150: Year-End AI Review—Models, Markets, Applications, Meaning | Solo
Summary
The most important change on the model side in 2025 was the convergence of reasoning, coding, multimodality, tool use and memory into a capability stack that can support Agents. DeepSeek R1 pushed Test-Time Compute onto the global stage through full open sourcing, a detailed technical report and the headline that its final training run cost just $5.57M. It also showed that young, small teams can produce frontier results at relatively low cost; Claude Code, meanwhile, demonstrated that “coding is not only the destination but also the means”—programming gives models hands and feet in the digital world.
Model competition is shifting from a single-algorithm race to the co-design of chips, systems, algorithms, products and organizations. Google’s vertical integration of TPU, JAX, AI infra, cloud, Gemini and applications; DeepSeek’s joint development of algorithms and infra; Alibaba’s integration of Qwen, Alibaba Cloud and T-Head; and Tencent’s placement of AI infra under 姚舜宇 all point to the same moat. DeepSeek once used more than 1,800 cards to handle over 600B input tokens and nearly 170B output tokens in a day, calculating an 84.5% gross margin. The controversy itself underscored the extremity of its systems optimization.
Agent applications are opening opportunities across general automation, vertical workflows and infrastructure, but control of the user entry point still sits with platforms. Claude Cowork and OpenClaw are bringing general agents within reach of ordinary users, while vertical Agents are entering law, design, film and other fields through proprietary data and workflows, shifting from seat-based pricing toward “paying for outcomes.” The restrictions placed on the Doubao phone assistant by WeChat, Meituan and others expose a three-way struggle among handset makers, super apps and AI OS providers. Whether Agents can call services, who captures advertising value and whether platforms remain necessary are all unresolved.
China’s tech giants are now betting simultaneously on models, products, traffic and ecosystems, with ByteDance temporarily holding the most complete consumer flywheel. Doubao became China’s first AI product with more than 100M DAU, while its overseas version Dola also surpassed 10M DAU. ByteDance has built a “one superpower, many strong players” system through model-product collaboration and Douyin distribution, but its biggest traffic product remains free; revenue mainly comes from tools such as TRAE and Coze. Alibaba is making a major push into life services through Qwen, taking daily marketing spend to RMB15M, while Tencent’s full integration of DeepSeek drove Yuanbao’s DAU up tenfold in a week. The Spring Festival also brings Doubao’s CCTV gala campaign, Yuanbao’s RMB1B red-packet campaign, Qwen’s RMB3B campaign and DeepSeek V4 as variables.
Valuations for foundation-model startups have risen sharply, but revenue and spending capacity remain two to three orders of magnitude below leading US companies. After listing, Zhipu and MiniMax reached market capitalizations of RMB79.4B and RMB128.4B, while their reported revenue during the relevant periods was roughly $27M and more than $53M. OpenAI says its 2025 revenue was well above $13B; Anthropic was reportedly at $4.5B. 程曼祺’s summary was that an IPO “is not a reward for the winner, but the starting drumbeat for the next round of competition”—and even a “safety net for staying alive.”
Application companies have broader commercial validation than foundation-model companies, but only about 40–70 global AI-native companies have truly crossed $50M in ARR. Cursor, Perplexity, Midjourney, ElevenLabs and Harvey cluster around coding, content creation, law, education and healthcare. Among teams with Chinese roots, Manus was acquired by Meta for roughly $2B, Genspark reportedly reached $50M in ARR, and OpenArt generated about $70M with roughly 20 people. Application teams can remain small and focused, get acquired or move down the stack toward models; they do not all need to play the foundation-model “Go Big or Go Home” game.
Embodied intelligence has attracted capital pricing far ahead of current revenue and deployment, and the 2026 listing wave will test how far into the future the market is willing to pay. Chinese embodied-AI and robotics financing reached RMB73.5B in 2025, versus RMB18.2B for companies still training large models. After Agibot acquired control of Shangwei, the latter’s market cap briefly rose from just over RMB3B to RMB69B, though that did not amount to a backdoor listing. The sector remains stuck on data, models and embodiments: there is no consensus on whether simulation is a low-cost path or “a giant pit”; Optimus’ dexterous hand lasts only about six weeks and costs more than $6,000 per hand; and scaled deployment in factories, stores and homes remains unproven.
AI hardware could produce an entry-level device in glasses, or diversify over time around narrower needs such as recording, health, companionship and outdoor use. The value of glasses lies in being hands-off, always-on and able to capture real-world context unavailable to the major platforms, but displays, weight, battery life and heat remain unresolved. Plaud’s more than 1M units sold through call recording and AI summaries show that a clear use case matters more than treating “hardware as a vessel for models.” The harder question ultimately concerns people: AI job creation and white-collar layoffs are happening simultaneously. Repetitive work can be “AI-ified,” but human meaning may not transfer seamlessly from work to life—“A large model cannot live through life for you.”
Deep dive
1. DeepSeek R1 Turned Reasoning Models Into a Joint Victory for Technology, Open Source and Narrative
On January 20, 2025, DeepSeek R1 and Kimi K1.5 launched on the same day. R1 became the world’s first reasoning model to reproduce OpenAI o1’s performance at a large parameter scale, and its impact later far exceeded K1.5’s—even the original o1’s.
R1’s breakout was not driven by capability alone: it fully open-sourced its strongest “full-fat” version, published a sufficiently detailed technical report and deliberately highlighted that its final training run cost only $5.57M. The first two won respect from researchers worldwide; the third triggered reactions from US political and business circles, Trump and Nvidia’s stock.
程曼祺’s account of the diffusion path is specific: first, R1 earned a small but highly favorable following among the world’s top AI researchers through performance and a spirit of sharing; then it broke into the mainstream during the Spring Festival and “went abroad before coming back home.” There was also an element of frustration: “An AI model that every American says is impressive” naturally carries an extra halo in China.
R1 also changed capital’s view of the R&D threshold. 陈天桥 backed the creation of MirrorMind after seeing that frontier-model investment might be far lower than expected. The case showed that a “small and very young R&D team” could also produce world-class results at relatively low cost.
2. Reasoning, Coding and Multimodality Together Form the Agent Model
The capabilities required by an Agent can be broken into five categories: complex reasoning and task planning, coding, multimodal understanding, tool use, and memory that can preserve long-term context and know when to invoke it. They do not all need to be trained into the foundation model; they can also be assembled through external systems and protocols such as MCP.
The core improvement in o1 and R1 was multi-step reasoning, moving more computation into the model-use stage through Test-Time Compute scaling. DeepSeek further showed that reasoning does not necessarily depend on Monte Carlo tree search or supervised fine-tuning on human-written chains of thought; it can also emerge through reinforcement learning from scratch on a base model.
Claude 3.5 and subsequent models made coding another central track, and Anthropic launched Claude Code in February 2025. 程曼祺’s key judgment was that “it is not just an AI coding product; it is actually a general agent.” In the digital world, knowing how to code is like having strong hands and feet.
OpenAI launched Codex in April, xAI launched Grok Code Fast in August and Google launched Antigravity in November, showing that every leading model company is doubling down on coding. Coding can serve programmers directly, but it can also become the general execution layer through which an Agent conducts research, browses the web and automates tasks.
3. Native Multimodality Is Moving Models From Chat Windows Into Operable Environments
Lovart’s Chat Canvas demonstrates the product value of multimodal understanding: a user selects an area of a generated design and asks the Agent to modify it in natural language. The model must understand the image, the selected spatial region and the textual intent simultaneously, rather than simply generate a new image.
GPT-4o and Gemini 1.5 opened the era of native multimodality. Gemini 3 and Kimi 2.5 later followed the same model-based approach to text, images and voice. Compared with stitching together multiple single-modal modules, this gives Agents more coherent observation and action capabilities.
Visual generation has developed into a relatively independent model track, represented by 可灵, 海螺, Veo 3 and Nano Banana. Nano Banana and NotebookLM originated at Google Labs, which also staffed content editors, designers and other creative roles and once turned “photos into desktop figurines” into a Xiaohongshu trend.
Anthropic, Kimi and DeepSeek have not made multimodal generation an equally high priority. The difference reflects variations in technical judgment, application goals and resource allocation. Because foundation-model companies do not cover every modality, video, image and voice models still leave independent room for startups.
4. Co-design Is Becoming an Organizational Moat for Model Companies
One explanation for Google’s comeback is co-design: from software libraries such as TPU and JAX to AI infra, cloud platforms, Gemini and upper-layer applications, Google can optimize across the full technology stack. Alibaba’s shorthand—“Qwen, Alibaba Cloud and T-Head”—describes the same vertical-integration logic.
Beginning in the second half of 2025, Qwen started hiring its own infra talent instead of relying entirely on Alibaba Cloud personnel. Tencent also placed AI infra under 姚舜宇, hoping to “connect infra, algorithms and products end to end,” shortening iteration cycles and reducing internal friction.
DeepSeek has involved infra in algorithm design since its founding. If an algorithm is difficult to implement reliably at the systems level, it may be rejected outright. 梁文锋 can work hands-on across the stack because model training today still looks more like a large engineering project than pure scientific research detached from implementation details.
During its open-source week in February 2025, DeepSeek released infra results for five consecutive days, closing with the V3.2 reasoning system. Former intern 王子涵 gave a pointed explanation for why the company keeps open-sourcing flagship models: either “the boss does not want to make money and wants to benefit society,” or it wants to do something bigger, such as become an industry standard.
5. DeepSeek’s Reasoning Ledger Suggests Systems Optimization Has Entered an Extreme Competition
DeepSeek used a complete 24-hour data sample to estimate that more than 1,800 cards supported over 600B input tokens and nearly 170B output tokens. Converted using GPU rental rates and official model pricing at the time, the result implied a remarkable 84.5% gross margin.
游洋, founder of third-party AI infra company 路辰科技, served as the program’s opposing voice and challenged the figure directly: “Impossible.” His argument was that the calculation did not adequately account for peaks and troughs in usage and therefore could not be treated as steady-state operating gross margin.
袁进辉, founder of SiliconFlow, argued that many suppliers cannot reach this level mainly because the V3 and R1 architectures differ sharply from mainstream models. One possibility is that the team first designed a specialized architecture and then solved stable training and inference; another is that it designed the model backward from systems capabilities. Either way, the approach depends on tight algorithm-infra collaboration.
6. Both Attention and Compute Competition Are Shifting From Point Performance to System Architecture
The computational complexity and memory footprint of the original attention mechanism constrain context length, and those limits ultimately land on the physical realities of GPU compute, interconnect and storage. Sparse and linear attention are therefore not merely algorithmic issues; they necessarily extend into systems and hardware.
FlashAttention was an early example. It understood GPU memory-access characteristics, reordered the computation sequence of standard attention to improve efficiency and reduce memory use. 杨松灵, a core author of DeltaNet, also taught himself kernels after starting as an algorithm researcher, explicitly setting optimization goals that were “more GPU-friendly and scalable for training.”
Huawei’s CloudMatrix 384 supernode and Nvidia’s NVL72 both show that compute competition has moved from single-chip performance to multi-chip interconnect and system optimization. 徐明杰 also believes AI computing costs are shifting from being compute-heavy to memory-heavy. Combining chip design, interconnect and complex software capabilities will reinforce “the strong getting stronger.”
7. The Next Learning Paradigm Is Bringing the Industry Back to Research
The dominant paradigm remains massive-data pretraining followed by supervised fine-tuning or reinforcement learning on smaller, higher-quality task data. 伊利亚·苏茨克维尔, 哈萨比斯 and 勒昆 are searching for the next step because they are more inclined to define AGI through learning mechanisms rather than surface-level performance.
The capabilities they seek include few-shot learning, generalization by analogy, lifelong continuous learning and the ability to genuinely understand or even discover physical laws. In a November 2025 interview, 伊利亚 said the past few years belonged to scaling, “and now we are returning to research, returning to the research phase.”
Researchers are looking again at animal and human intelligence. Sakana AI, founded by Llion Jones, has proposed a continuous thought machine in an effort to approximate brain mechanisms; 马毅 defines intelligence as learning the regular, predictable structures of the world and is exploring learning systems with biological-style closed-loop feedback.
The human brain consumes only about 20 watts, a stark contrast with the energy required to train large models. 田渊栋 believes the next paradigm should be “simpler and more elegant,” and explainable rather than a black box. He says, “If an LLM alone can achieve AGI, then humanity’s future is bleak,” because that would mean the laws of the universe could perhaps be captured in an uninterpretable way.
8. World Models and Continuous Learning Put Scientific Discovery at the Top of the Intelligence Test
The advanced form of continuous learning is scientific discovery: a system proposes hypotheses like a scientist, designs experiments, executes them in virtual or real environments and then uses data to validate the hypotheses. 哈萨比斯 wants a system that does more than answer a question—a closed loop that can pose its own questions, obtain feedback and continue learning.
World models can provide the experimental environment. DeepMind’s Genie 3 can generate explorable 3D spaces, and SIMA 2 can move through and extend those environments. But current versions still cannot truly manipulate objects: they cannot pop a balloon, pick a flower or predict the consequences according to physical laws.
程曼祺’s philosophical question is whether models could become “AI inventors” if they not only use technology but improve technology themselves. This extends the idea of technology reproducing itself described by Brian Arthur. In the past, that process had to pass through human inventors; now the intermediary itself may begin to automate.
9. Coding Agents Are Bringing the Minimum Efficient Scale for Personalized Software Down to One Person
“Coding is not only the destination but also the means”: Cursor serves programmers, while Lovable lowers the barrier to Web coding. Claude Code works directly in the command line; Claude Cowork and OpenClaw use simpler interfaces to put the same capabilities in the hands of nonprogrammers.
TRAE Solo, 蚂蚁灵光, Macaron, Uwear and MuleRun’s Agent Builder all address personal automation. One-off needs can produce a PPT or webpage; repetitive workflows can “use an Agent to build an Agent,” turning software that was once too niche to develop into a personal application.
程曼祺’s own example is having an Agent track the social media of technology professionals every day, categorize the posts, summarize them weekly and automatically adjust the topics and names it follows as the industry changes. No one would previously have built such software for a single reporter’s needs, and she could not code herself. Now the development cost is approaching an acceptable range.
Anthropic’s MCP and Skills standards further reduce the assembly barrier. Skills are essentially reusable collections of prompts and tool descriptions. 陈凯杰, founder of Macaron, therefore argues that after scaling data, parameters and compute, the next phase will be “scaling agents.”
10. Multi-Agent Collaboration Starts With Role Division and Small Social Experiments
Moltbook has been described as “the Facebook for AI.” Every member of the network is an Agent, making it a small experiment in collective intelligence: when large numbers of agents gather and interact, will structures and behaviors emerge that are absent from any individual model?
Inspired by it, 黄东旭 built a small version to help himself write code, assigning three different roles to work together. The animation Agent OiiOii and film-production Agent MovieFlow likewise configure roles such as screenwriter, scriptwriter and storyboard artist according to real production workflows.
李笛 founded 明日星城 to explore multi-agent frameworks. 程曼祺 connected these experiments to “Plaything” from Black Mirror: the little figures in the game follow rules and interact continuously until they develop superintelligence. Real products remain far from that science-fiction scenario, but “Agent sociology” has already become an experimentally testable question.
11. The Agent Market First Needs Supply Before Distribution and Trading Can Work
Uwear once treated Web coding as a new creative medium, arguing that creation would naturally produce a sharing community. MuleRun was betting on an Agent marketplace. Both later shifted from platform positioning toward tools, reflecting the fact that there are still too few users creating Agents frequently.
Uwear’s homepage changed from showcasing user projects to a large dialog box that directly guides users to build project-management or e-commerce data-analysis tools. MuleRun 2.0’s Agent Builder likewise emphasizes creation itself. When supply is insufficient, a fuller shelf only makes the emptiness more visible.
陈宇森’s further judgment is that future marketplaces should not display products on shelves like Taobao. They should be conversational: users describe a goal, and the system immediately finds, modifies or generates the appropriate Agent. The marketplace entry point may therefore remain an Agent rather than a traditional app store.
12. The Core Barrier for Phone Agents Is Not Just Capability but Platform Economics
ByteDance released a preview of the Doubao phone assistant in December 2025, but its access to WeChat, Meituan Waimai and other services was quickly restricted. Super apps do not want Doubao to operate “over the top,” and they certainly do not want to become downstream services that another AI entry point can call at will.
Phone Agents have therefore created a three-way contest among handset makers, super apps and AI OS companies. Efficiency use cases such as ordering food, booking flights and hotels, and comparing prices are the most likely to be restructured by Agents. Entertainment platforms such as Douyin, Xiaohongshu and Bilibili may be less affected because users already want to consume the content themselves.
Advertising is the harder question. If Agents inside apps become more numerous and real people less numerous, who will see feed ads? During the transition, can AI also be influenced by advertising? Should most revenue go to the Agent directly serving the user, or to the platform that owns the transaction and content inventory?
Alibaba has a supply advantage in life services after connecting Qwen to e-commerce, instant retail, food delivery, travel, ticketing and ride-hailing. Tencent controls the massive WeChat Mini Program entry point. Whether Agents will dissolve platforms or simply form a new layer of platforms remains to be seen.
13. The Opportunity for Vertical Agents Is to Create Specialists, Not Replicate Generalists
陈冕, founder of Lovart, divides applications into Office and Adobe on the production side, and search, social and general entertainment on the consumer side. Coding and general Agents are the new Office; image, video, audio, comic-drama and animation Agents are the new Adobe.
He chose the Adobe direction because foundation-model companies mainly want to create “high-IQ generalists,” while vertical applications build designers and creators on top of that base. The survival condition is to use general capabilities fully without doing something “that one generalist can already do directly.”
爱与智能, a legal Agent backed by Mingze Investment, chose small-claim bank bad-debt litigation. Legal costs could previously exceed recovery proceeds; an Agent can screen cases with a high probability of winning and generate legal documents, but it still cannot replace appearing in court. The company itself is a law firm, does not sell software to peers and instead takes the cases directly, charging by outcome.
This model moves AI software from “selling tools and seats” toward “selling outcomes.” The program noted that Sequoia made a similar observation in May 2025, but whether a vertical team can bear responsibility for the result depends on its control over workflows, data and delivery—not merely its ability to call a model.
14. The Agent Toolchain Is Following Model Upgrades, Closing the Gap Between “Can Use” and “Works Well”
Everything between a foundation model and a complete Agent can become infra: RAG, prompt engineering, context engineering, reinforcement-learning environments, evaluation, state checks, voice, memory and other components have already gone through several rounds of iteration.
Henry Ying of MOE Capital summarizes the toolchain’s development as six major waves. Each time model capability advances, older methods fail or expose a new gap, and tool companies help developers turn theoretical capability into a stable product.
The fine-grained division of labor in US software provides a reference for this ecosystem. Even ordinary-user-invisible layers such as observability provider Datadog and identity company Okta can produce companies with more than $2B in annual revenue. Existing software processes may also be rebuilt by AI from the ground up.
15. The Sora App Shows That Generation Capability Does Not Automatically Create a Consumer Platform
Sora App’s distinctive advantage was the combination of OpenAI’s leading video model at the time and Cameo interaction. After authorizing a face, users could generate characters resembling themselves and co-star with friends, prompting a wave of “AI YouTube” speculation.
陈冕’s optimistic case was that Cameo would motivate users to recommend the product to friends, making Sora both a tool and a social product. 曹越, founder of SuneAI, applied a stricter standard: a new platform must at least create a new content form and a new distribution chain, and Sora had achieved neither at the time.
Subsequent data supported the cautious camp. Sensor Tower estimated Sora App’s day-30 retention at below 8%, far below the program’s cited figures of 42% for TikTok and 38% for Instagram. Many users still generated content in Sora and then distributed it through WeChat Moments, Xiaohongshu, Douyin or Kuaishou.
Low overall retention does not mean nobody creates on the platform long term. Sora Stats serves active creators and tries to help them become the MrBeast of the platform. Its groups include teachers and Uber drivers, showing that creators are not limited to technology workers or AI influencers.
16. Japanese AIGC Users Show the Clearest Cultural Affinity for Consumer AI
程曼祺 follows Kago Masumaru, who never posts realistic Cameo content and instead uses Sora to explore anime culture and visual effects. The creator behind the account is 松丸彗吾, a Japanese puzzle-game producer and younger brother of influencer 松丸亮吾. By the time the program was prepared, the account had more than 100,000 followers, approaching Sam Altman’s 140,000.
Sam Altman once specifically praised the “remarkable creative output” of Japanese users. The program attributed it to the country’s long-standing ACGN and fan-creation culture: AI tools amplify existing creative habits, ideas and communities without first having to educate users on why generative content matters.
Sora therefore presents a dual reality. It did not rapidly reach the mass market as early expectations suggested, but it established stable use cases among specific cultural and creator communities. Whether consumer AI can cross the chasm depends not only on model performance but also on existing content cultures and distribution relationships.
17. Non-Productivity AI Applications Are Still Searching for Entrances in Communication, Input and Companionship
Intent was founded in San Francisco by 25-year-old 陈春雨 to serve the local Spanish-speaking population by embedding large-model translation into chat software. It is challenging a communications market that appears mature but still contains substantial cross-language friction.
Typeless takes a different route through voice input. 程曼祺 believes its accuracy, text correction, structural organization and long-term memory for frequently used names and proper nouns are clearly better than the built-in functions of WeChat and Feishu. The difference is not dramatic, but it sits on top of a high-frequency essential need.
Companionship, socializing, gaming and entertainment did not begin in 2025. Glow, Talkie and 猫箱 had already validated interaction with virtual characters. 陈冕 predicts that 2026 will be the breakout year for AI-to-C, but Sora’s retention also shows that novelty is not the same as sustained use.
18. AI for Science Is Moving From Accelerating Computation to Automatically Proposing and Testing Hypotheses
深势科技’s early route was to use machine learning to simplify first-principles quantum-physics calculations. DPMD, which 张林峰 worked on during his PhD at Princeton, increased computational efficiency without sacrificing accuracy, allowing methods such as the Schrödinger equation, density functional theory and molecular dynamics to scale from microscopic systems toward broader applications.
The second phase was to use deep learning and generative AI to address scientific problems directly, including AlphaFold’s protein-structure prediction and 深势’s Uni-Fold. These capabilities could later be packaged into the drug-development computing platform Bohrium.
As large language models matured, the target expanded to the full research workflow: reading literature, proposing hypotheses, designing experiments, running them and validating results. Another direction is “using AI to improve AI,” with Agents taking on part of the work of AI researchers—a form of technology stepping on its own left foot with its right.
Brian Arthur defines technology as a means of achieving a goal: old technologies can be recombined into new technologies, and new technologies create new needs that improve themselves. If an AI inventor becomes an intermediary with far broader information access, far more compute and no need to sleep, the question shifts from “can it?” to “are we ready?”
19. ByteDance Used Independent Organizations and Talent Density to Launch a Saturation Attack on Models
ByteDance separated its AI organization relatively independently. Internally, it consists of Flow for products, Seed for models and Storm for backend support, seeking to escape the inertia of a company with more than 100,000 employees by operating more like a startup. Doubao sits within Flow, while core model R&D sits within Seed.
In the summer of 2023, ByteDance had considered investing in MiniMax and 阶跃星辰. 张一鸣 then made his position explicit: large models should be built in-house, and could be done well. He and senior executives subsequently met people intensively; in one case, when an investor visited a paper’s author, the answer was reportedly, “I just met with 一鸣.”
ByteDance first recruited young technical leaders including 蒋路, 秦宇佳, 黄文浩, 丁明 and 周畅, then brought in 吴永辉, former research vice president at Google DeepMind, in February 2025. The organizational logic was to gather leaders in their ascent first, then find a senior technical manager who could “hold the room.”
Multiple strong teams also created intense internal competition, including parallel teams working on video generation. One technical leader cited by the program distinguished between “horse racing” and “boxing”: in horse racing, a loss leaves another round; in boxing, the loser exits. The real danger is permanently assigning people and tasks before the direction has been validated.
20. Alibaba’s Model Strength Grew From a Peripheral Team Into a Core Force
周靖人, Alibaba’s AI chief, simultaneously serves as CTO of Alibaba Cloud and head of the Qwen laboratory, and became a partner in 2025. Under him, Qwen handles the core language model, 通义万相 handles image and video generation, and another team works on multimodal interaction.
Qwen, led by 林俊旸, who was born in the 1990s, initially grew from a relatively peripheral position inside the company. The small, elite team first focused on open-source language models, then built community influence and a developer ecosystem. It later expanded into embodied VLA, built its own infra and launched Qwen Chat to showcase its latest capabilities.
周靖人 gave Qwen considerable autonomy and was also the earliest advocate of open-sourcing Alibaba’s large models. Unlike Seed, which was built from scratch, 通义 emerged from the merger of several DAMO Academy labs and lacked sufficiently independent compensation incentives. Broad pay and promotion increases at the end of 2024 were interpreted as a response to market conditions.
Several veteran leaders also left 通义 in 2025, but Qwen’s open-source ecosystem, team initiative and relatively stable leadership became important assets in Alibaba’s AI strategy.
21. Tencent Used 姚舜宇 as a “Catfish” to Reset Its Model Organization
姚舜宇, a former OpenAI researcher and the author of ReAct, joined Tencent in autumn 2025. In December he became chief AI scientist in the president’s office while also taking responsibility for AI Infra and the large-language-model division. Born in 1998, only recently out of his PhD and with limited management experience, he represented an unusually aggressive appointment for Tencent.
After joining, he recruited intensively, including 王炳轩, a researcher of his generation brought in from DeepSeek. He also redefined the target, arguing that Hunyuan had focused too heavily on benchmarks—being strong on leaderboards did not mean offering a good real-world experience—and pushed co-design across infra, algorithms and products.
Unlike ByteDance, which recruited young technical leaders first and then brought in a veteran manager, Tencent introduced young researchers first as the organizational variable. Some called 姚舜宇 a “catfish,” but the program also drew a boundary: DeepSeek is not young only; 梁文锋’s coverage and understanding of every part of the large-model stack are equally critical. Tencent’s restructuring will show more results only in 2026.
22. Model-Product Collaboration Determines Whether User Scale Converts Into a Data Flywheel
曹越 of SuneAI took an organizational lesson from Sora App: OpenAI can send product demand gradients back into the model, an example of end-to-end vertical integration. By that standard, ByteDance has the tightest model-product collaboration among China’s three major tech companies, with Flow and Seed showing a strong shared-service mentality.
ByteDance’s three AI goals at the beginning of 2025 were to explore the upper bound of intelligence, explore new UIs and strengthen economies of scale. But every large-model call consumes compute, making the business more like manufacturing with unit costs. More free Doubao users also mean higher costs, so the logic of “build once, sell many times” cannot be applied directly.
Doubao 1.5 Pro explains economies of scale through a user-data flywheel: identify problems, mine data, perform human-machine labeling, run rapid post-training and improve the real-world experience. The two ends of the flywheel are “large-scale user feedback” and “actual user experience”; user data does not automatically make the model smarter.
闫俊杰, founder of MiniMax, once warned that “you should never apply the mobile-internet playbook to building a large model.” ByteDance is still trying to transfer the A/B testing and operating infrastructure accumulated in recommendations, search and advertising. Doubao’s growth shows that the approach works, but it does not eliminate compute costs.
23. Product Restructuring at Alibaba and Tencent Exposes the Coordination Cost of Large Organizations
Tencent grouped Yuanbao, QQ Browser, Sogou Input and intelligent workstation ima under CSIG, while the Hunyuan model team remained in TEG. Models and products therefore sit in different business groups, while WeChat remains independent and every change must withstand the constraints of a billion-plus users.
WeChat even blocked group sharing of Yuanbao’s RMB1B Spring Festival red-packet link, citing a crackdown on excessive marketing and incentivized sharing. The episode showed that an internal AI campaign cannot automatically cross WeChat’s governance boundary.
Alibaba initially placed the Tongyi Qianwen app and model under Alibaba Cloud, then moved the To-C team to the intelligent-information business group managed by 吴嘉. The product went through two renamings—Tongyi Qianwen to Tongyi to Qianwen—while flagship status alternated between Quark and Qwen.
Quark took over in 2025 with its “AI Super Box,” but was constrained by habits built around a decade-old product: new users came to try AI, while existing users still mainly used cloud storage and question solving. After Qwen regained flagship status and focused on life services, the program’s assessment of the repeated adjustments was blunt: “Some problems are being treated only where the pain appears.”
24. Doubao Leads Consumer AI, While Qwen and Yuanbao Bet on Ecosystem and External Models
Doubao is China’s first AI product with more than 100M DAU and has remained in first place for an extended period. Its challengers have shifted from Kimi in 2024 to DeepSeek during the 2025 Spring Festival and then to Yuanbao after it integrated DeepSeek. Dola, the overseas version, has also surpassed 10M global DAU.
ByteDance uses a “one superpower, many strong players” strategy: alongside Doubao are TRAE, Coze, 豆包爱学, 猫箱 and 即梦. Douyin gives Doubao content distribution, from an argumentative persona that banters with users to fashion advice. Creators can even build accounts around related content and then bring new users back.
The commercialization model is mismatched with traffic: the largest product, Doubao, is free, while AI revenue mainly comes from tools serving professional users. Doubao has experimented with triggering product cards and service recommendations inside conversations. A key 2026 question is whether commercialization will damage its approachable, everyday persona and user experience.
Alibaba connected Qwen to e-commerce, food delivery, travel and ticketing, with peak marketing spend reaching RMB15M per day. Some insiders were pessimistic, arguing that buying traffic at scale before the product was mature would make it difficult to compete with Doubao’s already-established position in users’ minds.
25. Yuanbao Used DeepSeek to Fill Its Model Gap, Pushing Customer-Acquisition Budgets to Extremes During the Spring Festival
Tencent’s in-house Hunyuan was relatively weak, which paradoxically meant the company faced the fewest constraints around integrating DeepSeek. After Yuanbao, QQ and WeChat all integrated it, Yuanbao’s DAU rose tenfold in the final week of February. The result showed that a leading third-party model can rapidly change an application’s competitive position.
Yuanbao Pai brought WeChat and QQ social relationships into the AI product: friends entered through a link into a group containing the Yuanbao assistant. It attempted to turn Tencent’s strongest relationship graph into an Agent scenario, but WeChat’s restrictions on the red-packet link also showed that group resources cannot be called on unconditionally.
The Spring Festival scoreboard featured Doubao winning the CCTV gala, Yuanbao committing RMB1B in red packets and Qwen subsequently announcing RMB3B. The biggest variable was whether DeepSeek V4 would launch before the holiday and how it would perform. Even if ordinary users could barely perceive differences among the top models, the DeepSeek brand could still materially benefit the product most tightly bound to it: Yuanbao.
26. Foundation-Model IPOs Are Financing Tools, Not Trophies Signaling the End of the War
Zhipu and MiniMax listed one after the other in January 2026, making China the first country to produce a “first listed large-model company” ahead of the US. As of February 7, the date cited by the program, their market capitalizations were RMB79.4B and RMB128.4B, or roughly $11.4B and $18.5B.
The valuation gap with OpenAI and Anthropic narrowed from two orders of magnitude to one, but the revenue gap remained two to three orders of magnitude. Zhipu reported RMB190M in revenue for the first half of 2025, and MiniMax more than $53M for the first three quarters; OpenAI said its full-year revenue was well above $13B, while Anthropic was reportedly at roughly $4.5B.
All leading model companies remain deeply loss-making. They must continue training while paying compute costs for every call. The program cited Bernstein’s estimate, inferred from Microsoft’s financial statements, that OpenAI’s quarterly loss could reach $12B. A fixed subscription model can especially obscure the marginal cost generated by high-frequency users.
One reason Chinese model companies listed earlier is that less money remains available to support large models in private markets. 程曼祺 wrote that an IPO “is not a reward for the winner, but the starting drumbeat for the next round of competition.” Her more direct judgment, omitted from the report, was that “an IPO is also a safety net for staying alive.”
27. Chinese Model Startups Are Playing a Harder Game Without a Hardware Upgrade Dividend
On the same day MiniMax announced its listing date, Kimi announced a $500M financing round. 阶跃星辰 also completed a large financing, while DeepSeek had backing from 幻方. But since the second half of 2023, it has become difficult in China to establish and finance a new foundation-model company.
New labs continue to emerge in Europe and the US: SSI, Reflection, Thinking Machines Lab, Advanced Machine Intelligence Labs and Core Automation, founded by Jerry Tworek after leaving OpenAI. One of Tworek’s reasons was that high-risk frontier research had become difficult to advance inside a large organization.
Unlike mobile internet, large models have created new software but have not yet brought an iPhone-like mainstream hardware wave or an App Store-like natural distribution channel. Startups still live on the Web and phones; channels, traffic and growth methods are controlled by incumbents, while Chinese tech giants generally choose to build in-house rather than acquire.
These founders are competing for the position of the next generation of technology companies by building models directly. The upside is high, but so is the risk of falling short. For foundation-model companies, “surviving without reaching sufficient scale can itself count as a failure.”
28. $50M in ARR Is the Initial Commercial-Validation Threshold for Application Companies
The program estimates that roughly 40–70 AI-native companies globally have ARR of at least $50M, concentrated in AI coding, content creation and verticals such as law, education and healthcare. ARR annualizes subscription revenue over a given period and is generally above actual revenue for fast-growing companies.
The list includes Perplexity, Cursor, Lovable, Midjourney, Higgsfield, Runway, ElevenLabs, HeyGen, Synthesia and Harvey. They prove that real revenue has emerged at the application layer, but the number of companies that have crossed this scale is still far smaller than market enthusiasm suggests.
Among teams with Chinese roots, Manus was acquired by Meta for roughly $2B in December 2025. Some members of its more-than-100-person team had followed 肖弘 through a decade of entrepreneurship, and a substantial proportion held options. 程曼祺 called it a “fairy tale”: it may not have reached the destination the founder originally imagined, but investors and employees still received a return.
Genspark said in September 2025 that its ARR had reached $50M; OpenArt reached $70M with roughly 20 people; Lovart passed $30M four months after launch; and PixVerse’s full-year revenue was reportedly above $40M.
29. Small Teams Can Achieve High Productivity, but Foundation Models May Move Into High-Value Applications
The ranking maintained by Henry Shi tracks companies founded less than five years ago with fewer than 50 employees, ARR of at least $50M or revenue per employee above $1M. Its latest version includes 44 companies, but only 14 have actually crossed $50M in ARR.
OpenClaw was launched by Austrian developer Peter Steinberger, with most of the core development completed by one person. It is an extreme example of “few people, outsized impact.”
Foundation-model companies may move into large, high-value areas such as coding, general Agents, marketing, education and health. Application companies can establish themselves in gaps and later train models downward, as Perplexity and Cursor have done, or remain small and focused and seek acquisition. They do not all need to “Go Big or Go Home.”
30. Globalization Has Become a Question of Corporate Structure and Survival
HeyGen initially built teams in San Francisco and Shenzhen. Its Chinese entity, 诗云科技, was dissolved as early as December 2023, a choice made before AI had fully entered the mainstream. Manus moved its headquarters to Singapore in June 2025, adopting a more balanced structure.
After Meta acquired Manus, China’s Ministry of Commerce said it would review compliance with relevant departments. Scrutiny of a high-profile company is not surprising, but MirrorMind showed that even an early-stage team operating for less than a year can have its fate suddenly changed by legal and cross-border structures.
MirrorMind’s domestic team first built Agent RL post-training and then added pretraining, releasing MirrorThinker 1.5 in January 2026. Less than half a month later, its head 戴继峰 left. The program explained that 陈天桥 could no longer continue supporting domestic R&D because of legal-compliance concerns, leaving the team to raise independently or find another route.
Founders must judge who will control the leading models, whether a scenario truly requires the most advanced model, and whether to prioritize performance and scale while tolerating losses or pursue profitability in stages. Open-source communities may eventually carry technology across borders, but “how long is long term” is enough to determine the fate of an entire cohort of companies.
31. Embodied-AI Financing Is Running Far Ahead of Technical Consensus and Commercial Deployment
Total Chinese financing for embodied AI and robotics reached RMB73.5B in 2025, compared with RMB18.2B raised by the several companies still training large models, including the IPOs of Zhipu and MiniMax. The steady stream of weekly financing news shows that capital has already made embodied AI the next major track.
Figure, the most highly valued US company in the category, had a latest valuation of $39B—roughly one-twentieth of OpenAI’s. Chinese leaders in embodied AI and models, by contrast, are valued in the tens of billions to more than $100B. Galbot reported a $3B valuation, while Unitree’s Series C valuation exceeded RMB12B.
New teams continue to enter because frontier companies worldwide are still in the experimental phase, with no stable consensus on VLA, data or embodiment routes. This differs from the large-model startup wave of 2023, when everyone at least understood the rough process of pretraining and fine-tuning and could initially set the goal of reproducing GPT-3.5 or GPT-4.
Automakers also have three advantages: technical talent, manufacturing supply chains and factory scenarios. XPeng and Li Auto are both increasing their investment in humanoid robots, reusing accumulated autonomous-driving capabilities, testing robots in auto factories and strengthening their technology brands.
32. Government Funding, Manufacturing Strength and Exit Expectations Are Driving Up Chinese Embodied-AI Valuations
Embodied AI fits local governments’ preferred investment-attraction profile: visible hardware, production lines and jobs. By the end of 2025, nearly 30 embodied-AI training facilities had been built or planned nationwide, with robots purchased in batches and operators hired to teleoperate them for data collection.
Training facilities could become customers and revenue sources for some companies, while representing R&D costs for others. Companies could even manufacture their own robots, build their own data-collection centers and use the resulting data for their own models. The program also warned against “training facilities” that are actually industrial-park real-estate projects.
China’s supply chain can mass-produce embodiments at low cost. Unitree brought the starting price of a humanoid robot down to roughly RMB100,000, making it a common development tool at universities and research institutes worldwide. It has also reached several hundred million yuan in revenue and profitability. Selling to schools could create a developer ecosystem beginning at the PhD stage.
Manufacturing companies are more welcome in China’s secondary market, making the exit path appear clearer than in software. Unitree is pursuing a STAR Market listing, and several embodied-AI companies are reported to be planning Hong Kong listings in 2026. The problem is that most have no scaled applications, have yet to prove recurring revenue and remain loss-making, while capital is already trading on possibilities several years out.
33. Agibot-Related Trading Shows That the Secondary Market Is Paying an Irrational Premium for Embodied-AI Concepts
After Agibot acquired 29.99% of STAR Market-listed Shangwei, the stock hit successive daily limits. Its market cap rose from just over RMB3B to a peak of RMB69B and remained above RMB55B during the program. Its original core business was mainly wind-power materials.
The transaction did not amount to a backdoor listing. That would also require a change of control, asset injection into the listed company, substantive replacement of the main business and post-injection profit and revenue metrics meeting IPO standards. Agibot had completed only the first step.
Several companies are preparing to list, Hong Kong market conditions may fluctuate and companies remain broadly loss-making. Together, these factors make the 2026 embodied-AI listing wave a major suspense point. Early-stage losses do not by themselves prove that a company is poor, but neither do they automatically justify a high valuation.
34. Embodied-AI Data Has No Unified Scaling Path, and Simulation Is the Biggest Fault Line
Current data methods include real-machine teleoperation, Sim2Real simulation, Internet video, UMI wearable collection and self-play through which robots adjust based on failure. The methods complement one another in cost, accuracy, quality and coverage, and most companies use a mix.
Tesla’s Optimus initially relied more on teleoperation, then shifted toward more self-play and simulation, with limited human calibration. Galbot believes teleoperation is too expensive for startups to bear and therefore emphasizes simulation; Hillbot also treats high-quality 3D simulation and reinforcement learning as core routes.
Teams including TARS, 千寻智能, 原力灵机 and 自变量 remain cautious or outright negative about simulation-based training for embodied models. The mildest description is “not very workable”; the most decisive is “this is a giant pit.”
The core dispute is not whether simulation is useful, but how large the gap is between generated environments and the real world, which capabilities can transfer and how much real-machine data is required to close the gap. If data cannot be scaled cheaply, embodied AI will struggle to replicate the data scaling seen in language models.
35. VLA, End-to-End Systems and World Models Are Options, Not Established Standards
VLA stands for Vision-Language-Action: visual and language inputs produce robot actions. The mainstream method usually starts with an LLM, obtains a VLM and then trains a VLA. 千寻 proposes adding touch to form VLTA, while 原力灵机 argues that the base model could be a video model rather than an LLM.
“End-to-end” essentially means replacing as much rule-based programming as possible with deep neural networks. Some in embodied AI expect one network to perform every operation; others advocate hierarchy, with a high-level model handling complex planning and other models executing at the action level, rather than forcing different time scales into one system.
Generative world models attempt to predict the next state from the current state. In 2D, the result resembles an indefinitely extending video; in 3D, it resembles Genie 3. A complete version must also support interaction with objects and predict changes according to physical laws. Current systems remain far from that point.
36. Dexterous-Hand Lifespan Shows How Far Embodiments Remain From Commercial Durability
One key reason for the delay to the third-generation Optimus is that its dexterous hand lasts only about six weeks, is prone to damage and cannot be repaired locally. The entire hand must be replaced, at a cost of more than $6,000 each. Even without discussing intelligence, that fails the durability and maintenance-cost requirements of real commercial settings.
Embodiments also face the lifespan and stability of arm and leg joints, as well as system-level problems involving weight reduction and battery life. One supplier repeatedly affected by delays complained: “Old Ma’s credit score is now so low he can’t even borrow a power bank.”
Optimus is a window into the highest level of robotic embodiment, but it also reminds capital that model demos can update quickly while mechanical structures, materials, supply-chain quality and mass manufacturing move at different speeds. Embodied intelligence must cross two maturity curves at once: software and hardware.
37. Robot Sales Have Arrived, but R&D and Performance Still Account for Much of Real Demand
Agibot says humanoid robots have reached a 5,000-unit production scale, while Unitree says actual sales exceeded 5,500 units in 2025. Public figures make it difficult to determine the specific customers, uses and proportions by scenario, so they cannot be equated directly with the commercialization of autonomous productivity.
Current applications can be divided into R&D, performances, industrial work, commercial or household services, and companionship. Selling to training centers, universities and labs reflects real demand and should not be mocked in the industry’s early stage. R&D demand, however, is temporary and has a ceiling.
China Mobile placed a humanoid-robot order worth more than RMB120M, including roughly RMB78M for Agibot and RMB46M for Unitree, with delivery scheduled for 2025–2027. Uses include inspection and meter reading in data centers, marketing displays in exhibition halls and embodied-data collection.
Performance rentals have reportedly recouped their costs in “two weeks to one month,” but dance, martial-arts and boxing videos often show a human operator nearby. The business proves that hardware can generate revenue; it does not prove that robots already understand and execute movements autonomously.
38. Industrial Applications Are Huge, but Humanoids Will Not Easily Dominate Them
New robots have the best chance in areas that are difficult for traditional automation: flexible objects such as wiring harnesses and fabric, fine force control for electronic assembly and insertion, small-batch, high-SKU production with pronounced seasonal swings, and mobile tasks such as construction and interior finishing.
Industrial environments are highly fragmented. Different sectors have different requirements for cost, precision, safety, cycle time, stability and payload, making customization unavoidable. A general-purpose robot needs both generalization and an embodiment that is durable, stable and cost-controlled.
Semiconductor manufacturing, bottled beverages, automotive welding, dispensing, loading and unloading, and warehouse depalletizing are already highly automated. Mature dedicated machines may remain superior to new humanoid or humanoid-like robots in cost, reliability, cycle time and payload, both in the near term and potentially over the long term.
The program also described an industry story in which “you buy my robot, and I give you orders,” while the supplier uses the concept to push up its stock price. Embodied-AI companies, suppliers and shareholders appear to form a “three-way win.” Such a loop can create revenue, but it does not necessarily prove that end users are willing to keep paying.
39. Homes and Companionship Make Value Easier to Demonstrate, but Have Not Reached Scaled Deployment
Homes and commercial services may be friendlier to startups because they face less competition from existing automation. Sanctuary AI and 1X have demonstrated actions such as picking up a wine glass and folding socks. Della Robotics is targeting US laundromats and hotels for towel folding, but the market remains at the prototype and pilot stage.
Companionship does not require a breakthrough in complex task planning. Automatic following, motion control, filming, patrol and language interaction can be combined. Unitree’s consumer robot dogs have sold tens of thousands cumulatively but have not truly broken into the mainstream after years, showing the distance between usability and mass-market need.
维他动力’s V-Bot focuses on following, carrying, filming and home patrol. It reported 6,540 presale orders in half a month, but deposits remain refundable until formal order confirmation. Lock-in and delivery data in March 2026 will be closer to a measure of real demand.
Several industry participants predict that 2026 will be the first year of scaled deployment. 程曼祺’s test is simple: over the next ten months, will we see only “expectations,” or judgments grounded in orders, delivery, repeat purchases and autonomous operation?
40. AI Glasses Are Competing for the Context Entry Point, Not Just Adding Another Screen
AI glasses are expected to become the next entry-level hardware after smartphones because they are both hands-off and always-on. Users can speak directly to issue tasks, while the device stays close to the visual and auditory centers throughout the day, capturing images and sound from the human point of view.
Snapshot capture and recording are only the current functions. The long-term value lies in collecting physical-world data unavailable to existing devices. 洪晓平 put it most sharply: “If you want to disrupt Apple, Google and Microsoft, you need to obtain the context they cannot obtain.”
Looki uses a portable camera that can be worn on the chest, capturing throughout the day and automatically summarizing the footage into videos or comics. Like glasses, it begins as a sensor platform. Big tech companies have already occupied much of Internet data, so startups are trying to find a larger new data pool in the offline world.
Voice is suitable for input but not the most efficient way to receive information; without a display, headphones are sufficient for many scenarios. Displays create their own problems around clarity, field of view and motion sickness, while batteries, sensors and chips add weight, constrain battery life and generate heat.
41. Entry-Level Hardware Has Not Converged, Making Lightweight Iteration More Realistic Than a One-Shot Perfect Product
Vision Pro represents the fully featured, all-at-once route, but its weight and price show that current technology cannot yet support a “perfect product.” Ray-Ban Meta began with simple functions such as music, calls and photography; after the second generation improved the experience, cumulative sales exceeded 3M units.
The third-generation Ray-Ban Meta Display added a screen in September 2025. The program also cited media reports that Apple could launch lightweight glasses more like the second-generation Ray-Ban Meta in the second quarter of 2026, without a display, to reduce engineering risk.
Google plans to launch glasses with partners, with Gemini providing model capabilities and hardware partners building the devices—roughly the path of a glasses OS. Partners include Chinese company XREAL. Before the entry point takes a final form, models, operating systems and hardware brands could each occupy different layers.
Another possibility is that there will never be a single smartphone-like entry point. AI could remain distributed across phones, cars, glasses, headphones, recording devices and specialized hardware, with each product perfecting one high-frequency task.
42. Plaud Shows That Specialized AI Hardware Must First Solve a Specific Pain Point
Plaud is a thin recording device attached to the back of an iPhone. It uses vibration-based audio capture to address the difficulty of recording iPhone calls. Because it is bound to the phone, users do not need to remember to carry an extra device, naturally increasing usage frequency.
It then uses large models to process meetings, interviews and other scenarios, generating summaries and to-dos. The first generation launched in June 2023, before ChatGPT’s breakout. Plaud therefore captured a hardware timing window while quickly adding generative-AI capabilities.
Plaud says cumulative global sales have exceeded 1M units. The company now has ten partners, with a team that includes both serial entrepreneurs and product-engineering talent from phone manufacturing and consumer electronics. Hardware delivery requires a much broader capability set than model calls alone.
The case supports 潘宇扬’s objection: it is not enough to say that “AI hardware is a vessel for large models” without explaining the use case and user need. Hardware ultimately has to prove itself through feature fit, usage frequency and willingness to pay.
43. Health and Companionship Hardware Can Resonate With Niche Users, but Crossing the Chasm Remains Difficult
Odyssey AI’s necklace uses multimodal vision to record food intake and provide health recommendations. Oura Ring has sold 5.5M units cumulatively and reached a valuation above $10B. Its relationship with large-model technology may not be strong, “but this valuation may have quite a lot to do with it.”
Looki did not meet any clear need for 程曼祺 herself, but podcast host 树羊 used it consistently and recommended it to peers. The contrast captures the current state of many AI products: they can genuinely resonate with one group, but it is unclear how niche word of mouth becomes a mass market.
Fuzo Fuzo attracts young women through cat-like eyes, character design and emotional feedback. 程曼祺 criticized its high latency and inability to distinguish speakers in person, then saw a diary entry in the app: “曼祺 thinks I’m too stupid. I need to work hard to improve.” She felt a twinge of guilt and marveled that the product was “quite good at playing on people’s emotions.”
The problem for companionship products remains sustained use. Appearance, personality, IP and emotional design can drive the first conversion, but whether users have the time to interact over the long term determines whether the product is a novelty toy, a form of content consumption or a stable relationship product.
44. Shenzhen’s Hardware Flywheel Runs on Engineers, Supply Chains and Global Niche Markets
Shenzhen companies including Insta360, 拓竹, Strata and HyperShell all use AI, but do not treat “AI-native” as their starting point. They begin more pragmatically with user needs and implementable technology, reflecting another view of hardware: software matters, but the final deliverable is a complete product.
Many founders came from DJI, including 拓竹’s 陶冶, Strata’s 洪晓平 and teams building yard robots, outdoor energy storage and AI guitars. Companies cluster around Xili and DJI’s Sky City in Shenzhen and maintain close links with Shenzhen Institute of Artificial and Intelligence and Robotics and Dongguan’s Songshan Lake robotics base.
李泽湘’s ecosystem shifted from To B toward To C because young engineers can more easily enter consumer markets through innovation and product definition. B2B relies more heavily on commercial relationships, organizational management and experience. Once a consumer brand takes shape, it pulls the supply chain along, and the supply chain supports the next round of product experiments.
A common route is to target high-income markets in Europe and the US, earn high gross margins through innovative experiences and quality, and launch first on Kickstarter. Demand clusters around makers, outdoor mobility, entertainment and household services, including 3D printing, energy storage, smart instruments, lawn mowers and pool robots.
45. The Real Barrier for Shenzhen Companies Is Crossing From Niche to Mass Market While Defending a New Category
The first hurdle is expanding from niche to mass. Drones found a clear aerial-photography use case, while 拓竹 built the MakerWorld model community in parallel so users who could not design models could still use 3D printers. Hardware and content ecosystems expanded the market together.
The second hurdle is competitive strategy. Any attractive category draws imitators, and robot vacuums quickly became a red ocean after innovation spread. Insta360 initially chose to build profit and capabilities in vertical markets. 刘靖康 said: “In the long run, fierce competition is inevitable, but my initial hope is that it arrives as late as possible.”
By 2025, Insta360 had formally launched a drone, while DJI released a panoramic camera. The two companies were now attacking and defending across multiple imaging categories. Hardware success is not a one-off hit; it is long-term management of new categories, channels, supply chains and competitive timing.
46. Sky-High AI Compensation and White-Collar Layoffs Are Happening in the Same Cycle
San Francisco is simultaneously filled with Meta offers worth hundreds of millions of dollars, startups raising tens or even hundreds of millions at launch, wealth stories from stocks and investments, and arguments over whether OpenAI’s $500B valuation is justified. The AI world is “prosperous and restless,” with grand technology and sudden-wealth narratives coexisting.
The other side is career anxiety among data scientists, HR professionals, operators, designers and software engineers. At one event, multiple attendees asked how data scientists should transition. 立正’s answer was brutally direct: “Don’t become a data scientist.”
Amazon announced 14,000 layoffs in October 2025 and another 16,000 at the end of January 2026. Layoffs.fyi estimates that Google, Meta, Microsoft and other major technology companies collectively cut roughly 60,000–80,000 jobs in 2025.
China began aggressively expanding AI hiring from February 2025 after the DeepSeek boom, with monthly postings continuing to rise by double digits month over month. At the same time, demand for middle-management roles fell by roughly 25%. Alibaba’s morale improved alongside AI and its stock, producing the internal slogan “Make Ali Great Again.”
47. The Effective Way to Use AI Is to Start With Real Pain Points and Iterate Repeatedly at Work
卡兹克’s workflow scrapes daily AI news, has a model rank its importance and selects several dozen items. He also looks for content with “few followers but many likes,” uses AI to analyze its distribution characteristics and uses the findings to inform story selection.
A more concrete example is organizing offline events. Previously, several people would spend days screening registration comments and grouping participants according to rules such as who could or could not sit at the same table. AI can identify high-quality intent and handle complex constraints, compressing repetitive work.
His learning method is “learning by doing”: start with a real need and keep experimenting. If you do not know where to begin, start with the part of your work you hate most and do not want to repeat—“AI-ify anything you do three times.”
程曼祺 also acknowledges that despite spending the year reporting on and trying large numbers of AI tools, she has not felt meaningfully freed from work. The industry’s heat has generated more story ideas, while face-to-face conversations, interviews and carefully considering communication remain difficult to replace. A friend reassured her: “That’s a good thing—it means AI won’t replace you so quickly.”
48. AI Will Ultimately Force People to Find Sources of Meaning Beyond Work
胡泳 has argued that AI’s major challenge for ordinary people is to reconsider that “the meaning of being human does not lie in work.” The reality is that most people’s identity, sense of achievement and meaning still come from careers and from answering the question, “What do you do?”
程曼祺 does not fully believe the promise that once AI frees people from work, they will naturally enjoy life. Her analogy is: “It’s like banning dating from elementary school through university, then expecting you to get married as soon as you enter society.” People have not been adequately trained to find what they love, enjoy life or experience passion.
田渊栋 imagines that the initial phase of substitution will bring emptiness: work results become easier to obtain, reducing the motivation to act; knowing that no amount of effort can outperform AI may lead people to give up altogether. Society may then shift toward competing on creativity, individuality and scarcity, but not everyone has those talents, and the pain will not disappear immediately.
His long-term outlook is more optimistic. If AI takes on most work that can be turned into skills, education and evaluation systems may no longer revolve solely around professional competence. People could turn toward activities driven by genuine interest, producing an explosion of occupational diversity. 黄东旭’s endpoint is simpler: life is about experiencing the journey—“after all, a large model cannot live through life for you.”