Vol.85: DeepSeek One Year On, Ahead of Yuanbao’s RMB10B—厚雪长坡 Crossover

Summary

• DeepSeek R1 did not merely rewrite one model leaderboard; it changed China’s identity from AI spectator to core player. As the US unveiled “Stargate” and its $500B investment, R1 created a clash between a single training run costing several million dollars, open reasoning methods, and performance close to leading closed-source models—an apparent gap of several thousand times; Nvidia fell as much as 17% in one day, wiping nearly $600B from its market cap, while 冯骥’s description of R1 as “potentially a technology achievement on the scale of national destiny” became the emotional flashpoint.
• China’s model war will not end with one dominant player, but new entrants can no longer appear out of thin air; the likelier outcome is “a few dozen” companies each defending its own territory. 庄明浩 believes some Chinese companies have used open source, engineering efficiency, and local-market scenarios to reach “80 to 90 points” with investment on the order of 1% of OpenAI’s, with paths spanning To C, To B, government, embodied intelligence, and advanced manufacturing; capital markets are forcing the giants to fill in chips, cloud, models, applications, and even robots—“the market rule is that you need all of them.”
• The technical variables worth tracking most closely in 2026 are not version numbers, but whether reinforcement learning, multimodal fusion, and online learning can keep lifting the capability curve. 庄明浩 estimates that at least roughly 80% of pretraining data may already have been used, though pretraining engineering still has room to improve; reinforcement learning is perhaps only 40%-50% complete, with expert data, benchmarks, and reward hacking still to be addressed. Further out lies continual learning, in which models continuously receive feedback and update themselves—“pretraining is oil; post-training is new energy”—and if continuous learning works, it could deliver a leap comparable to the rise of nuclear power.
• Agents have not failed; the spotlight has simply shifted from “watching AI work” to the infrastructure beneath the waterline. Manus gave users their first direct look at AI opening webpages, clicking buttons, searching for information, and making PPTs, but real deployment requires AI-native browsers, databases, identity, permissions, and protocols such as MCP and ACP; because these capabilities implicate platform interests, liability, and commercial ownership, the move from language to action “cannot be solved suddenly in the short term,” making infrastructure such as Browser Use the clearer opportunity.
• Valuations for leading US model companies have entered a right-side market that prices in the endgame first and waits for delivery later. OpenAI’s expected valuation of at least $1T against roughly $20B in revenue implies about 50x PS; Anthropic is valued at roughly $350B-$370B, while xAI has been priced at around $230B on roughly $500M in revenue, with the core investment case reduced almost to “we trust Musk.” With Google already worth $3T-$4T and the entire leading AI ecosystem modeled at roughly $20T, relative valuation can lift the second-place company and the whole supply chain together.
• Industry fundamentals are still moving higher, but stock returns can no longer be treated as equivalent to technological progress. TSM raised its next-year capex outlook from roughly $49B to $52B-$56B and said it was “very satisfied” with actual AI demand after a three-to-four-month survey covering “customers and customers’ customers”; 庄明浩 sees the supply chain’s most sober participant offering clearer confirmation of demand, but that cannot erase volatility from financial leverage and crowded trades. “The technology will definitely iterate, and AI tools will keep getting better, but the stock price does not necessarily have to rise.”
• The battle for AI entry points and personal usage have both entered the monetization phase: products must fight to become platforms, while individuals must turn them into daily leverage. Doubao has surpassed 100M daily active users; according to 庄明浩’s account, ChatGPT “should have been” at 800M weekly active users around last October, with rumors now putting it at 900M, and both are believed to have developed a retention “smile curve”; Alibaba is the most aggressive, Tencent is using its deep balance sheet as a backstop, and ByteDance is betting on multimodality and Spring Festival traffic. The most practical rule for individuals is, “if you do something more than 5 times a week, try handing it to AI”; investors can also ask it to play the other side of the trade, but information convergence, risk appetite, and betting discipline must ultimately remain their own.

Deep dive

1. DeepSeek R1 Turned China from Spectator into Core Player

• 七一 dates the starting point to DeepSeek R1’s release on January 20, 2025; 庄明浩 remembers the real public-opinion flashpoint as the evening of January 26, when 冯骥 wrote after 5 days of testing: “DeepSeek may be a technology achievement on the scale of national destiny.”
• Before that, Trump had just announced the $500B “Stargate” project involving OpenAI, SoftBank, and Oracle. On the other side was a Chinese model with a single training run costing several million dollars and performance that was “about the same, or even looked better”—a narrative gap of several thousand times.
• Nvidia then fell as much as 17% in one day, wiping nearly $600B from its market cap. 庄明浩’s summary: at least in the eyes of the public and media, China “suddenly became a core player.”

2. R1’s Shock Came from the Reasoning Leap—and from Fully Exposing the Method

• 庄明浩 uses OpenAI’s layers to explain the technical significance: Chatbot is L1, reasoning models are L2. OpenAI released O1 around the boundary between Q3 and Q4 2024, but it was closed source, leaving peers worldwide to attempt their own reproductions.
• Built on V3, released at the end of December 2024, R1 used reinforcement learning to produce visible reasoning traces. DeepSeek published not only its training and implementation methods, but also applied the same method to open-source models such as Llama, proving that performance could improve there too.
• That made “open source” one of the most important keywords in Chinese AI in 2025. Chinese open-source models of different sizes occupied a large share of the top 10, and even the top 20, on many leaderboards, forcing OpenAI to open up some of its non-optimal models for adaptation.
• 庄明浩 believes that once OpenAI completed the basic paradigm’s journey from zero to one, competition shifted toward engineering, implementation, cost, and efficiency—precisely the areas where Chinese engineers are more likely to excel.

3. 2025’s Main Line Ran from R1 and Manus to Google’s Multimodality

• Looking back over the year, 庄明浩 says Manus emerged in March and Agent discussions continued until around May; after US equities fell sharply in April on tariffs and other factors, the market gradually recovered, though the S&P 500’s path to new highs was anything but smooth.
• After the much-anticipated GPT-5 launch, the model was “not quite as ideal” and did not open up the kind of huge lead seen in previous generations. By Q3, Google was steadily strengthening its multimodal position with Veo 3, Nano Banana, and Nano Banana Pro; GPT-5 and other models followed toward year-end.
• DeepSeek was continually chased by expectations for R2: one round of rumors emerged in May and June, and another intensified in September and October around V3.1 and its suspected “final version.” The market thought V4 would come next, only to receive V3.2 a week later.
• Entering early 2026, DeepSeek’s successive papers focused more heavily on model architecture, leading 庄明浩 to guess that V4 would “most likely” arrive around the Spring Festival. He also raised the possibility of skipping an intermediate version and moving directly to the next generation, while emphasizing: “We don’t know.”

4. User Scale and the “Smile Curve” Show AI Applications Have Crossed a Threshold

• Doubao surpassing 100M daily active users is a major App milestone. According to 庄明浩’s account, ChatGPT “should have” reached 800M weekly active users around October 2025, with subsequent rumors putting it at 900M.
• More important is the signal that both are believed to have achieved the so-called retention “smile curve”: retention normally declines over time, so a renewed rise implies a qualitative shift from the compounding of user scale, brand, and functionality—the product may have crossed a threshold.
• After ByteDance secured sponsorship of the Spring Festival Gala, 庄明浩 judged that Doubao’s daily active users that night “will definitely be at a peak”; whether it breaks 200M is still unknown. Alibaba, meanwhile, began visibly increasing To C promotion for Qwen in Q4.
• 七一 asked how Tencent and Yuanbao would respond. 庄明浩 believes DeepSeek’s new version, Doubao’s Spring Festival traffic, and Qwen’s continued investment will concentrate the pressure on Tencent, 姚顺雨, and 张小龙, though the division of labor remains unclear.

5. China’s Model Market Will Not Be Reduced to 3-5 Players; “A Few Dozen” Is More Likely

• Asked whether the outcome would be 3-5 companies or 50-500, 庄明浩 gave a range of “a few dozen.” New entrants can no longer emerge from nowhere, but models are not a single platform, so one company is unlikely to dominate outright.
• Hundreds of large models have been registered in China, though far fewer are still under active development. Beyond the internet giants and frontier startups, Baidu, Xiaohongshu, Bilibili, Xiaomi, Meituan, iFlytek, and companies serving To B and government clients are all searching for their own scenarios.
• China has also avoided being locked into the US-style single mega-financial narrative. Open source, engineering implementation, To C, To B, embodied intelligence, and advanced manufacturing can all compound, creating far more paths than “taking one road to the end.”
• 庄明浩 uses the bluntest but most explanatory number: some leading Chinese model companies spent roughly 1% of OpenAI’s money and achieved “80 to 90 points,” while fitting China’s conditions better.

6. The Value of Agents Is That They Let Users “See AI Working” for the First Time

• In 庄明浩’s framework, L1 Chatbot, L2 reasoning, and L3 Agent are no longer sequential stages that replace one another; they are being developed simultaneously and layered together like a Gantt chart.
• Language models can only converse, but the technological revolution ultimately has to “do work” and raise GDP like the steam engine did. The industry must therefore move from language to behavior, and from conversation to action.
• Manus was most impressive not because it announced the Agent concept, but because it let people see AI open webpages, access them, click, collect data, and make PPTs. “This is what the Agent everyone expected should look like.”
• 七一 noted that a similar effort by Doubao was quickly restricted, which only reinforced the importance of the visual stimulus: ChatGPT let people see machines produce text, R1 let them see machines think, and Agents let them see machines execute.

7. Agent Hype Cooled Because the Real Work Is Hidden Underwater

• Language models mainly handle compute, algorithms, and data. Agents must connect to webpages, databases, permissions, ride-hailing, food delivery, and other external environments, making their complexity closer to autonomous driving; one model company cannot solve it in the short term.
• Today’s browsers, databases, CAPTCHAs, and SMS verification were designed for humans. If AI is to take on more work, it needs AI-compatible browsers, databases, identity, and access permissions; Browser Use is rebuilding precisely this layer.
• Websites often identify AI traffic as crawlers or attacks, which is why protocols such as MCP and ACP have attracted attention. 庄明浩’s reminder from internet history: HTTP, FTP, and SMTP existed before consumer interfaces, whereas with AI the interface arrived first and only then exposed the missing underlying protocols.
• The harder issue is incentives. Early internet protocols did not belong to any one company; today’s protocols are promoted by commercial companies. “Why should I use your protocol? I want to build a protocol too.” Liability, revenue, and control naturally slow progress.

8. DeepSeek’s Strongest User Impact Was Turning “Thinking” into a Visible Interface

• The DeepSeek App initially had no deep-thinking mode, then added online search. The 2 buttons created 4 combinations, and 庄明浩 considered the experience complete only when deep thinking and online search were both enabled.
• Once the model slowed down and displayed how it broke down a problem, users could see for the first time how it decomposed a question into a logical, structured task. More strikingly, it often surfaced angles users had not considered themselves.
• AI beating Go, driving, or playing StarCraft was not something ordinary people necessarily encountered every day. Now it can write emails, analyze problems, and organize talking points—“think like you”—which feels entirely different to carbon-based life.

9. Claude Proved Text Understanding; Nano Banana Pushed “Say It and It Happens” into Vision

• 庄明浩 recalls that by the end of 2024, Claude could write in “The Wall Street Journal style.” That required the model to understand the genre’s structure, expression, and development—not merely generate words.
• In the first half of 2025, GPT-4o triggered a wave of Ghibli-style profile images, likewise demonstrating the model’s abstract understanding of color palettes, styles, and visual elements.
• With Nano Banana, Nano Banana Pro, and Veo 3, ordinary users could measure progress through “say it and it happens”: the question was no longer whether an image looked similar, but whether the model could directly align abstract intent, common sense, and visual output.

10. One Elementary-School PPT Showed How Multimodal Models Extract Story Structure

• 庄明浩 was preparing an AI class for sixth-grade elementary students, hoping to explain both how models work and how to use them for learning. He wanted content the students could understand, ideally built around a familiar story.
• The first version used Naruto. The model mapped the Sharingan’s copying ability to AI learning and connected the “Copy Ninja” to the model’s capabilities at the level of logic; his son pointed out that perhaps half the class did not know Naruto.
• After switching to Eggy Party, Gemini accurately understood the Chinese game IP. The child then worried that the teacher might dislike a game-based topic, so they ultimately chose Journey to the West.
• The final metaphor was: “You are Tang Sanzang, the model is Sun Wukong, and you need the tightening spell to control it.” Data, algorithms, and compute were placed inside Taishang Laojun’s alchemy furnace; the model generated most of these mappings, and they fit the story logic closely.

11. Language, Images, and Video Are Merging from “Three Tables” into One Model Core

• In the past, language, images, and coding were treated as 3 parallel lines. PPT tools usually generated an outline first, then applied a template and added images, like separate workshops assembling a product.
• 庄明浩 has changed his view: the future may have only “one table”—multimodal in, multimodal out, with a single core doing the understanding in between. Qwen’s “three in, three out” means text, images, video, or audio can all be inputs, with 3 types of outputs supported.
• Pure language models are also filling in the gaps themselves: Zhipu released its first image model, and DeepSeek released an OCR model, showing that the industry believes many problems can no longer be solved through text alone.
• In serious search and reporting, 庄明浩 considers Gemini very strong at organizing, collecting, and processing information because search ranking, source importance, and information convergence are Google’s longstanding strengths. ChatGPT puts more emphasis on “understanding you” and emotional value, but still needs third-party capabilities for serious search.

12. xAI’s Valuation Is Essentially a Vote for Musk’s Entire Empire

• xAI has been nearly as aggressive as OpenAI in building compute and uses X’s real-time corpus for social interaction, companionship, and product synergies. 庄明浩 guesses that xAI and X are “very likely” to merge further in 2026, though this is only a projection.
• Tesla, SpaceX, space-based data centers, and other narratives can be linked into a closed loop, allowing Musk to keep raising capital. When discussing xAI, “you cannot talk only about xAI itself.”
• At the scale 庄明浩 cites, OpenAI generates roughly $20B in annual revenue, SpaceX about $9B, and xAI about $500M. Their valuations, however, are roughly $500B for OpenAI’s previous round, about $700B for the current round, $350B-$370B for Anthropic, and $230B for xAI.
• 七一 asked why xAI could command $230B with revenue an order of magnitude lower. 庄明浩’s answer was direct: investors can only write, “We trust Musk”—“go ahead, do your thing.”

13. Meta’s Problem Is Not Money; Its Model Positioning and Cloud Monetization Are Both Unstable

• 庄明浩 sees Meta as more of a product, application, and distribution company than a company known for foundational technology. After the Llama team went through architecture changes, leaderboard controversies, and other disputes, the role and objectives of its internal AI lab remained unsettled.
• As Microsoft, Google, Amazon, and other cloud providers put capex in the $100B range into related infrastructure, Meta followed suit without a public cloud business to absorb the investment. “That $100B is effectively floating in the sky.”
• Zuckerberg aggressively recruited talent in the middle of the year, resembling the assembly of special forces or mercenaries, then put young leaders in charge of chasing a curve that was already rising rapidly. The result was that Llama was nearly silent in Q3 and Q4.
• The positive factor is that AI has clearly improved advertising efficiency, a point Meta, Google, and Chinese internet companies can all make. But that only explains better performance in the legacy business; it does not answer whether Meta must become a top foundational-model company.

14. The Five-Layer Cake Is Becoming a Mandatory Scorecard for Tech Giants

• The market has begun comparing companies layer by layer across chips, data centers, cloud, models, and applications, sometimes adding embodied intelligence and robots. Google is strong because “it has every layer, and every layer is strong.”
• Microsoft, Amazon, Meta, and even Apple are being asked which layer they lack. Amazon’s space, robotics, and warehouse-efficiency efforts are also being placed in this complete-loop framework. 庄明浩 is unsure whether the standard is reasonable, but “this is how the market evaluates you.”
• Alibaba, ByteDance, Huawei, and Baidu are likewise being placed into the matrix of cloud, data centers, chips, models, and applications, though the strength and priority of each layer differ.
• Listed startups such as Zhipu and MiniMax currently occupy only 1 or 2 layers, mainly foundational models and applications. The AI war may continue for another 5 years, and founders must approach being “hexagonal warriors.”

15. New Lab and DeepSeek Are Searching for Paths Beyond Big Capital

• In Q4 2025, the US saw a wave of New Labs founded by researchers from OpenAI, Google, and Meta. Some even called themselves “Labs” rather than companies, securing high valuations and large financing rounds before having products.
• Capital is betting on a new R&D paradigm: funding and physical requirements along the existing path have been pushed to their limits, while leading companies have placed the overwhelming majority of money, people, and compute into the first 2 categories of research, leaving an entrepreneurial window for a third level of exploration.
• DeepSeek is even more unusual. Its connection to quantitative trading runs deep, making its commercial role difficult to explain using the standard model-company template. It can attend the top-tier gatherings, or not show up, while maintaining its own pace.
• Borrowing a joke from an article, 庄明浩 says US model companies might also consider launching quantitative funds to generate cash flow. His assessment of DeepSeek: “It has a bit of the Dugu Nine Swords feel”—unconstrained by mainstream commercial logic.

16. Pretraining Is Not Over, but Reinforcement Learning Remains the Main Battlefield for Capability Gains

• 庄明浩 estimates that at least roughly 80% of usable pretraining data may already have been fed into models. The remaining data may not be digitized, may be difficult to process, or may be too low quality. But Google’s improvement also shows that data cleaning, filtering, and pretraining engineering still have room to unlock gains.
• He calls 2025 “the year of reinforcement learning.” Progress was fastest in mathematics and coding because the answers are verifiable: math is either right or wrong, and code either runs or does not, giving the reward signal natural clarity.
• Fields such as medicine, law, and finance have no single correct answer. Senior doctors, lawyers, and traders must define what counts as good within human cognition, making data labeling and expert feedback important again.
• He estimates that reinforcement learning may be only 40%-50% complete. The industry still needs to build environments, create scoring systems and benchmarks, and prevent reward hacking: “Give me a KPI, and I will always find some way to game it.”

17. Online Learning Could Be the Next Paradigm Leap; Token ROI Determines Who Survives Until Then

• Current models are trained at discrete points, frozen, released, and then followed by the next generation. Online learning or continual learning would let models update continuously from real-time feedback; the logic is feasible, but the engineering and resource demands are more complex.
• 庄明浩’s metaphor is that pretraining is oil: powerful but finite. Post-training is new energy, which can continue to substitute for it. If autonomous continual learning becomes real, it would look more like the leap brought by nuclear power.
• For model companies, the question becomes token ROI: should limited compute and R&D resources go to pretraining, reinforcement learning, or continual learning, for how long, and with what capability gains and revenue returns?
• 杨植麟 of Kimi, 唐杰 of Zhipu, and others have discussed similar questions. 庄明浩 believes this is no longer just a choice of technical route, but a new resource-allocation problem for model companies.

18. Tokens Becoming 10x Cheaper Every Year Does Not Automatically Give Frontier Applications Gross Margin

• Some argue that the token cost of the same model falls by an order of magnitude every year. 庄明浩 emphasizes that this is a static comparison: the old model becomes cheaper a year later, but a stronger and more expensive new model appears by then.
• In 2024, many applications justified having no gross margin by saying costs would fall to one-tenth the following year. But as long as capabilities have not converged, frontier products must keep calling the best models, and gross margin will remain elusive.
• The potential inflection comes when certain scenarios begin to converge: enterprises no longer need the largest, newest model, and a smaller, cheaper model is sufficient. Only then does the cost curve truly translate into commercial profit.

19. World Models Aim to Give Robots a Virtual Reality for Repeated Trial and Error

• 庄明浩 explains world models as virtual environments that obey real-world physics: slopes have gradients and speed differentials, stones have weight, objects sink in water, and a glass dropped on the floor shatters.
• Gaming is the most obvious use. More important is generating training data for autonomous driving and humanoid robots. A specialized surgical robot faces a limited range of materials and movements, while a general-purpose caregiver robot must “go anywhere and do anything.”
• AI-generated simulated environments let robots climb mountains, cross water, and grab objects repeatedly at low cost; crushing a cup causes no real-world loss. The physical-AI path represented by 李飞飞’s team is targeting this layer.
• Another visual path starts from generating images and video, aiming for real-time rendering and instruction-based modification—for example, making a person in a podcast scene stand up and knock over a chair. Which path matters more? 庄明浩’s answer remains: “Who knows? In any case, a lot of people are working on all of them.”

20. Vertical Applications Will Grow, but the Real Value Still Lies in Moving from Tools to Platforms

• To B can be segmented by function, such as customer service and sales, or by industry, such as healthcare, finance, and law. These sectors are large and contain valuable proprietary data, so they will be early adopters of AI.
• US revenue rankings already include vertical companies in law, finance, sales, and video editing alongside foundational-model companies. 庄明浩 expects opportunities to continue emerging in both To B and To C in 2026, with multimodality unlocking more scenarios.
• Over the past 3 years, AI applications have mostly been tools, and tools are usually not large enough. Leading companies will inevitably try to build platforms in social, content, e-commerce, and other forms, creating network effects and business models; ChatGPT’s addition of advertising is just one step in that process.
• The industry value chain is currently an “upright triangle”: chips and infrastructure at the bottom are the largest, applications at the top the smallest. Traditional IT should have the largest application layer. Nvidia has already reached roughly $4T; if the value chain inverts, the upper layers must open up much more valuation space.

21. OpenAI’s $1T Valuation Is Pricing in Its Endgame Position Ahead of Time

• 庄明浩 applies the “China 1% rule” to both models and GPUs: leading Chinese model companies achieve 80 to 90 points at roughly 1% of the cost, and their valuations are also about 1% of US peers. Leading domestic GPU companies are valued at roughly 1% of Nvidia’s $4T-$5T.
• OpenAI’s expected valuation of at least $1T against roughly $20B in revenue implies about 50x PS. The market is not pricing current revenue; it is assuming that leading AI companies could eventually be worth roughly $20T in total, then estimating how much OpenAI can capture.
• OpenAI’s external projections put 2030 revenue at roughly $200B-$300B per year, while Google already generates about $100B in a single quarter. Subscription, advertising, e-commerce, API, and other revenue streams are not impossible to count.
• 庄明浩 believes private and public markets will “fully price in expectations that can be calculated.” As with Kuaishou before its 2021 listing, comparing Bilibili, Kuaishou, and Tencent created valuation transmission that lifted the entire group of assets.

22. The Cloud Alliance Has Gone from Exclusive Ties to “A Total Mess in Northwest Shanxi”

• Amazon’s stock once rose about 12% in a single day after a Q3 or Q4 earnings report, driven by AWS growth returning above 20%. Previously, Microsoft’s cloud business and public cloud were growing around 30% per quarter, while AWS had fallen to the low double digits.
• The early structure was Microsoft tied to OpenAI, while Amazon and Google invested in Anthropic. 庄明浩 says Microsoft and OpenAI had an “amicable breakup” in November; once exclusivity restrictions loosened, Microsoft also participated in Anthropic’s financing.
• Google then increased its investment in Anthropic and signed large TPU and cloud agreements. Nvidia, Microsoft, and others also crossed into financing rounds, while the old camp boundaries were repeatedly broken apart by capital and compute contracts.
• The risk first landed on highly leveraged New Cloud companies such as CoreWeave and Nebius, then spread to Oracle’s future cash flow and AMD’s data-center expectations. By early 2026, the giants’ share prices were relatively stable; the hottest names remained memory companies such as Micron and SanDisk.

23. The Phone Remains AI’s Most Realistic Entry Point; Apple Has One Core Question Left for 2026

• 七一 observed that Google is tied to Apple and Microsoft to OpenAI; after Doubao’s application was restricted domestically, it might also seek a handset partner. 庄明浩 agrees that AI is unlikely to create an entirely new giant hardware category in the short to medium term and will remain dependent on the phone as its central vessel.
• That is why Chinese giants resumed the “entry-point war” in Q4 2025, replaying the mobile-internet competition of more than a decade ago. Whether an unexpected device, rival, or decision-maker will emerge remains unknown.
• Google was among the strongest performers of the Magnificent 7 in 2025, while Apple was relatively the weakest. If Apple is instead the leader by the end of 2026, AI must have produced a huge marginal change—not merely allowed the existing business to remain stable.
• The difficulty is that Apple’s business is too deep and Cook is not aggressive. Who succeeds Cook, what AI stance that person takes, and how quickly it is executed constitute what 庄明浩 sees as “the only Apple question for 2026.”

24. Alibaba Is the Hungriest, ByteDance the Most Focused, and Tencent Is Buying Time with Its Deep Reserves

• 庄明浩 believes the valuation repair in Chinese internet stocks over the past year was broadly rational, but AI alone cannot explain the entire gain; sentiment around US-China tensions also provided a boost.
• Alibaba is the most aggressive: it is raising the standing of Alibaba Cloud, integrating Qwen with its businesses, and pushing the Qwen App toward To C in Q4. The old message was “Qwen empowers every business”; now it looks more like every business is being loaded into Qwen, with the person in charge holding the “imperial sword.”
• After organizational changes, ByteDance separated model and application teams and made Doubao independent. Its path is closed source, multimodality, and To C. Even as DeepSeek changed the battlefield, ByteDance kept its objective clear, and Doubao had previously appeared not to require extravagant promotional spending.
• Tencent is more like Apple: its legacy businesses in games, advertising, and payments are too stable, giving it the “right to be resilient.” It has Yuanbao, integrated with DeepSeek, improved its existing businesses with AI, and brought in 姚顺雨, but how 姚顺雨 and 张小龙 will work together remains a key variable.

25. Giants’ AI Postures Are Ultimately Shaped by Founder Temperament and the Capital-Market Clock

• 七一 asked whether it is better to be aggressively anxious or stable and understated. 庄明浩 believes there is no universal answer: companies remain expressions of founder temperament, whether Tencent, ByteDance, Alibaba, Pinduoduo, Meituan, JD.com, or Baidu.
• AI efficiency gains in Tencent Meeting, Docs, search, Official Accounts, and games can “hold the floor,” but are “not enough to raise the ceiling.” For companies with deep reserves, the floor is already valuable, but short-term capital still needs a clear flag.
• Alibaba’s culture is naturally suited to launching campaigns continuously and concentrating forces to win, making it a better fit for the sentiment of an AI market. Tencent’s corporate objectives, however, are not solely about serving the share price; investors and the company occupy fundamentally different positions.
• As for whether wunderkinds or 30- and 40-something veterans are more reliable, he offered no abstract answer. Tencent’s decision to choose young researchers at this moment itself shows that management is not unconcerned; its level of anxiety is simply moving dynamically.

26. The AI Rally Has Spread from the Magnificent 7 to the Entire Supply Chain

• 庄明浩 cites a set of statistics he calls “COTWO,” dividing stocks into 3 groups: the Magnificent 7, non-Magnificent-7 AI companies, and non-AI companies.
• The Magnificent 7 led in 2023. They still led in 2024, but by a narrower margin. In 2025, non-Magnificent-7 AI companies took the lead, including chips, energy, AI software, AppLovin, Snowflake, and Salesforce, while the Magnificent 7 rose roughly in the low-teens range.
• China’s diffusion appears to lag by about 1 year: the past year was led by technology giants, while the next phase may turn to applications, infrastructure, and edge battlegrounds. The giants may keep rising, but with more limited gains.
• This is not a stock-picking conclusion. The US has already rotated into memory, and China is trading similar links in the chain. As the theme reaches the downstream end, investors must distinguish earnings delivery from relative-valuation expansion.

27. The Key to the Bubble Debate Is Not “Whether It Exists,” but Whether the Market Is at the Midpoint or the Summit

• The AI bubble began attracting discussion in Q2 2025 and peaked during the most aggressive circular financing in Q4: Nvidia invested in AI, OpenAI invested in Broadcom, Broadcom invested in AMD, Oracle invested in other companies, and capital and orders continuously crossed.
• By the time the program was recorded, the debate had actually cooled, and the market was beginning to accept that “existence is its own justification.” Almost all major US financial institutions had an upward outlook for 2026, but the concentration of consensus could itself become a risk.
• A technological revolution must pass through a noisy early phase. What cannot be determined is whether the market is at the midpoint or the summit. If it is the midpoint, the rally can continue; if it is the summit, a meaningful correction comes first.
• 庄明浩 observes that drawdown-and-recovery cycles are compressing from years to 6 months or a quarter; the April 2025 selloff was recovered quickly. So even another major correction might not necessarily become a long-term problem spanning calendar years.

28. TSM’s Capex Increase Is the Supply Chain’s Calmest Confirmation Yet

• 庄明浩 lays out the chain in order: users, applications, models, cloud, data centers, chips, foundries, and only then raw materials and silicon.
• TSM raised its next-year capex outlook from roughly $49B to $52B-$56B. Its chairman continued to stress that investment was “very prudent,” saying the company had spent the previous 3-4 months surveying customers and customers’ customers to verify whether AI demand was real.
• The results made him “very satisfied.” 庄明浩’s inference is that the most downstream, rational, and even extremely calm company in the supply chain has offered a clearer demand outlook, seemingly enough to give the industry trend a stronger confirmation.
• But 七一 and 庄明浩 retained an element of irony: application gross margins are not credible, models are telling stories, and cloud and Nvidia are also being questioned, so the market can only keep believing down to the foundry and then the raw-material suppliers. Supply-chain confirmation does not eliminate valuation risk in financial assets.

29. The AI Industry Is Still Moving Higher, but Financial Narratives Make Stocks Far More Volatile than the Technology Curve

• 庄明浩 refuses to summarize a complex system as “this time is the same” or “this time is different.” Both sides can find 100 factors, and the weighting can still be rewritten suddenly by a single unexpected event.
• The more basic data continues to improve: user scale, retention, payment rates, per-user token consumption, and expansion into new application scenarios. Looking only at technology and industry, AI remains on a positive trajectory.
• Once financial narratives, leverage, and long-term valuations are layered on top, the price curve will not be as stable as commercial progress. 七一’s summary is blunt: “The technology will definitely iterate, and AI tools will keep getting better, but the stock price does not necessarily have to rise.”

30. The Best AI Strategy for Ordinary People Is to Outsource Repetition and Keep Judgment for Yourself

• 庄明浩 offers a simple threshold: if something at work or in daily life is repeated “more than 5 times a week,” try having AI do it, even if it replaces only one step.
• For information workers, AI is already “limitlessly strong” at summarizing, processing, translating, and visualizing. Even if machine translation of a long English document is imperfect, it can be turned into an infographic or PPT to provide a rapid grasp of the whole.
• Investors can ask AI to play the other side of the trade: “If you were my counterparty today, how would you think about it?” It may not be right, but it can consistently surface overlooked factors and is easier to act on than a generic “brainstorm.”
• AI search has made finding fundamental information easy. The real difficulty is convergence. Investment style, risk appetite, and discipline come from repeated gains and losses; that personal framework cannot be outsourced to a model.

31. No One Knows Yet Whether AI Will Open or Close Children’s Creativity

• 七一 asked whether a child who can generate a PPT or creative work with one sentence is opening creativity or shutting it down. 庄明浩 answered candidly: “No one knows.” People naturally take shortcuts; the search-engine and electronic-device eras already showed that, and AI makes it unavoidable.
• Parents can provide a trustworthy environment in which children can experience, participate, and feel. He mentioned hackathons, AI camps, and the idea of having AI video practitioners, documentary directors, and teachers help children complete a film or documentary in 7 days.
• 七一 recalled that his daughter, a senior kindergarten student, can already use AI by voice. When an assignment asked her to use a simple drawing to represent an abstract word such as “lose,” she seemed to draw several scattered cats to express it, solving a problem even her parents could not draw.
• 七一 closed by linking the Spring Festival to the cycle of AI narratives: 2023 was the first A-share AI-concept wave, 2024 was Sora, and 2025 was DeepSeek. By the 2026 Spring Festival, “the atmosphere has been built up to this point—something big is bound to happen.” That remains an expectation and forecast with uncertainty attached.