entire market.

Even without directly impacting the United States, it would disrupt American supply chains globally. By industry estimates, China's strategic petroleum reserve capacity has expanded to four months; combined with diversified energy sources from new energy development (data centers increasingly use green power), it is relatively secure. The same cannot be said for Japan, Korea, and others.

There are two trackable indicators for judging whether the cycle has turned.

The first indicator is "watch only the leader": as long as the top company's stock can still hit new highs (Cisco in 2000, for example), the capital cycle is still ascending; once the leader fails to make new highs, the cycle may have entered decline.

The second indicator is the gold-dollar seesaw: when markets expect dollar strength, capital prioritizes interest-bearing, highly liquid dollars; when the dollar shows weakness while risk concerns persist, capital rotates to gold.

At the tail end of a cycle, the hardest money to earn is the last wave.

When broad asset classes enter high volatility or decline phases, it is difficult to expect any subcategory to rally against the trend, just as during the four-year real estate downturn, virtually no commercial real estate could sustain independent gains for long. But peaking does not mean immediate decline; broad assets will continue wide-range fluctuation until liquidity truly reverses one day.

As for where and when black swans emerge, it is genuinely hard to predict.

Zooming out further, this financial cycle has an even larger backdrop.

After the dollar decoupled from gold in 1971, central bank money printing no longer faced hard constraints, and central banks grew increasingly aggressive in using balance sheet expansion to intervene in crises. There is a book called The Hand of Money, describing precisely how central bankers used their "magic hand" to manufacture liquidity. Whenever tides rose and fell, they applied their tools.

The quantitative easing (QE) after the 2008 global financial crisis was implemented by the United States in three and a half rounds over three and a half years, basically printing a bit more whenever it proved insufficient — relatively cautious. 2020 was different: rate cuts to zero plus unlimited QE right from the start, completing in seven or eight months what previously took three and a half years.

The money printed in 2008 should have been gradually unwound during the rate hike and balance sheet reduction cycle that began in 2017, but the Trump administration's interference with Federal Reserve policy interrupted this contraction. Then the 2020 pandemic叠加了一轮规模更大的放水，结果是"至少一次半的泡沫"堆在了一起。后果是：除中国外全球通胀难抑；资产价格高企（75万亿美元的美股市值，对应的只是30多万亿美元的GDP）；以及贫富分化加剧（有资产者受益更多，没有金融资产的人没法受益）。

由此悬着两把达摩克利斯之剑：如果泡沫破裂，理论上要"破一次半"，届时全球的一二级市场都会受到冲击；但央行行长们已经习惯了自己那只干预之手，很可能更快出手，再堆起一个更大的泡沫，代价则是长期滞胀与难以逆转的社会分化。

AI技术能解决当前困局吗？

这显然是个过高的期待，而这种期待本身，又放大了AI叙事的泡沫。需要区分的是：AI技术有价值，不等于AI公司及其股票就有价值、或值那个价。热点的形成和估值的弹性，很大程度上受流动性影响——在流动性的上行期，一些线性增长的技术，也能获得非线性增长的估值。

撇开技术本身，从金融视角看，由技术创新驱动的金融周期通常分为两个阶段：

在一个技术周期的前半段，一般是搞金融的人竭尽全力去讨论技术；

在周期的后半段，则是搞技术的人竭尽全力在讨论金融。

具体到AI，开头提到的那些日常变化——朋友圈从讨论模型参数转向讨论股票涨跌，硅谷的工程师从讨论技术转向讨论估值与期权——就是这种阶段切换的写照。

从这个角度看，这轮AI产业周期正在进入后半段。就像爬山，有上山，就有下山。

那么，由AI技术创新引发的金融周期，不同阶段的投资标的有什么不同？

第一阶段，技术创新刚起步，大家投的是最大的技术创新节点，也就是大模型。

第二阶段，大模型阵营大致成型之后，大家投的是这项技术"最有想象力"的应用：在美国是通用Agent（或者此前常说的AGI），在中国还多一条赛道——具身智能机器人。前者替代人在数字世界的工作，后者替代人在物理世界的工作。但"最有想象力"的另一面，就是没那么容易落地。

只有当前两个阶段走完，大家不再关心谁能讲最大的故事，而开始关注"谁能用技术真正赚到钱"时，才会进入第三阶段。

我认为，AI正处在技术投资从第二阶段向第三阶段过渡的临界点。从2026年5月开始，我个人判断拐点已经临近，于是在基金内部鼓励同事们转向"谁能用上AI、并且靠它赚到钱"的方向。

支撑今天AI相关产业这些"高楼大厦"的底层，是全球AI资本开支的持续投入。高盛的数据显示，到2026年底，全球AI领域的累计投资总额将达到1.8万亿美元。这样的资本开支能否持续？答案或许要交给时间检验，但眼下已经能看到一面镜子和五个信号。

这面历史的镜子，是2000年的思科。

互联网泡沫时期，市值创下全球纪录的不是互联网商业模式公司，而是做交换机的基础设施公司思科，最高时达5500亿美元。当时的逻辑是：互联网再怎么有泡沫，总要建基础设施、总要买交换机，何况思科手里还有几百亿美元的在手订单和客户定金。

思科在2000年还推出了一项业务：向购买设备、建设基础设施的互联网公司提供设备租赁和贴息贷款。是不是和最近我们看到的一些新闻很像？泡沫破裂后，最大的打击恰恰来自那几百亿在手订单——其中大部分不再执行，客户宁可损失定金也不再扩张，思科市值一年之内跌去了80%多。这里没有影射任何一家公司的意思，只是回顾一段早年的故事。

如果说思科的例子比较特殊，我们不妨再看看全球资本开支正在转折的五个信号。

第一个信号是资本市场态度的逆转。

In Q1 2026, major global AI tech giants and mid-sized internet companies, caught up in an atmosphere where "whoever's aggressive gets the premium valuation," announced aggressive data center capex plans one after another. Q2 was the "most FOMO" quarter across the entire supply chain, with the heaviest spending: everyone believed memory prices would definitely rise, so they stockpiled inventory and poured in real money.

But since Q2 earnings began rolling out in late July, the wind seems to have shifted from "whoever spends aggressively is impressive" to "whoever's aggressive gets punished." For mid-to-large internet companies whose first-half capital expenditures exceeded free cash flow, the market has applied a discount — and even Google wasn't spared.

On July 22, Alphabet reported Q2 earnings: revenue up 24%, cloud up 82%, but capex doubled year-over-year to $44.9 billion, free cash flow turned negative for the first time (-$5.9 billion), and full-year spending guidance was raised to $195–205 billion. The next day, the stock fell 7.1%, wiping out roughly $300 billion in market cap in a single session.

The second signal is cost pressure.

In Q2, major tech companies' gross margins were relatively insulated from price hikes because the products they sold still used low-cost inventory stockpiled before Q1. By Q3, that inventory was depleted. Combined with NVIDIA raising AI server prices due to surging memory costs, cost pressure began transmitting to the end market in earnest.

From public reports, Apple raised iPhone prices twice in the Japanese market. Huawei, Xiaomi, HONOR, and other domestic Android players announced price increases on multiple products in early September. The impact of raw material inflation is beginning to show.

The third signal is infrastructure falling short of expectations.

The wave of data centers announced in the United States in early 2026 has been broadly delayed or scaled back due to stricter environmental reviews, state government approvals, grid and supporting power generation facilities, and votes by local residents. In Q1 alone, more than 75 projects (with a total value of $130 billion) were canceled or postponed. By Q2, opposition to data centers was still growing — so much so that President Donald Trump responded at the end of August. In his view, data centers are the golden goose.

Currently, of the 12GW of U.S. data center capacity announced for 2026, 60% (7GW) has been canceled or delayed, with only about 5GW actively under construction.

The fourth signal is the "lease instead of buy" strategy being deployed by giants.

NVIDIA is partnering with third-party financial firms to help customers obtain loans or guarantee off-balance-sheet lending, then leasing GPUs to users. A Financial Times commentary noted this is essentially "old-fashioned vendor financing" — NVIDIA is writing checks for customers to buy more product, or to afford product they otherwise couldn't. Cisco and Nortel used the same method in the six months before the dot-com crash.

The fifth signal, related to this, is the giants' cash flow.

Barron's published an article asking: NVIDIA's earnings are explosive, but where's the cash flow? Based on the latest report, net profit looks stellar, but free cash flow was cut in half quarter-over-quarter, receivables surged (up roughly $24.6 billion in six months) — far above historical levels — and long-term borrowings quadrupled in a single quarter.

Sentiment in secondary markets transmits to primary markets, because the wealth effect in primary markets ultimately depends on secondary markets.

This year, China's primary and secondary markets have shared something in common: it's not a broad-based rally. Valuation growth has concentrated in areas where AI intersects with Chinese industrial policy — robotics, Embodied Artificial Intelligence, world models, quantum computing, controllable nuclear fusion. Hot topics rotate quickly too, accumulating large gains in very short periods.

In the last month or two, these popular sectors have shown signs of cooling in the primary market.

First, new money is drying up.

Previously, robotics companies could be pushed to valuations in the tens of billions because enough capital was flooding in. Now, state-owned capital has tapped the brakes to varying degrees due to Document 54. The tens of billions in guidance funds from the National Development and Reform Commission and ultra-long special treasury bonds also mandate investing early, investing small, and staying below specific valuation thresholds. This has created a gap for the larger funding rounds needed in mid-to-late stages. Dollar funds were aggressive in the first half, but after catching up on deployment pace, some have begun to wait and see.

Second, the transmission of wealth effects is reversing.

Companies that went public in 2025 have seen their lock-up periods expire in the second half of 2026. People are discovering that stories worth hundreds of billions of HKD at IPO have shrunk to tens of billions or just over a hundred billion at lock-up expiry — the wealth effect fell far short of expectations.

Third, the money-making effect of Hong Kong IPO "lottery" subscriptions is cooling.

Hong Kong IPO subscriptions are usually not a high-probability way to make money. Yet in the first half of this year, the first-day破发 rate for Hong Kong new listings was only 14.3%, the lowest in three to five years, attracting large amounts of short-term speculative capital to play the IPO lottery. To use an imprecise analogy: when subscribing to new listings is almost a sure bet and you can earn several tens of percent in two weeks, who's willing to wait a year to earn 10%?

Therefore, when the破发 rate returns to "normal levels" (around 50%), this can be seen as an indicator of whether Hong Kong IPOs are returning to investment fundamentals.

Since July, this metric has begun to shift: 17 new listings in July, with a first-day破发 rate exceeding 40%; only 2 companies listed in August, slowing the pace. After an approximately two-week dry spell in late August, the only two new listings in the first week of September both broke on their first day, each falling 20% in their debut week. As of mid-September, 4 of 5 new listings broke on the first day, an 80%破发 rate. This may mark the beginning of the end for the money-making effect of Hong Kong IPO subscriptions.

To summarize, the cyclical coordinates of the AI industry in the current macro environment can be characterized as follows:

The financial cycle is in the late stage of Phase 2 (zero-sum game). Technology investment is at the inflection point between Phase 2 (the most imaginative applications) and Phase 3 (who can make money). The self-reinforcing mechanism of capex that props up tech valuations — as long as spending doesn't stop, revenue and stock prices can feed each other, and valuations can be sustained — has begun to show cracks reminiscent of 2000.

At the tail end of every bubble, someone always trots out that old line: "This time is different."

What's different between this AI cycle and the last internet cycle? The location of the burn is indeed different.

In the internet era, the vast majority of people had never used the internet. Internet companies had to subsidize users to bring them aboard. Money was burned on the front end, to educate users. Later, the market demanded these companies prove they could make money in the current period. When the burning stopped, some companies collapsed. Today, users are already universally on Mobile Internet. Many companies simply need to add AI features to their products. Money is mainly burned on the back end — that is, on large model-related technology.

But different burn locations don't necessarily mean different outcomes. If one day the market demands that large model companies prove their current-period profitability, they too will have to cut costs and boost efficiency: reduce capex, find ways to get paid for every token.

This is already beginning to happen. Whether domestic or international, large models have already started charging, or are considering charging.

Once large model companies begin charging for tokens, an important difference between large models and the internet will become prominent: the internet runs almost entirely on static databases — process once, reuse many times, with very low marginal cost per service. Large models, by contrast, must genuinely consume compute for every incoming question, calculating the answer in real time, then dynamically calculating again when outputting the response. The cost structure gap between the two is substantial.

So if the market truly demands that foundation model companies make money, what the outcome will look like is a question worth following closely.

Identifying that a cycle is in its late stage does not equal being bearish on the technology. A bubble bursting is merely a financial cycle concept; having a bubble doesn't mean the technology lacks value.

From the steam engine becoming a symbol of the Industrial Revolution to its widespread use, several decades passed. The internet experienced several bubble bursts, yet that hasn't prevented it from being ubiquitous at the application layer today. Saying AI technology has a bubble is not to deny its profound impact on business and society.

Where to look after the inflection point? In an upward capital cycle, people invest in narratives, whoever has the bigger story. Past the inflection point into a downward cycle, the technology narrative matters less — people invest in who can make money.

Because primary markets can't trade continuously, we need to make a judgment first: if you believe we're still in an upward capital cycle, you can keep investing in the big stories. If you believe the inflection point is near, you should invest more in those who can make money.

So the question becomes: what can make money?

Chinese AI applications are one answer. Previously, China typically lagged in the first half of technology cycles, only taking the lead in the application phase. This AI round is different: China is already not far behind in the technology phase, and has even more strengths in the application phase.

China's advantages lie in its well-developed digital infrastructure, complete industrial chain, numerous industrial segments and scenarios, plus industrial policy drivers. Since the second half of 2025, industrial policy has already begun steering toward AI+ applications.

In August 2025, the State Council issued Opinions on Deepening Implementation of the "Artificial Intelligence+" Initiative. On April 28, 2026, a Politburo meeting explicitly required "comprehensive implementation of the 'Artificial Intelligence+' initiative." In the July 30 Politburo meeting's technology discussion, "Artificial Intelligence+" was placed in a very important position, with emphasis on developing new forms of intelligent economy and improving AI governance systems.

The AI+ story may not be as grand as robotics, nor is it the current investment hotspot. But when specific industries add AI, the stories of improving quality and efficiency, cutting costs and saving money, are equally worth investing in.

Regarding the path of AI technology spillover, I have a preliminary judgment: today's heat is concentrated in data center leasing and services, semiconductor equipment, and chips — areas directly related to AI. But the world is ultimately fair, and economies are ultimately diverse. No large economy can be driven by a single axis alone, so opportunities will emerge in all directions. I believe market heat will spread to broader domains — AI+ consumer, AI+ services, AI+ industrial and manufacturing — diffusing toward directions that can both use AI and make money.

Another angle on the answer: invest in assets that are large, cheap, and currently perhaps unloved.

Whether in primary or secondary markets, one lesson verified over many years is: buy when unloved, sell when crowded. So look at "where it's large, where it's cheap, where it's unloved." We can afford to be optimistic: large and cheap means it can't stay unloved forever.

I'll give two examples. One is the intersection of biomedicine and AI.

FreeS Fund has invested in multiple rounds of the biopharma company Inming Biotech, which spent six-plus years building over a dozen First-in-Class pipelines with entirely novel targets, five of which have entered Phase I/II clinical trials. This company worked for six years, yet its valuation is roughly the same as a world model company we invested in just six months ago this year, whose valuation has already risen more than tenfold.

It sounds somewhat unfair, but the world will eventually be fair. So starting in May, I've been internally encouraging colleagues to invest more in biomedicine, especially at the intersection of biomedicine and AI.

In August, the AI and biomedicine industries received an encouraging piece of major news: Moderna used AI screening to develop the world's first personalized melanoma cancer vaccine, and its stock nearly tripled in a day. It employed a technology whose story isn't grand enough, but that is genuinely AI-related, to complete the core screening work — AI was genuinely useful here.

From an industry perspective, Chinese pipeline-focused pharma companies are actively pursuing license-out deals, which explains why China's biomedical license-out value has surged this year — already accounting for two-thirds of the global total. This trend will pressure or incentivize more drugmakers to accelerate, broaden, and improve their pipeline R&D: faster, more extensive, and higher quality. To achieve this, they'll exhaust every available efficiency tool, and AI is naturally among them.

Another example is SaaS.

SaaS has never been a particularly favored sector, and since large language models took off, there's been growing chatter that AI will make SaaS obsolete. I see it differently: SaaS companies possess their own data assets, customer bases, and workflow integrations. Many will inevitably evolve toward AI, becoming a new generation of software service companies.

To summarize: financially, this AI boom represents the massive liquidity of 2020 that was forced into dollar-denominated assets in 2022, with ChatGPT serving as the convenient rationale. By Q3 2026, incremental capital will be nearly exhausted, global liquidity will tighten, and dollar-dominated capital markets will enter the late stage of zero-sum competition. Tech giants' AI capex is showing cracks reminiscent of Cisco in 2000, and hot sectors in the primary market are already cooling.

Regardless of whether this time is truly different, only when the large-model-driven financial cycle enters its second half will the market begin focusing on who can actually make money with AI, rather than who can tell the grandest AI story. That's when real opportunities in AI applications will arrive: widespread, substantial, and sustained.

(The views expressed above are personal opinions for reference only and do not constitute any form of investment or financial advice.)