Vol.86 One Generation of Technology, Two Systems: 181 Pages Reconstructing the AI Industry in 2025
Summary
2025 was not about China and the US using two generations of technology, but about the same frontier-model constraints being amplified into different paths by two industrial and capital systems. The US deepened its vertical stack across chips, data centers, cloud, models and applications, with the giants locked in mutual dependence; China used open-source models, multimodal applications, internet distribution and policy-backed hard-tech financing to pursue at close range. 庄明浩’s summary: “One generation of technology, two systems.”
The center of gravity in model R&D has shifted from pretraining to verifiable reinforcement learning, with the 2026 contest moving further toward Agents, autonomous learning and AI for Science. OpenRouter data shows reasoning models already account for more than 50% of usage; RL environments, task design, outcome scoring and expert feedback are becoming new infrastructure for model companies, while Mercor took ARR from $1M to $500M in 17 months. “Every company’s choice on this actually reflects its taste.”
Language, coding, multimodality and world models are converging onto one “main table,” but the US big three are pursuing radically different commercial entry points. Google is building a full-stack advantage around Gemini, image, video, NotebookLM and TPU; OpenAI is betting on To C, user memory and “adults want it all”; Anthropic is betting on “coding plus everything,” with Claude accounting for more than 60% of coding calls on OpenRouter and forecast to reach positive cash flow earlier.
Revenue is arriving first where technology has crossed a threshold, not in every product carrying the Agent label. AI coding formed an approximately $4B market in 2025, while healthcare represented roughly $3.5B of incremental opportunity; OpenAI data shows coding usage up 11x and healthcare up 8x, after which both ChatGPT and Claude broke Health out as a standalone category. Independent coding companies are simultaneously being squeezed by first-party tools from model vendors, showing that a market can be real without startups capturing all the value.
The biggest risk in the US AI bull market has shifted from whether demand is real to who finances $2T of data-center investment, how the assets depreciate and whether the middle layers can earn a profit. Large data centers already in operation total roughly 2.26GW, versus 35GW under construction or committed; Meta’s 2026 Capex of about $116B is approaching its free cash flow, while Oracle’s investment requirements for a five-year, $300B OpenAI order have forced a market rethink. a16z’s blunt math: by 2030, AI revenue may need to reach $800B-$1.1T to support the planned investment.
One year after DeepSeek R1, China has made open source the main battlefield of the US-China model race. By mid-2025, Chinese open-source models had surpassed US models in Hugging Face downloads; Qwen later overtook Llama, and the leading open-source slots were almost entirely occupied by DeepSeek, Qwen, Kimi, Zhipu and MiniMax. Epoch AI puts the historical average gap at 7 months, while 庄明浩 believes the current gap is smaller. “After V4 ships, most likely everything will have to be rewritten”—the market’s way of showing DeepSeek “respect.”
China’s application-layer battle is far from settled by billion-user figures, and the 2026 Lunar New Year will test retention, ecosystem orchestration and customer-acquisition efficiency at once. Doubao is rumored to have passed 100M DAU, Qianwen announced more than 100M combined MAU across its app and web products, and Baidu claimed 200M monthly active users for Wenxin in its PR language; Yuanbao committed RMB1B in red envelopes, Baidu RMB500M, Doubao and Volcano Engine sponsored the Spring Festival Gala, and Ant moved its Five Blessings campaign to A-Fu. Old conclusions such as “Chatbots are not entry points” and “paid acquisition is meaningless” are being reopened.
Both capital systems are concentrating money in the leaders, but the US is underwriting trillion-dollar platforms while China prices domestic GPUs, commercial space and frontier models at roughly 1%-2% of US leaders. AI accounted for about 48% of US financing events in 2025 and more than 60% of large rounds; in China, roughly 1.43% of projects raising at least RMB1B absorbed more than 40% of capital, while the dollar share fell from about 10% in 2021 to 2.5%. The ultimate risk remains the question 庄明浩 keeps returning to: “Wishful thinking, or the natural consequence of the facts?”
Deep dive
1. 2025 Was Not an Annual Update but the Simultaneous Unfolding of Two AI Systems
庄明浩 used 181 pages of slides to review 2025 because the industry “basically turns over every three months”; only three months after his early-November summary, the narratives around models, products, capital and US-China competition already needed to be rewritten.
“One generation of technology, two systems” means that China and the US face the same frontier-model constraints and are both adding compute, capital and model investment, but once the technology enters industry, organizations and markets, infrastructure, commercial entry points and capital-exit paths begin to diverge.
The US story runs from models to products, products to use cases, and industries to capital; China’s runs through technological breakthroughs, super-app distribution and the capital narrative. This is not a simple ranking exercise, but a way to observe how different systems absorb the same technology.
2. AI’s Capability Frontier Is Expanding Irregularly; One Failed Task Cannot Disprove the Whole
庄明浩 borrowed an image of “human capabilities as a circle”: AI initially had scattered protrusions, then exceeded human limits in some capabilities while remaining far behind in others. “AI still can’t do X” and “AI already exceeds humans at Y” can therefore both be true.
He believes we are roughly in stages 3-4: some capabilities have surpassed humans and may continue expanding across most tasks, but obvious gaps remain. AGI would be the point at which this irregular shape fully covers the circle of human capabilities.
3. Reasoning Models Turned 2025 into a Model Year That Iterated by the Week
From DeepSeek R1 to GPT-4.5, GPT-5, Claude 4/4.1, Gemini 2.5/3, Sora 2 and Suno V5, followed by GPT-5.1 and GPT-5.2 at year-end, frontier-model launches approached a weekly cadence. On the video and image side, there were also Veo 3, Nano Banana, Kling and Hailuo.
The first keyword was the reasoning-model explosion. The path began with OpenAI’s o-series launch in September 2024, was amplified by DeepSeek R1, and became standard equipment for leading vendors in 2025. By OpenRouter’s measure, more than 50% of models are now reasoning models.
The technical chain has expanded from “pretraining-human alignment” to pretraining, post-training and reinforcement learning. The corresponding organization now includes pretraining, post-training, inference services and hardware teams; the R&D paradigm directly shapes the company’s boundaries.
4. Verifiable Reinforcement Learning Is Becoming New Heavy Infrastructure for Model Companies
庄明浩 believes pretraining remains important, but 70%-80% of the usable opportunity may already have been “mined.” As marginal returns decline, more compute and talent are moving into post-training, especially verifiable reinforcement learning.
The hard work is not simply “doing RL,” but building environments, designing tasks, setting scores, validating outcomes and closing the engineering loop. Coding and mathematics lead because their answers are easier to verify; vertical SaaS often has to build environments from its own business data.
Suppliers have split into model companies, vertical-product companies and a new generation of data-service providers. After the Meta-Scale AI deal reshaped the market, companies such as Mercor absorbed the demand; 庄明浩 describes them as the AWS of the LLM era. Mercor grew ARR from $1M to $500M in 17 months.
These “data laborers of the new era” hire not only ordinary annotators but also lawyers, finance professionals, industry specialists and scientists. 庄明浩 cited Surge AI CEO Edwin Chen: “Post-training optimization is really an art.” Which experts to choose and which answers to reward are themselves matters of a model company’s “taste.”
5. Agent Task Duration Is Replacing Single-Task Benchmarks as AGI’s Practical Yardstick
Sequoia’s “2026: This Is AGI” summarizes the path as follows: pretraining solves knowledge, post-training solves reasoning, and Agents now solve sustained execution. The key measure is not merely whether an Agent answers one question correctly, but how long it can autonomously complete a complex task.
He kept one recruiting example: AI can find an RL conference talk, identify the speaker, analyze the person’s social-media influence and likelihood of leaving, shortlist candidates and send emails. A traditional headhunter might take several days; an Agent could perhaps finish in 30 minutes. “The ability to solve problems” is therefore closer to AGI than an abstract definition.
OpenAI’s L1-L5 moves from Chatbot and reasoning to Agent, innovator and organizer. If autonomous learning becomes real, L4—the “innovator”—could come within reach. That also explains the rise of AI for Science and the emergence of dedicated research products in early 2026.
6. New Labs Are Betting on the Next Research Paradigm, Not Near-Term Revenue
Silicon Valley’s new consensus is “autonomous learning.” Ilya Sutskever’s call to “open a new era of research” points in the same direction: beyond current scaling, post-training and Agents, there may be a new architecture, learning mechanism or research paradigm.
A wave of new labs emerged over the past six months, many founded by researchers from leading model companies, and raised enormous sums before proving commercialization. Investors are betting that one of them may discover the “next-generation paradigm.”
庄明浩 offers no definitive answer. Companies founded by Ilya, Mira Murati and others could make breakthroughs in foundational research, or stumble as teams churn; at least in 2026, they should not be expected to deliver commercial results quickly.
7. Language, Coding and Multimodality May Ultimately Share One Main Table
He once treated language, coding and multimodality as three separate tables, but changed his view in the second half of 2025. The performance gap among the top 10 language models may be only about 3%, versus more than 30% between first place and sixth or seventh in video, and more than 70% in Agents. Resources naturally flow toward the areas with greater marginal returns.
More important evidence comes from product experience. NotebookLM does not merely know the plots of Naruto or Eggy Party; it can map their world-building to AI principles and visualize the result. It can also understand the story of The Three Little Pigs from photos and video of a pop-up book, including the three-dimensional effect of “wooden sheets or sticks attached to the book.”
This convinced 庄明浩 that text, images, video, voice and coding are not parallel capabilities but different outputs of the same intelligent system. “Once it breaks through a certain line, it looks like it just takes off—and keeps accelerating.”
8. World Models Are Splitting into Real-Time Interaction and Physical Simulation
The real-time interaction camp is pursuing something close to infinitely long video: users issue instructions during generation, and seconds later the scene continues in the new direction. 庄明浩 cites PixVerse’s recent world-model demo, which he sees as closer to interactive content and gaming.
The physical-simulation camp, represented by World Labs and DeepMind, aims to make generated spaces obey real-world physics. World Labs’ cooperation with Chinese robotics companies is designed to turn these spaces into virtual laboratories, providing robotics and autonomous driving with data that is scarce and expensive to collect in the real world.
He still stresses that the field is “very far” from the ideal state. But if simulation quality continues to improve, embodied-intelligence training may at least partially reduce its dependence on real-world data collection.
9. The US Big Three Have Shifted from Same-Table Competition to Differentiated Bets
Google’s chip is multimodality plus the full stack: Gemini, images, video, NotebookLM, Chrome distribution and TPU reinforce one another. 庄明浩 calls its multimodal lead “in a different league,” which is why the market began discussing at the end of 2025 whether TPU could erode Nvidia’s share.
OpenAI is betting on To C and user memory. ChatGPT is trying to understand users’ personalities, preferences and professions, then adjust its answers accordingly. It believes that an experience capable of “understanding you” can improve retention before extending into browsers, communities, e-commerce, advertising, infrastructure and hardware.
Anthropic’s bet is the most concentrated: “To B coding.” Claude Code, Skills and Cowork extend outward from programming into general knowledge work, giving the company a clearer enterprise-market spine than OpenAI.
10. The Chatbot War Did Not End in 2024; Gemini Is Rewriting the Share Map
As of mid-January 2026, the traffic chart cited by 庄明浩 showed ChatGPT, DeepSeek, Claude and Copilot weakening while Gemini and Grok grew. Gemini had moved from low-single-digit share to roughly the teens or low twenties.
ChatGPT traffic still grew about 50% in 2025, but Gemini grew about 560% off a low base. After Gemini 3 launched, its scale briefly reached roughly 38%-40% of ChatGPT’s. The old conclusion that “ChatGPT has already won” has therefore weakened materially.
OpenAI “hit the red alert” after Gemini 3, not only because of benchmark pressure but because its consumer entry point could be eroded. 庄明浩 expects the 2026 competition between the two in Chatbots to become even more intense.
11. OpenAI’s “Adults Want It All” and Anthropic’s “Coding Plus Everything” Create an Operating Split
OpenAI is simultaneously building consumer applications, browsers, communities, e-commerce, advertising, infra and potential hardware. 庄明浩 summarizes this as “adults want it all,” but even with a valuation approaching $1T and a team of several thousand, it remains a huge organizational challenge for a startup.
Anthropic is expanding outward from Claude Code. By the figures he cited, Claude Code may be the fastest application in history to reach $1B in ARR, while Claude already accounts for more than 60% of coding calls on OpenRouter.
The market also circulates claims that more than 90% of Claude Code’s code is written by Claude and that Cowork’s code is completed by Claude Code. 庄明浩 sees this as a sign that coding is approaching “self-evolution,” while retaining the qualification that “humans may have participated a little.”
The financial paths also differ. Forecasts he cited suggest Anthropic could reach positive cash flow first, in 2027, while OpenAI may not do so until 2030. The IPO race, however, could force OpenAI to pull forward its original 2027 timeline to Q4 2026.
12. Technology Crossed a Threshold in Coding First, Then Made Healthcare the Second-Largest Incremental Market
AI coding “created” an approximately $4B market in 2025. Cursor, Replit, Claude Code and Codex kept setting new speed records for reaching $100M in ARR, compressing the path from two years and 18 months to nine months or even six.
Traffic to independent companies broadly declined for a time from Q3 as first-party coding tools from model vendors surged. It recovered in January 2026, showing that demand is real, while leaving unresolved whether the model vendor or the independent product ultimately captures the value.
Healthcare represents roughly $3.5B of potential incremental AI-market opportunity. OpenAI usage data shows coding up 11x and healthcare up 8x. Once a channel becomes mature enough, it gets separated out like autos and real estate on a portal website; ChatGPT and Claude have both begun breaking out Health as a standalone category.
庄明浩’s 2026 extrapolation includes enterprise-Agent implementation, AI-native vertical applications, Web Coding and deeper multimodal enterprise penetration. Data access, privacy, payments and security, however, will move from secondary considerations to prerequisites for deployment.
13. The US AI Industry Has Become a Complete Free-for-All
Chips, data centers, cloud, models, enterprise software and consumer applications are layered on top of one another. Google, Microsoft, Amazon, OpenAI, Meta, xAI, Nvidia, Apple and Anthropic are both one another’s customers and cross-investors through cash, servers, model agreements and chip contracts.
In this structure, Google is closest to a full-stack winner: its in-house chips, cloud, infrastructure, models, consumer entry points and distribution are all strong. Microsoft and AWS have stronger cloud positions but weaknesses in chips or models; OpenAI and Salesforce are strong in models and applications but weaker underneath.
Software has entered its fourth wave. ChatGPT, Perplexity, Midjourney and Manus represent AI-native products; Figma, Notion and Canva belong to the previous generation of cloud software, but their existing scale and AI uplift have opened new windows for growth and public listings.
14. AI Has Pushed US Equities into a $20T-Scale Industrial Wave
庄明浩 cites the “three-times rule”: the combined market caps of the leaders in PCs, the internet, mobile internet and cloud were approximately 3x the prior generation each time. Cloud is worth about $6.5T, implying a potential AI aggregate of roughly $20T; the existing giants already approach that level in aggregate.
Roughly three years after ChatGPT’s launch, AI companies in the S&P 500 had gained about 185% in market value, versus only about 28% for non-AI companies. As of January 26, 2026, the S&P 500’s total market cap was approximately $62.3T and still setting new highs.
The 10 largest US companies by market cap gained an average 24% in 2025; the Magnificent Seven gained about 22%, versus roughly 17% for the S&P 500. Google rose 62.4%, Broadcom 51%, Nvidia 37.8% and Tesla 38.5%; even Apple, the weakest among the leaders, gained about 6%.
The 10 largest companies together were worth roughly $25T, or about 40% of the S&P 500. Optimists point to an average PE of about 30x, far below the 80x-100x seen during the dot-com bubble; concentration risk has not disappeared.
15. The Primary Market Recovered, but More Capital Was Compressed into Fewer AI Companies
US AI projects accounted for about 48% of financing events in 2025 and absorbed more than 60% of large-round capital. Series A, B and C rounds and billion-dollar rounds all improved; only seed rounds under the traditional definition declined, partly because round sizes themselves have become “inflated.”
OpenAI raised about $41B last year, Anthropic about $32B, Scale AI about $14.8B and xAI about $12.8B. Their corresponding valuations were roughly $500B, $350B and $230B; based on revenue, xAI’s PS multiple reached approximately 230x.
AI companies take roughly two years on average to become unicorns, and more than 100 have been added over the past three years. IPO exits remain far weaker than in 2021; Figma, Cove and others offer only limited improvement, while the unicorn inventory continues to build.
M&A has therefore become an alternative exit. 庄明浩 puts Nvidia-Groq, Meta-Manus, Google-Windsurf and Meta-Scale AI among “hollowing-out acquisitions”: pay heavily to obtain teams or core capabilities while avoiding the antitrust resistance of a full acquisition.
16. Data Centers Are Moving from Megawatts to Gigawatts, While Storage Captures the First Scarcity Premium
The three major cloud providers grew quarterly revenue by more than 20% on average in 2025. Microsoft and Google were mostly above 30%; AWS was briefly in the low teens before returning to roughly 20%. New clouds such as CoreWeave and Nebius are using higher leverage to absorb incremental demand.
Extending cloud and new-cloud construction plans through 2030, the estimates cited by 庄明浩 imply cloud revenue of $300B-$400B. The US data-center network under development is “more insane than the railway network of its time.”
Large AI data centers already operating total roughly 2.26GW, versus 35GW under construction or committed. New AI data-center demand is beginning to exceed traditional data-center demand and is propagating through networking, power, energy, cooling and storage.
Among the best-performing S&P 500 companies in 2025, SanDisk, Western Digital, Micron and Seagate ranked near the top. Storage supply is highly concentrated, and new fabs and production lines can take two or three years to go from construction to shipments. If demand continues to be revised upward during that period, a shortage could shift from a short cycle to a long-term trend.
17. The Real Risk in $2T of Capex Is Financing, Depreciation and Cash-Flow Mismatch
The market consensus for data-center investment over the next 4-5 years is roughly $2T. The technology giants themselves may cover only a little more than $1T; the rest must come from depreciation recovery, private capital, credit and bonds.
Meta raised 2026 Capex to approximately $116B, approaching its free cash flow. Amazon, Meta, Google and Oracle, despite being leaders, began issuing enormous amounts of debt from Q3-Q4 2025. The market is therefore more concerned about the debt side than the equity side.
GPUs may be depreciated over 6-7 years for accounting purposes, but the rapid iteration of A, H and B series means their economic lives may be only 3-5 years. The larger the investment, the harder it will be for demand strength to conceal the future pressure from depreciation and amortization on the income statement.
Oracle’s trajectory is the clearest example. Its stock surged after confirming a five-year, $300B OpenAI order, then fell steadily back to the level before the contract announcement once the market realized that the cash flow required for construction in 2026-2028 was insufficient.
18. A Smooth Demand Chain Does Not Mean Every Layer Captures Value
Users run AI apps, apps call models, models run on clouds and data centers, data centers buy Nvidia chips, and the chips are manufactured by TSMC. Token usage multiplies along the chain; demand itself is not the most questionable link.
Nvidia sells scarce compute at gross margins above 70%, while TSMC benefits steadily from its foundry position. Cloud providers still have revenue growth to support them. The hardest layer to explain is models and applications: inference is expensive, gross margins are thin, and the incremental revenue from a new model may not even cover training, infrastructure, headcount and promotion.
a16z’s calculation sets a harsh hurdle. By 2030, if the planned investment is to earn roughly a 5% return, AI companies need annual revenue of approximately $800B; at roughly 15%, they need $1.1T. OpenAI currently has about $13B in revenue, leaving a gap of an entirely different order of magnitude.
TSMC raised next year’s Capex to approximately $52B-$56B and said, “If we do not treat this prudently, it will be a disaster for TSMC.” After spending 3-4 months speaking with customers and customers’ customers, its chairman concluded that demand was real and that the discussions had “made him extremely satisfied.” 庄明浩 still asks: “Are we halfway up the mountain, or already at the summit? I don’t know either.”
19. DeepSeek R1’s Biggest Impact One Year Later Is Changing How the World Accesses Model Capability
R1 launched on January 20, 2025, and had just reached its first anniversary at the time of recording. It not only shook US VC, but also offered less-developed economies a relatively affordable path to important model capabilities. 庄明浩 believes its global impact “is still underestimated to this day.”
DeepSeek’s path ran from V3, R1, V3-0324 and R1-0528 to V3.1, V3.2, OCR, a mathematics model and OCR 2. In the first half it “fought relentlessly on reasoning”; midway through the year it advanced the V and R series in parallel and gradually fused them, while later papers increasingly pointed toward architectural breakthroughs.
The market’s anticipated R2 may be built on V4, or V4 itself may unify reasoning and Agent capabilities. 庄明浩 insists on completing his summary before V4 because “after V4 ships, most likely everything will have to be rewritten.” That is the market’s way of showing DeepSeek “respect.”
20. Chinese Open-Source Models Have Gone from Catch-Up Players to a Global Core Supply Base
In June and July 2025, Chinese open-source models surpassed US models in Hugging Face downloads. By around Q3, 庄明浩 believed Chinese models had also begun to lead the global open-source field, while the US had never managed to reopen a decisive gap.
Epoch AI research says Chinese models have lagged the US frontier by roughly 7 months on average since 2023. Since that figure averages across 2-3 years, he infers that the current gap should be less than 7 months, without exaggerating it into a comprehensive Chinese lead.
Qwen overtook Llama in downloads and may have exceeded the combined total of all other open-source models in December alone. The latest leaderboard’s leading open-source slots are concentrated among Chinese teams including DeepSeek, Qwen, Kimi, Zhipu, MiniMax and Xiaomi.
Open source is no longer a specifically Chinese path. OpenAI, Google, Nvidia, Microsoft, IBM and several European model companies are all active. The real competition is now open and closed source in parallel, with a fresh “battle of the heroes” inside each camp.
21. Chinese Model Companies’ 2026 Roadmaps Closely Mirror the US
Zhipu says it will continue exploring the upper bound of intelligence. Known directions include scaling, foundation models and coding Agents; unknown directions include entirely new architectures, extremely long context, knowledge compression, memory mechanisms, continual learning, multimodality, AI for Science and robotics.
Kimi is focusing on token efficiency and long-context memory. 庄明浩 calls it one of the few companies able to be “self-indulgent” in foundational research: it is not yet public and does not have to answer quarterly financial questions, while already having raised enough capital.
Coding and Agents are now common investment priorities. GLM-4.7 improves open-source coding, while Kimi K2.5, MiniMax, DeepSeek and Qwen are advancing in parallel. Qwen’s launch title, “From Question to Action,” captures the move from language generation to execution.
Strip away the logos and Doubao, Yuanbao, Qianwen, Kimi, Manus, Coze and MiniMax Agents are increasingly difficult to distinguish. Model vendors are therefore packaging experience and workflows. The industry joke evolved from “this pen was made by AI,” through MCP, into “this pen was made by a Skill.”
22. Billion-Level Daily Active Users Have Not Ended China’s Chatbot War; They Have Raised the Entry Bar
By the end of 2024, the market was already saying “Doubao won.” DeepSeek R1 briefly reached the top ranks in 2025, then fell back because it did not accept mass To C traffic, allowing Doubao to retake the lead. By year-end, Doubao was rumored to have passed 100M DAU and showed a ChatGPT-like recovery in retention.
Alibaba announced more than 100M combined MAU across the Qianwen app and web, while Ant promoted Lingguang and A-Fu. Baidu claimed in its PR language that Wenxin had reached 200M monthly active users. “Whether you believe it is another matter,” but the figures at least show that every player remains willing to invest in the entry point.
The 2026 Lunar New Year became a new stress test. Yuanbao committed RMB1B in red envelopes, Baidu RMB500M, Doubao and Volcano Engine became Spring Festival Gala sponsors, Qianwen and other products appeared to have no budget ceiling, and Alipay’s Five Blessings campaign moved to A-Fu.
庄明浩 therefore revisited the old narrative. Models may not consume every application, user data may have a flywheel after all, Chatbots may count as entry points, and paid acquisition may not be meaningless. The key is whether users stay after acquisition and whether the ecosystem can be mobilized.
23. China’s Multimodal Sector Is a Shared Advantage for Models and Applications, Not a Peripheral Boom
Chinese teams occupy roughly 11 of the top 30 image-model slots and about 19 of the top 30 video-model slots, with Kuaishou, ByteDance, MiniMax and PixVerse appearing in force. Around 15 of the top 20 overseas applications are also related to images or video.
The application layer is even more vibrant: Jimeng, Kling, PixVerse, Vidu, Hailuo, Meitu and CapCut are competing across multiple layers. 庄明浩 believes that if multimodality is treated as a separate table, China has the edge in model supply, application iteration and overseas expansion.
Yet every company is ultimately moving toward a unified architecture. Qwen lists “unified multimodal understanding and generation” as a goal; Zhipu, which had previously done relatively little in multimodality, released its first image model, while Kimi K2.5 natively supports vision, text, thinking mode and Agent tasks.
The 2026 competition among ByteDance, Alibaba and Tencent will therefore extend beyond Chatbot share for Doubao, Qianwen and Yuanbao into complete multimodal systems spanning images, video, cloud and distribution.
24. AI Hardware, Companionship and Games Are High-Risk Samples of China’s Application Boom
AI hardware was one of the biggest consensuses in Chinese VC in 2025, covering toys, recording devices, cameras, learning machines, musical instruments and humanoid robots. Nearly 40 Chinese AI-toy companies exhibited at CES alone; AI toys became the “consensus within the consensus.”
Social companionship appears stagnant in the US. 庄明浩 cites Tolan, whose MAU was only about 200,000 at year-end; Character.AI still held more than 95% share, but the market barely grew. China, by contrast, saw denser experimentation with products including Catbox, Xingye, Nieta, Dream Dimension and Second Me.
AI plus otome games was called by one guest “the most pragmatic and best-funded direction of 2025.” The combination of early Agents and the mature payment model of romance games produced projects such as EVE, Xingmian and Infinite Valley. 庄明浩’s reservation is pointed: “Many founders lack reverence for games.”
Borderline products proved willingness to pay while exposing regulatory boundaries. In the first related domestic case, the product had roughly 110,000 registered users and 24,000 paying users, generated RMB3.63M in cumulative membership fees, and its developer received a four-year prison sentence. Whether to go overseas or stay domestic, and where to put the guardrails, cannot be answered by growth data.
25. The New BAT Are Reassembling Cloud, Models and Entry Points into Single Companies
ByteDance’s advantage is the coordination between Doubao and Volcano Engine. Monthly average daily Token volume at Volcano Engine rose from roughly 4T at the end of the prior year to 12.7T in April, 16.74T in May, 30T in September and 50T at year-end; 庄明浩 summarizes this as 11x growth over the past year. Headphones, the Nubia Doubao phone and Anker’s “recording bean” show that its hardware ambitions remain alive.
Alibaba’s story is “Make Alibaba Great Again.” Alibaba Cloud, Qwen, T-Head and e-commerce services form a more complete vertical chain. The market has also heard that Qwen was trained using Alibaba’s in-house chips, a logic similar to Google supporting Gemini with TPU.
Two presentation slides show the reversal in Alibaba’s organizational logic. In 2023: “All Alibaba products will connect to large models in the future.” In 2026: Taobao, Alipay, Fliggy, Amap, Youku, Damai, Cainiao, Alibaba Health, 1688 and Freshippo “team up to serve the Qianwen app.”
Tencent Cloud reportedly achieved its first profitable quarter in Q4 2025, according to an article by Leiphone, while the Hunyuan team also reorganized. Tencent is not yet playing as continuously as ByteDance or Alibaba, but 庄明浩’s view is simple: “Tencent will not stay still.”
26. Ant, Tencent and Baidu Are Betting Separately on Health, Social and the Revaluation of Existing Assets
Ant deserves separate treatment. Its CEO calls Lingguang the exploratory logic and A-Fu the “must-win logic.” Years of accumulated healthcare payments, services and offline operations may be the groundwork for A-Fu; Ant Investment is also regarded as one of the most aggressive investors in AI applications.
Tencent’s Yuanbao Party puts AI inside group chats. It can adjust the mood, execute tasks, recommend films and music, organize meetings, and let users watch Tencent Video or listen to QQ Music together. The experience resembles Clubhouse while continuing Tencent’s earlier experiments in gaming and social spaces.
Red envelopes, group chats, music and video appearing together means Tencent is repackaging its old ecosystem capabilities into an AI entry point. Xinhua called it “Tencent entering the arena”; 庄明浩 sees it as a replay, in the AI era, of the Spring Festival red-envelope war more than a decade ago.
Baidu released Wenxin 5.0, advanced the spin-off of Kunlunxin and reorganized Baidu Wenku and Baidu Netdisk into a new business group. With AI added, old businesses such as Wenku have regained growth. Baidu’s stock more than doubled over the past 52 weeks; its market cap was about $55.4B at recording, meaning AI has at least delivered one capital-market revaluation.
27. China’s Hard Tech Is Priced at 1%-2% of US Leaders; Policy and Scarcity Substitute for the Income Statement
GPUs, commercial space, embodied intelligence and frontier models are synchronized hot sectors in China and the US. China has Moore Threads, Biren, MetaX and Enflame; commercial-space companies including i-Space, LandSpace, Galactic Energy and Space Pioneer; and robotics companies such as Unitree, Galbot and Agibot.
庄明浩 proposes a “1%-2% valuation principle.” Nvidia at roughly $4.5T implies $45B-$90B for Chinese GPUs; SpaceX at an expected $1.5T implies $15B-$30B for Chinese commercial space; OpenAI at roughly $830B implies RMB55B-RMB110B for Chinese frontier-model companies.
The principle does not depend on current revenue or profit. It depends on policy opening the gate and the China-US gap potentially narrowing from two orders of magnitude to one. Borrowing an A-share-style headline, 庄明浩 jokes: “No one understands large models, GEO, GPUs, storage, chips and optical communications like the A-share market.”
28. China’s Primary Market Recovered, but Funding Sources, Concentration and Exit Venues Have All Changed
By IT Juzi’s measure, Chinese primary-market financing recovered sharply in 2025 and may even have exceeded 2021. The dollar share, however, fell from roughly 10% in 2021 to 2.5%; 庄明浩 says the decline may exceed 80%, with dollar capital clearly retreating.
Artificial intelligence, healthcare, integrated circuits, robotics, new materials, new energy, advanced equipment, the low-altitude economy, energy storage and commercial space were shared hot sectors. Approximately 1.43% of projects raising at least RMB1B absorbed more than 40% of financing, showing the same concentration logic as the US.
IPO volume recovered from the 2024 trough to roughly 120 companies. In 2021, about 84% chose the A-share market and 11.4% Hong Kong; by 2025, the split was 41.5% A shares and 41.9% Hong Kong, with Hong Kong overtaking for the first time. The queue was rumored to contain 300-400 companies.
If only hard-tech exits are counted, China’s 2025 exit value even exceeded the US. But the Hong Kong exchange’s “best of the best” requirement also signals that the 2026 window may tighten; the number of companies waiting cannot be treated as the number certain to exit.
29. The Most Likely Endgame Is Not One-Sided Victory but Two Standards Formed by Intersecting Compute and Capability
庄明浩 closes the US-China competition with a four-quadrant framework. If the US leads in both compute and model applications, it wins the AI war; if China leads on both ends, it creates a distinctly Chinese AI era. Neither extreme is necessarily the most likely.
A more realistic split may be that the US is stronger in model capability and adoption but constrained by power, energy and construction costs, leading it to build one standard with its allies. Chinese models may be somewhat weaker but serve another group of countries through lower compute costs, deployment efficiency and customized solutions.
This is not a static conclusion. Open source may overtake closed source, applications may regain value, and model companies may fail to cover their costs. 庄明浩 ends with “the AI revolution is not yet successful” and reminds himself to continue as the “Sima Qian of AI,” because the next model launch may invalidate today’s judgment.