Summary

This round of “Chinese asset re-rating” has already delivered many leading Chinese stocks’ full-year return targets of 30%-40% in just over a month, with some up roughly 60% or more. Anyone chasing the rally should first guard against becoming the bagholder. 庄明浩 observes that Chinese stocks often get a window in February and March; once “East rises, West falls” becomes a consensus view across the internet, fund managers may decide, “We can call it a day.” He does not know whether the pullback will last days or weeks, and says he personally has no stock account.
DeepSeek’s breakout was driven by more than open source and low cost: the Lunar New Year mood, product experience and distribution timing combined to turn it into a “national-fortune” event. R1 launched on January 20, U.S. mainstream media picked it up on January 24, and 冯骥 opened a post with “national fortune” just before New Year’s Eve on January 26. Many users were also seeing a visible “thinking process” for the first time while being able to turn on deep thinking and web search together. “For users, this experience was entirely new. It was magic.”
The window for actionable AI entrepreneurship is reopening: the technology is still advancing, but its capability frontier has become more defined than it was over the past 2 years. From late 2022 through the second half of 2024, a team could spend 3 months building a product only for the underlying model to swallow the original plan. Now founders can work through “technology—use case—product—delivery—iteration” again. Robot vacuums with robotic arms, companion toys and AI glasses all suggest that the real moat is often scenario definition, hardware cycles and delivery in messy real-world environments—not simply plugging into a model.
Integrating DeepSeek does not mean the model war is over; the big platforms are using it to “borrow the fake to cultivate the real.” After WeChat added DeepSeek search, some users were routed to Yuanbao, which nevertheless defaults to Hunyuan. Tongyi Qianwen also launched its own reasoning model, while as of the recording Kimi, Doubao and Alibaba’s Tongyi had yet to integrate DeepSeek into their main apps. 庄明浩 likens reasoning models to Level 2 autonomous driving: “If you cannot get here, do not think about what comes next.” Once past that threshold, however, each company will retain proprietary models for its business, data and security needs.
DeepSeek has split the AI capex narrative into competing paths. The guests cited 2025 spending plans of $80B for Microsoft, $75B for Google, $60B for Meta and $100B for Amazon, plus the $100B proposed by Trump—roughly $420B in total versus about $220B the prior year. But Microsoft canceled 2 data centers and Satya Nadella questioned the actual economic growth, prompting the market to trade a “downward spiral.” Palantir’s 10% one-day drop was a sign of high-consensus assets losing support.
For now, AI is more likely to cut costs and raise productivity in existing businesses than to create broad-based new revenue and jobs; getting there may require a restructuring of organizations, business models and the way people connect. 沈帅波 worries that if the gains accrue mainly to chip sellers, data-center builders and a handful of model companies, AI could further shrink middle-class employment and purchasing power. 庄明浩 invokes the electricity revolution and a “compute—connectivity” framework to argue that the productivity dividend may take much longer to arrive.
The primary market still offers episodic profits, but compressed valuations have reduced the margin for error, and exits are harder than sourcing deals. Chinese teams have already built plenty of wrapper products to around $10M ARR; “if there is a shell, fill it” is not shameful. But going from obscurity to a multibillion-dollar valuation in 6 months leaves later-round investors, buyers of old Unitree shares and MiniMax investors exposed to the risk that 5 years of expectations have been priced in at once. RMB funds favor humanoid robots with physical delivery, while dollar funds lean toward applications; whatever the investment, the constraints around IPOs, M&A and buybacks cannot be ignored.

Deep dive

1. The hottest phase of the Chinese-stock re-rating may also be the end of the year’s trade

庄明浩’s framework, based on years of observation, is seasonal: Chinese stocks in Hong Kong and the U.S. often offer an opportunity early in the year, concentrated in February and March, after which “the year’s rally may already be over.” This is not a firm forecast of how long the current move will last; it is a reminder to first identify where you are in the cycle.
More concerning is how quickly consensus is spreading. When financial media are saturated with talk of “East rising, West falling” and a “re-rating of Chinese assets,” and Morgan Stanley’s chief China equity strategist is voicing the same view in U.S. media, 庄明浩 treats the heat itself as a signal—not a reason to chase the narrative further.
Hedge-fund incentives offer a more direct sell-side logic: if the S&P 500 returns roughly 15%-20% for the year, delivering 20%-30% is already good, and 30%-40% in a bull market is more than enough. Many leading Chinese stocks have already delivered those numbers in just over a month; some are up roughly 60% or more. The rational response may be, “We can call it a day; take the next 10 months off.”
The evening’s selloff was not a single-factor event: U.S. technology stocks fell broadly, Trump’s new investment memorandum added restrictions on Chinese stocks and Chinese investment in the U.S., and the market had simply “gone up too much and had to correct.” 庄明浩 made no call on whether the adjustment would last days or weeks. He stressed that he does not trade stocks and has no stock account, so he has no position to defend.

2. DeepSeek became a Lunar New Year event before being formalized as a technology narrative

庄明浩 reconstructs the distribution chain in detail: DeepSeek released R1 on January 20, and 梁文锋 attended a symposium with the premier that evening; U.S. mainstream media began paying attention on January 24, followed by domestic media and public accounts discussing V3 and R1; on January 26, 2 days before New Year’s Eve, 冯骥 opened a Weibo post with “national fortune,” pushing the technology story into mass sentiment.
The Lunar New Year was not incidental. China’s mobile internet industry has treated the holiday as a strategic battleground for more than a decade, combining deliberate campaigns such as WeChat’s red packets with products that go viral by accident. With many people already on holiday or traveling home, the “national fortune” label fused with the holiday mood and sent the story beyond DeepSeek’s ability to control.
Open source, low cost and model capability were the foundation, but the product form mattered just as much. OpenAI o1 had already demonstrated reasoning, but it required payment and had access barriers; most people had never used it. DeepSeek let many users see a “thinking process inside double quotation marks” for the first time—another magic trick after ChatGPT’s question-and-answer magic.
The key product move was allowing deep thinking and web search to be enabled simultaneously: it addressed the model’s stale-data problem while bringing reasoning into the oldest and most universal internet use case, search. 庄明浩’s summary: “For users, this experience was entirely new. It was magic; it was something they had never seen before.”

3. AI startups regain a full product cycle as the technology frontier begins to converge

From late 2022 through the second half of 2024, large models were in a period of technological explosion. A traditional social company might launch a project based on the capabilities available at the time and spend 3 months finding an AI use case, only to discover at delivery that the underlying technology had “flown somewhere nobody knew”—swallowing the product work. Relaunching for another 3 months could repeat the cycle.
庄明浩 says the change today is not that technology has stopped iterating, but that the capability frontier is “converging more clearly.” Founders can finally answer, in sequence: what the technology can do, which user needs they understand, which scenario to choose, how to shape the product, how to deliver it, and how to keep updating it—rather than endlessly chasing a moving model snapshot.
The underlying needs of users, especially individual users, have not disappeared because of AI. The opportunity is therefore not abstractly to “do AI,” but to extend familiar use cases. “From technology to scenario to product to continuous updating and iteration” sounds close to the classic internet playbook again—and means execution and demand judgment matter once more.

4. Robot vacuums with robotic arms show the distance between model capability and real-world delivery

At this year’s CES, several leading robot-vacuum makers independently added robotic arms to new products. 庄明浩 says that if he had to name one reason, it would be that vision models had become capable enough. Vacuuming and mopping had previously remained 2D, single-task activities; a robotic arm pushes the product into 3D space and multi-task decision-making.
A genuinely useful arm must recognize, remember, grasp and assess weight. It may need to put socks in the washing machine, return shoes to the shoe cabinet and throw trash into the bin. Only then do product tradeoffs begin: whether to wipe walls, pick up trash or sort and organize it; whether the object is metal or plastic, folded or exposed; how much weight to grip; and what price and market to target.
CES is only the starting point. Manufacturers gave initial test units to technology creators such as 韩路 so they could be placed in complex home environments, iterated and tested for accidents—the same logic as autonomous driving needing to accumulate enough miles. Between “the technology is good enough” and an end-to-end product that can be delivered at scale lies a large amount of unglamorous work.
The example also answers where founders should compete. The number of bottom-layer model players is limited, but the middle ground where models become reliable hardware, software or services remains wide open. A real Agent is not just a chat interface; it may be a robot that must bear the cost of mistakes in the physical world.

5. Toys and glasses that look “stupid” often lose to hardware cycles, not to bad product judgment

An AI companion toy takes scenario selection, partners, tooling, pricing, marketing and production to move from project launch to market. If a team started in early 2024 based on the models available then, the finished product might not appear until the second half or year-end. Consumers would then see either an expensive product or one with “extremely stupid” capabilities. Some Japanese companion products sell for several thousand yuan or more, such as LOVOT, which is even more expensive.
AI glasses face the same timing mismatch. 庄明浩 expects a wave of Chinese manufacturers to launch products in Q1-Q2 2025, but most projects were started in mid-2024, before the latest improvement in reasoning models. The AI in the first generation “may still not be very good,” and will need to be filled in through subsequent iterations.
The good news is that the barrier to starting has fallen in parallel for individuals. Code, design, images, open-source solutions and To B cloud services are expanding what one person can do. Work that once required a team of 10 or 20 may now be handled by an independent developer who “plays director” or “plays editor,” leaving design, programming and operations to AI and service providers.

6. Big platforms are using DeepSeek’s traffic to train their own models

沈帅波 asked whether Kimi, Baidu, Hunyuan and Tongyi had been pushed out of the game by DeepSeek. 庄明浩’s answer was “borrow the fake to cultivate the real”: DeepSeek’s attention must be “hijacked wherever possible,” or someone else will capture the traffic. What matters is where users are sent after integration and what the feedback is used to train.
WeChat Search routes some users to Yuanbao. Although Yuanbao offers DeepSeek, it defaults to Hunyuan on first launch and requires users to switch manually. Hunyuan has also launched a reasoning model, while Tongyi Qianwen released its own reasoning model around the same time. The hot capability is borrowed; the users, data and habits are retained.
As of the recording, among the main apps with their own models or development capabilities, only Kimi, Doubao and Alibaba’s Tongyi had not integrated DeepSeek. 庄明浩 called this a form of “luxury”: most of the world had already surrendered, while these 3 companies at least still believed their own model evolution had a chance. He was uncertain whether they would integrate it later.
Reasoning models are like Level 2 on the path from L1 to L5 autonomous driving: “The L2 war is nowhere near over, but L2 is a threshold.” Even companies such as Xiaomi and Li Auto will retain proprietary models for industry data, personalized needs and safety after integrating DeepSeek. Encountering DeepSeek just after announcing ambitions to enter China’s top 3 large-model companies is painful, but abandoning the effort is not an option.

7. DeepSeek has turned the AI capex narrative into a fight over competing routes

The dominant narrative over the past 2 years was a simple progression: spend more money on infrastructure, train stronger models, then use those models to drive growth. Before the Lunar New Year, technology giants were still raising 2025 spending along that path: Microsoft at roughly $80B, Google at $75B, Meta at $60B and Amazon at $100B, plus the $100B proposed by Trump. Planned spending totaled about $420B, up from roughly $220B the prior year.
DeepSeek suddenly showed the market that “at least on the surface, there may be another path,” potentially bypassing the logic of simply spending more. 庄明浩 was particularly surprised that Microsoft CEO Nadella voiced a negative judgment publicly at this moment. One interview even ran under the provocative headline that the relationship between OpenAI and Microsoft had broken down. Microsoft then canceled 2 data centers, naturally prompting the market to ask whether what it said 2 weeks earlier still held.
The previously unquestioned upward spiral has turned into disagreement, and possibly a downward spiral. The assets that had risen the most are likely to come under pressure first. Palantir, which provides AI systems to the CIA and enterprises, had a run last year “more妖 than Nvidia” and fell 10% that evening. 沈帅波 also noted that Jensen Huang has repeatedly argued demand will continue expanding, citing the trajectory of coal prices after the Industrial Revolution as support for the existing consensus.

8. AI has not yet proven inclusive growth; productivity may take a generation to arrive

沈帅波 pushes the question up to the macro level: companies use AI-driven price competition and cost cutting, end prices do not rise, and white-collar middle-class jobs are cut instead. If wealth flows only to chip sellers, data-center builders and a handful of model companies, while other industries see no revenue growth, purchasing power contracts further and the economy may become more hollowed out than the financial system.
庄明浩 responds with the electricity revolution: roughly 30 years passed from the invention of the electric motor to a genuine increase in economy-wide productivity. The first factories simply replaced a central steam engine with a more expensive central electric motor; the buildings, transmission methods and organizational structures stayed unchanged. Efficiency was released only after motors became smaller and reshaped factory layouts.
The “compute—connectivity” framework proposed by 张斐 of 5Y Capital points to the same conclusion. After computers came the internet; after energy comes ocean-going, infrastructure-like connectivity. AI may open a new era of computing or energy, but “nobody knows” what the next generation of connectivity will look like. We can only try and keep working at it.

9. AI first benefits incumbent giants, but whether Xiaomi’s gains count as an AI dividend remains disputed

庄明浩’s interim judgment is that AI is currently helping companies with stable existing businesses far more than entirely new startups. Giants already have businesses and use cases, so they can benefit directly from either cost reduction or productivity gains. Startups still have to prove both product and business model.
庄明浩 first called Xiaomi the clearest example, noting that its stock had not risen in 7 years after listing before later doubling, then acknowledging that “they made cars.” He subsequently said this wave was Alibaba’s—in his view, Alibaba has already evolved from an e-commerce company into an AI company. The disagreement over attribution is worth preserving.
For Xiaomi, AI at least makes the “people-car-home” story more complete: what used to be IoT is now AIoT, making a standalone hardware founder look like they are simply completing one link in Xiaomi’s ecosystem. Lei Jun’s personal image simultaneously spans hardware, the physical economy, technology, AI and autos, so his appearance at the entrepreneurs’ symposium was not accidental.

10. Strategic investing means protecting the “hard realities” while retrying the “poetry and distance”

庄明浩 defines himself as a strategic investor who starts with areas reachable in 1 or 2 steps from the existing business. His team worked in social and voice products, so it looks at AI companionship, AI emotion, AI social products and voice models, while also extending from software into hardware such as toys. Some capabilities are built internally; others are added through investments and partnerships.
He calls this the “hard realities at hand,” while the “poetry and distance” is what AI-native social products and games will actually look like. The metaverse was not fully explored in the previous cycle and proved less successful than hoped, but AI offers a chance to try again. He agrees with the principle: “If you believe a huge direction is correct, you should try it once every 3 years.”
Differences between the U.S. and Chinese ecosystems also determine where opportunities land. Since around 2010, the U.S. developed along To B, SaaS and cloud services, while China built capabilities in To C businesses such as WeChat, Alipay, Pinduoduo, ByteDance and Douyin. The two sides are beginning to converge in the AI era, but China may be the place to look for high-uncertainty consumer products.

11. Wrappers are not low-grade opportunities; clearly bounded products can already generate $10M in revenue

On whether a new giant company will emerge, 庄明浩 first points out that ChatGPT itself is already a killer app. It may not have network effects, but its brand effect is established. Based on past experience, if a U.S. industry produces a $100B company, China will “at least certainly produce” a similar $10B company. One or 2 leading AI startups may reach sufficient scale.
He opposes primary-market investors starting with the endgame. If investors in the early mobile-internet era had backed only the eventual winners, they might have ended up with ByteDance, Pinduoduo and Meituan. Saying “housing prices will fall one day,” or “everyone eventually dies,” may be correct but cannot guide stage-specific investing. “You make stage-specific money.” Once a company reaches an M&A or IPO milestone at a sufficient multiple, that can be enough.
Small businesses work too. Over the past 2 years, many Chinese teams packaged large-model capabilities into better user experiences and sold subscriptions or advertising overseas, reaching around $10M ARR. Their technical barriers may not be high, but they launch quickly, identify pain points accurately and execute well—proving that a delivery gap between models and users will always exist.
AiPPT is 庄明浩’s chosen example. Large platforms can all build similar functions, but a first mover with a good name that understands advertising and SEO can charge users both overseas and domestically. It does not need to satisfy professional creators who refine every slide; it only needs to help ordinary people prepare a weekly report or a child’s speech. As long as the product’s delivery matches its boundaries, “if there is a shell, fill it—there is no reason not to.” He also relayed 任心’s reflection that looking down on wrappers had been a mistake.

12. Not trading stocks is both a risk preference and a contrarian sentiment indicator

庄明浩 does not open a brokerage account not because he has been hurt, but because he is “afraid.” He believes he is extremely conservative by nature; if even he is emotionally pulled around by the market, “that is basically a signal of a top.” So “once you start thinking about getting in, hindsight will prove you went in to become the bagholder.”
He does not like playing Texas Hold’em himself, but enjoys standing nearby and watching because he does not want to enter the tense emotional arena while still enjoying the analysis of rules and human nature. After the Lunar New Year, he too briefly felt that Chinese stocks and U.S. stocks could “do anything,” but if he had chased them then, his account might have been down 10%-20% by the time of recording.
沈帅波 described being stranded in Queenstown after repeated flight cancellations, taking 3 days of trouble to return to China. “China’s AI industry is in full swing; Queenstown’s tranquility and distance could no longer contain my heart,” so he only wanted to get back quickly. He summed it up as Chinese people’s fear of “missing an era.” The rapid emergence of training courses, educational content and even books teaching people how to use DeepSeek became a commercial portrait of the mass frenzy he described.

13. When consensus valuations are built in 6 months, the real dead end is the exit

This is not unique to AI; it continues the post-pandemic liquidity wave. Some leading dollar funds doubled and redoubled from $200M to $800M while still investing in roughly 20 companies, lifting the budget per deal from about $10M to $40M. At their most aggressive, Sequoia and Hillhouse invested in roughly 300 deals a year, with “do not let the other side invest” becoming a KPI in hot sectors.
The result was a compression of the industry cycle from 2 or 3 years, or even 5 years, to 1 year and then 6 months. A company could go from obscurity straight to a $500M or $1B valuation, while AI layered consensus around the direction and the founder’s persona on top of each other. If a fund followed its old process of coffee meetings, team meetings, term sheets and investment committee approval, the second investor’s term sheet would already be on the table. The choice was all-in or out.
A paper valuation is not cash returned. LPs might ask the previous year why a fund failed to invest in Kimi, MiniMax or StepFun, only for the market’s direction to reverse the following year. Investors therefore care less about logos and paper multiples and more about DPI: “How much money did you put in, and how much did you return to me?”
The only exit paths are listing, acquisition, buyback or zero. Other than Zhipu, most leading model companies were funded by dollar funds. U.S. listings face a difficult environment; comparable companies in Hong Kong are valued at roughly HK$30B-HK$60B, while new-generation projects are entering at $4B-$5B valuations. Whether the A-share market will open the door, and for how many companies, is equally unknown. Giants are all building internally, making acquisitions more likely to take only the team, while buybacks require extremely strong commercialization.
Humanoid robots therefore fit RMB-fund preferences relatively well, while pure applications lean toward dollar funds—but this remains a “walled city.” Unitree’s old shares may already price in 5 years of technology, commercialization and IPO expectations. UBTech, the first humanoid-robot stock, has also fallen badly. Later-round MiniMax investors may go from “finally putting the logo into the LP deck” to worrying that they themselves are the final-round bagholders.
庄明浩 does not view the dead end as the end of the industry: “A great founder goes and solves problems that look dead.” But ordinary founders also need to confront reality earlier: whether they are building in China, Japan, Singapore, the U.S. or the Middle East; whether to move their family; and how to repay the first money raised within a few years—rather than assuming the next round and an IPO will inevitably arrive.

103: The Evolution of Attention Optimization: DeepSeek and Kimi Kerrisdale Capital's Sahm Adrangi on $ACMR's price dislocation between Shanghai and NASDAQ listings