Vol.94 If We Don’t Talk About OpenClaw Now, We May Not Need to Talk About It at All — Crossover with Zhiben Lun
Summary
- OpenClaw is not a revolution at the model layer; it puts models that are already close to “F1 engines” into a car ordinary people can drive. It packages capability through memory, preferences, tools, local files and IM connections, but its real value depends on how it is “raised”: basic searches or reports may not beat mature chat tools, while complex Coding, multi-stage delivery and acceptance processes create a clearer gap. “The driver still matters.”
- The first revenue to materialize this cycle is clearly coming from courses, while installation services and dedicated hardware are also absorbing the surge, with technical adoption and sentiment arbitrage happening at the same time. Media outlets are manufacturing anxiety that anyone who fails to get on board will be left behind, while intermediaries lower the barrier to entry; because OpenClaw is open source, big tech companies, KOLs, schools and even governments can all ride the wave. 庄明浩’s commercial judgment is blunt: companies with something to gain should “join in wherever possible”; “you have no choice when you’re in the game.”
- The real B2B disruption is not any particular “lobster,” but the possibility that natural language could drive the cost of building broad categories of software close to zero. CIO-like roles inside companies are therefore reassessing SaaS purchases and internal processes, but data, security, permissions, KPIs and organizational coordination mean enterprises will not migrate overnight; the only debate is whether the shift from watching to scrambling will happen suddenly or still require a longer implementation cycle. The enterprise question has moved from “should we use it?” to “how do we use it, where do we use it, and who does the work?”
- Token usage should not simply be described as a tool for model companies to harvest excess compute; it is the fuel for a market in which demand is growing faster than unit costs are falling. A single interaction has expanded from several dozen seconds to more than 10 hours, with Coding, search, images and video layered on top; 庄明浩 points to Sora 2 being shut down and domestic products limiting call volumes as evidence that supply remains tight. 孙冰洁 cites a maximum 17x gap between domestic and overseas prices and Chinese providers’ roughly 60% share of supply; low prices may amplify China’s enthusiasm, but they do not prove that providers are deliberately creating excess capacity.
- China’s OpenClaw craze is stronger because large numbers of users jumped directly from chat to action, while US users had already moved step by step through ChatGPT, Claude and Gemini toward more complex tools such as Claude Code and APIs. The leap has indeed moved mass awareness toward “letting AI do the work,” but the number of actual users may still be only several million—far from universal in a country with more than 1B internet users. “The whole internet is talking about it” may simply be an echo from a highly concentrated set of online communities.
- The secondary market has entered a phase in which industrial progress and share prices no longer move in a straight line together. Nvidia has beaten estimates for 2 consecutive quarters while its stock has barely moved, and its P/E is now below Walmart, Costco and the S&P 500 average; Microsoft at one point lost roughly one-third of its value, while storage, optical communications and other themes continued to take turns. Both statements can be true at once: “The market is always efficient, and the market is always excessive.”
- Asked whether AI names are value stocks or speculative stocks, 庄明浩’s answer was: “They are all speculative stocks; there are no value stocks.” Applications, models, cloud, data centers, GPUs and TSMC can all grow, but each also carries costs or capex; model companies’ roughly 50%–60% gross margins may not cover total costs, while Meta is directing roughly 50% of revenue to capex. The entire chain is underwriting a future that requires enormous capital and extensive prerequisites to arrive; investors can only choose their position based on their own risk tolerance.
- AI can amplify individuals and small teams, but once capability crosses a threshold it may also wipe out jobs almost instantly. Short-form drama moved from absorbing actors, screenwriters, and costume and makeup crews to AI animation and AI live-action drama in less than half a year—one example of what 庄明浩 sees as “unstoppable”; at the same time, early capital can use $50K worth of Tokens to let an individual or small team build something first and decide later whether to invest, allowing even a one-person company to become viable with AI. The workable response is not to search for a standard answer, but to “get your hands dirty” and figure out what you actually want to do and what game you are playing.
Deep dive
1. OpenClaw Adds the Product Wrapper, Not Model Capability
庄明浩 recalled that OpenClaw was formally released around the end of 2025 and only entered China’s concentrated internet discussion during the late-January-to-February Spring Festival period. In his recollection, within roughly a week of people returning from the holiday, major Chinese tech companies had begun turning the complicated installation process into one-click versions.
His central metaphor is that a large language model was originally like an engine; reasoning models such as DeepSeek R1 made the engine more powerful, while the Claude 4 series was closer to an “F1 engine.” Ordinary people cannot use an engine directly. OpenClaw provides a simplified “scaffolding car.”
This was therefore not a dramatic algorithmic breakthrough, but an architectural and engineering implementation that happened to catch the moment when model capability crossed the usability threshold. OpenClaw is merely the current flashpoint; B2B and B2C Agent implementation had already begun the previous year.
2. A Digital Employee’s Edge Comes Down to How You “Raise” It
OpenClaw puts memory, preferences, tool calls, local-file permissions and IM connections into a single architecture. The software installed by each user is the same, but scenarios, tools, permissions and usage patterns can make the final capability completely different. That is why the popular verb is not “use,” but “raise.”
If all you ask it to do is search the news, look up information or organize a report, it is not much different from ChatGPT or Doubao—and may even be worse than a chat product optimized for that specific task. The more complex the task, especially Coding and workflows involving multiple stages, deliverables and acceptance criteria, the more visible the value of the scaffolding becomes.
庄明浩 compares its initial state to a new employee at roughly the level of an ordinary college student: it does not know the company’s standards, upstream and downstream processes, databases, tools or delivery formats, and it does not understand the manager’s preferences. The user has to provide feedback, tune it and add permissions over time.
When 孙冰洁 asked how far it was from becoming a “digital employee,” the answer was not a list of capabilities but a transfer of responsibility: the ideal state does resemble an employee, but who raises it determines its ceiling. “The driver still matters. People still matter.”
3. Enterprises Have Moved Past “Should We Use It?” and Are Stuck on “How?”
OpenClaw currently appears to have more users on the consumer side, while B2B implementation will not move as quickly. But CIO-like roles inside companies have already begun reassessing data flows, internal systems and whether existing SaaS purchases are still worth maintaining.
The most aggressive change is coming from Coding. If individuals or companies can describe a requirement in natural language and have AI generate broad categories of software, the SaaS ecosystem built on software scarcity will be challenged. This is also central to how 庄明浩 understands the sharp first-half decline in US software and SaaS stocks.
Enterprise migration is still constrained by processes, data, security, privacy, permissions, KPIs and OKRs; a single installation cannot complete it. The disagreement is over timing: the aggressive view is that the market has moved from watching to scrambling, while the conservative view is that systematic implementation will still take time.
In corporate training sessions, he has observed that clients almost no longer ask whether AI “should be used.” Instead, they ask which workflow to apply it to, whether the goal is efficiency or direct revenue growth, whether to buy a standard model or a heavily customized solution, and whether to use a model vendor, a software company or build in-house. The anxiety has moved down to the operational layer.
4. Major Model Releases Have Compressed from Semiannual to Quarterly
孙冰洁 observed that ChatGPT, DeepSeek and OpenClaw all caught fire between the end of last year and the beginning of the new year. 庄明浩 believes the Spring Festival did amplify the spread of the story, but the dense release schedule also reflected a real version race among vendors.
From late January before the Spring Festival to mid-April, Anthropic released a mysterious new model and said it was so powerful that it dared not release it; the rumored GPT-6 was ready, 智谱 released GLM 5.1, and MiniMax released 2.7. Major-model cycles have compressed from 6 months to 1 quarter. If the pace continues, the “dense release period” will eventually disappear, because “every day will be a dense release period.”
5. The First Money Goes to Selling Anxiety; Open Source Turns the Wave into “Collusion”
Asked whether outsourced installation, training and hardware represented technical adoption, 庄明浩 first gave the most practical answer: “The first people to make money will definitely be the course sellers.” The same is true in the US. Media outlets package new technology as a revolution, the public worries about missing the train, and intermediaries fill the anxiety with courses, installation services or foolproof versions.
He describes the relationship as “a conspiracy” in quotation marks: once the wave has formed, cloud companies, internet platforms, model vendors, compute providers and chip companies would be wasting traffic if they stayed out. The optimal solution—even the only solution—is to “join in wherever possible.”
OpenClaw’s distinctive feature is that it is open source. ChatGPT and DeepSeek R1 belong to clearly defined commercial entities, making it difficult for outside institutions to use their logos as the main visual for a conference. OpenClaw, by contrast, allows KOLs, large companies, small companies, schools, tutoring institutions and even governments to tell their own stories. Once the snowball starts rolling, nobody can control it.
Books, courses and virtual hosts appeared within days of DeepSeek R1’s release, and OpenClaw was quickly turned into books as well. In 庄明浩’s view, its framework is not complicated enough to justify a book, but the number of products itself is a measure of the wave’s intensity.
6. Full-Permission Experiments Have Made AI Safety a Public Issue
OpenClaw was initially a tool built by an experienced programmer for himself, with the assumption that users could assess risks and set guardrails. When ordinary users suddenly poured in, that high-technical-literacy assumption failed. The original local version could even obtain full access to a computer.
Vendors had already been trying to give Agents a computer or browser, but startups and big tech companies had to account for security, cost and permissions, so they opened access only in bounded scenarios. OpenClaw set those concerns aside first and pushed the experience to its limit; the downsides became equally “exposed.”
After it went viral, versions were updated almost every 1–2 days, with many updates addressing security, privacy and permissions. Even people who had never used the product began worrying about files and data. AI safety was no longer just a narrow cybersecurity issue, but a new problem for individuals, companies, industries and governments with no ready-made answer.
7. Token Consumption Is Running Faster Than Compute Supply
In the engine-and-car metaphor, Tokens are the fuel, backed by data centers and electricity. If he had to choose one crude indicator for measuring an AI business, 庄明浩 would look at Token volume: it captures size, scale, growth rate and cash consumption at the same time.
The conclusion that “OpenClaw is merely helping model companies burn through excess Tokens” may look correct in hindsight, but he believes the ex ante reality is exactly the opposite: model companies were short of compute to begin with. AI is recreating the Moore’s-law and Andy–Bill-law dynamic—new compute gets consumed by more complex software demand.
Early chat interactions ran for only a few dozen seconds; advanced models can now execute continuously for a dozen-plus hours. Multiply that by Coding, search, image recognition and video generation, and consumption grows geometrically rather than merely reflecting longer question-and-answer text.
He cites OpenAI shutting down Sora 2 and domestic products limiting the number of calls as evidence of supply-side pressure. It is true that asking an Agent to check the weather or read a PDF can waste capacity, but “the user is not guilty.” Supply-demand distortions and positive reinforcement are happening together and cannot be reduced to a simple conspiracy.
8. China Jumped from Chat to Action, and the Hype Has Outrun the Real User Base
孙冰洁 cited data showing that the gap between domestic and overseas Token prices can reach 17x, with Chinese vendors accounting for roughly 60% of Token supply. Lower costs may make it easier for China’s ecosystem to scale, but 庄明浩 emphasizes that the key reason for the difference in enthusiasm is the divergence between the technology paths taken by mainstream US users and mass-market Chinese users.
US users had already moved gradually from ChatGPT, Claude and Gemini to more complex tools such as Claude Code and APIs. OpenClaw is more like a modest improvement for specialized personal use; B2B users may try it, but may not see it as a disruptive tool.
Large numbers of Chinese users were still operating within the conversational experience of Doubao, Kimi or Qwen. Only when Qwen began ordering milk tea and hailing cars did they feel the shift from chat to action. This wave made that transition more concentrated and more intense, pulling public awareness toward “doing things” and “doing work.”
But the denominator behind “the whole internet” needs to be unpacked. China has more than 1B internet users, while the number actually using OpenClaw through a computer or the cloud may be only several million. Several million people are enough to create the feeling that everyone is talking about it, but that is not mass adoption.
9. OpenClaw Itself Is Hard to Value; Its Ecosystem Impact Can Be Priced First
庄明浩 said the creator was quickly absorbed by OpenAI. OpenAI subsequently promised to continue supporting the open-source project in a foundation-like manner, but the creator himself did not join OpenAI. OpenClaw has no absolute parent-subsidiary relationship with OpenAI, making it difficult to calculate a narrow commercial value for the code team and software itself as one would for an ordinary company.
More important is the blurring boundary between open source and commerce. OpenClaw amplifies “one person’s value” to an extreme: the project does not generate direct revenue, but its influence, distribution, reach and ecosystem can drive hardware, Tokens, training and peripheral services.
The primary market is also beginning to accept staged valuations. A startup’s first model or product can be open-sourced and win capital recognition through influence and ecosystem effects. At some point it will still face the classic requirements of revenue, profit and cash flow—only by then, its scale may already be very large.
10. Early-Stage Capital Is Moving from Betting on One Path to Buying Multiple Attempts
庄明浩 cited a project his friend Koji is working on with ZhenFund: provide an AI individual developer or small team with $50K worth of Tokens, let them build something, and then decide whether to invest. Even a short conversation with the most expensive models can cost a fraction of a dollar, making Tokens a form of upfront incubation capital.
AI allows individuals or small teams to contemplate projects that previously required a medium-sized team, while medium-sized teams can attempt work that once required a large organization. Early-stage capital has not lost its relevance, but the funding needed to get started has fallen, allowing investors to support more experiments.
A single angel check might now let the same team try 6 times: spend a few months on one direction, switch if it does not work, raise more capital after receiving staged feedback, and potentially keep switching even after funding arrives. Low startup costs are changing the old constraint of “raise once, execute one plan.”
He jokes that the market will produce more “projects that die in the sunlight”: founders with strong résumés first raise tens of millions of yuan, then repeatedly launch, fail and pivot. The classic TMT path of scaling around one story from angel to Series A to Series B may no longer be the norm in AI entrepreneurship.
11. The AI Hardware Consensus and the “Outsider Genius” Narrative Can Both Be True
The biggest consensus in China’s primary market over the past year has been AI hardware. Pure software monetization and the B2B ecosystem are weaker than in the US, while most large-model investment has already been made. Supply chains, Shenzhen’s Huaqiangbei electronics market and the success of DJI, Pinduoduo? Wait no, Ta? Let’s retain names: DJI, Ta? Source: 大疆、拓竹、Insta360、小米 = DJI, Bambu Lab, Insta360 and Xiaomi. These examples naturally directed capital toward hardware.
“Ready” founders with big-tech experience were therefore chased, regardless of whether the product was a toy, a plush device, a display, an educational machine, a camera or a video camera. The group effect has a complete industrial logic, and 庄明浩 does not believe it is inherently wrong.
In her questioning, 孙冰洁 used the creators of breakout products such as OpenClaw, Manus and Cursor as examples of the “outsider genius” narrative. 庄明浩 believes the new battlefield genuinely gives individuals who were previously outside VC range more opportunities. The mobile internet had largely solidified after 2015 and 2016, but AI is once again amplifying individual capability. The hardware camp and the individual-developer camp may produce winners at the same time.
12. People Inside a Bubble Cannot Know Where They Are on the Curve
庄明浩 has a vague concern that the AI sector could end up like Web3 today: high valuations, heavy fundraising, firm technical conviction and a grand future ultimately turning into a small, “too highbrow for the audience” niche. 孙冰洁 explicitly pushed back, arguing that AI’s reach and mass acceptance are far greater than Web3’s and that the momentum is not even in the same order of magnitude.
His own position is closer to that rebuttal: AI is different from Web3 and the wave is still advancing. But he retains a mild concern because “when you are living through the present, you do not know which stage of the bubble you are in.” The curve can only be drawn 10 or 20 years later.
Market signals are equally contradictory. Tech giants broadly fell in 2026, Microsoft at one point lost roughly one-third of its value and may have been going through one of the worst quarters in its history; Nvidia beat estimates for 2 consecutive quarters while its stock barely moved. The apparent outcomes are either “a big move is brewing” or further downside, but nobody can confirm which in advance.
He once wondered whether another 50% decline in Nvidia would mark the end of the cycle, then found that its P/E had already fallen from roughly 40x at the peak to the teens, below 20x—effectively close to a halving. A single price milestone therefore cannot mark the end of a bubble; the adjustment may already be underway.
13. The Secondary Market Will First Lift Each Other Up, Then Make Up for the Decline Together
The US valuation framework still revolves around the KPIs of leading companies, changes at cloud vendors, the valuations of 2 companies preparing to list and ARR multiples for startups. Software and SaaS are in a major adjustment phase; China’s market lacks clear-cut high-quality targets, making it prone to pushing anxiety and thematic premiums onto companies that should not receive them.
A-shares and Hong Kong stocks may be a different story. “Pure-play” large-model companies such as 智谱 and MiniMax have posted huge gains, while companies with substantial AI revenue and respectable capabilities may still suffer sharp declines. As long as the massive capex premise remains intact, optical communications, optical transmission and even deeper-layer components can take over after memory.
庄明浩 uses the fact that Walmart’s P/E at one point fell below that of a contract manufacturer, while Nvidia’s fell below Walmart’s and Apple’s, to show that pricing can overshoot in either direction: “The market is always efficient; the market is always excessive.” Efficient means it will eventually correct; excessive means it must first cross the line.
智谱 rose on a new model. As long as the dominant sentiment remains intact, tomorrow’s OpenClaw, OpenCat or OpenBear could also become fuel, like “Tiyunzong,” the cloud-stepping technique in Jin Yong novels—using the left foot to step on the right and keep climbing. Once the dominant sentiment turns, the same logic becomes a “ladder-down” for cascading and catch-up declines. If forced to choose 0 or 1, he believes sentiment is still positive, but nobody knows when the turn will come.
14. So-Called Value Stocks Are Also Betting on an Extremely Expensive Future
Asked how to distinguish AI value stocks from speculative stocks, 庄明浩 answered: “They are all speculative stocks; there are no value stocks.” Investors buy applications, applications run on models, models rely on the cloud, the cloud sits in data centers, data centers are filled with GPUs and GPUs are manufactured by TSMC. The entire chain is betting that demand will continue to rise.
Application companies have to pay model and cloud vendors, so gross margins may be very low or even negative and require investment capital to keep them alive. Model companies must both serve users and train the next generation; even 50%–60% gross margins may not cover total costs.
Cloud companies, data centers, Nvidia and TSMC look more profitable, but they are also constantly expanding capex. Even TSMC, known for its discipline, is increasing investment. Its chairman has said the company surveyed customers and “customers’ customers” to confirm AI demand. That may strengthen confidence, but it cannot eliminate cyclical risk.
The real danger is revenue growth accompanied by a rapid decline in cash flow. Meta has no cloud business, yet invests roughly 50% of revenue in capex. Every technology revolution goes through a period of infrastructure overinvestment, but today’s constraints in memory, advanced packaging, power, optical transmission and land approvals make the endpoint of excess impossible to calculate.
15. Revenue Growth Cannot Rescue the Replacement Narrative, and Jobs Will Not Wait for Earnings Confirmation
The secondary market sets different tests based on a company’s stage. A newly listed model company may deliver double-digit-multiple growth in its first full fiscal year, but must show a new move the following year. Traditional implementation or data-center companies must use AI to lift gross margins by several points, while chip and foundry companies must make orders cross an entire order of magnitude.
Oracle is the classic example in 庄明浩’s view. This loser of the cloud era surged after winning a major OpenAI order, then fell back after the market realized that sacrificing gross margin to win orders could weigh on cash flow over the next 3 years. “It spent all of last year running in place.”
Software and SaaS companies can see both AI revenue and total revenue grow and still fail to overcome the entrenched logic that “AI will replace you.” Until that sentiment reverses, positive earnings feedback may not change the direction of valuation.
Job displacement is more abrupt. AI researchers, standing at the capability frontier, may be more worried than anyone else about being replaced by their own tools. Short-form drama had just revived actors, screenwriters, directors, costume and makeup crews and Hengdian, but once text, storyboarding, video and synchronized audio-visual capabilities became ready, the industry shifted rapidly to AI animation and AI live-action drama within less than half a year—“Bang. It swept everything away.”
16. A One-Person Company Has Leverage, but It Does Not Automatically Create a Business
If forced to choose, 庄明浩 is relatively pessimistic about the “one-person company.” AI, supply chains and the internet can indeed expand an individual’s boundaries without limit, but whether the real economy offers enough room to survive is a separate question. Being able to establish an entity does not mean the work has value or can support the founder.
The visible opportunities are still concentrated in a small number of service sectors, including self-media, course selling, teaching, graphic design and handmade-brand operators, with limited impact on the primary and secondary industries. 孙冰洁’s pushback is worth retaining: “How much hope for flexible employment are you shutting down with that way of thinking?”
郝景芳’s team offers a more realistic organizational model: a single integrated company of several dozen people was split into more than 10 atomic units of 4 or 5 people. Local teams can handle the core IP, educational courses and offline implementation, while also taking on other business; the course R&D team can likewise develop and sell other courses.
Here, AI functions more like a lubricant, reducing the coordination costs of course development, private-domain operations, sales and class organization. Activities that previously had to remain bound inside one company can be separated more easily. AI improves organizational boundaries; it does not replace the underlying business demand.
17. AI Anxiety Ultimately Lands on Mind, Body and Spirit—and Education
庄明浩 believes discussions about AI will most likely end up at 2 questions: mind, body and spirit, and children’s education. Both are fundamentally expressions of anxiety about the future life of oneself and the next generation. Technology experts talking about parenting in the AI era may attract clicks, but cannot provide a standard answer because everyone is living through the present.
His “honest non-answer” is: “Anxiety is the norm.” The response is not to carve out long blocks of time for intensive study, but to “get your hands dirty”—put models into daily life, perhaps by photographing meals and having AI count the calories, and experience the pleasure of building and playing. If it does not work, step back and try again later.
A summer-camp exercise asks children: “If you had no homework today, no extracurricular classes, and your parents made no demands at all, what would you want to do today?” Many children cannot answer. As AI moves closer to AGI, knowing what one wants and where to obtain spiritual satisfaction and positive feedback may matter more than mastering any particular tool.
OpenClaw itself is an example of this logic. After achieving financial freedom, its creator simply wanted to control 2 computers at home from WeChat or WhatsApp on his phone while away. He waited more than half a year for someone else to build it, then did it himself. A real, specific personal need that nobody had addressed eventually snowballed into a global wave.
18. The Next Stop Is Autonomous Evolution; Participants Can Only Step In and Out at the Same Time
The industry consensus formed at the end of last year was that the next stage for Agents would be autonomous evolution. The harnesses and environments built over the past few months are reducing human intervention and letting AI correct itself, verify each part of a task and search for solutions. Coding is close to ready; complex real-world tasks are not.
庄明浩 cites OpenAI’s 5-stage framework: L1 is Chatbot, L2 is reasoning, both of which have been crossed; L3 is Agent, which is underway. L4, the “researcher,” means AI can perform the work of an OpenAI researcher, while L5, the “organizer,” turns powerful individuals into an organized entity. AI for Science is already being discussed as a way for AI to handle researchers’ work on drugs, proteins and other fields, making L4 appear no longer out of reach.
Progress will not be linear. After capability surges forward, it will collide with the interests, regulation, policy, safety, privacy and permissions of the real world, then pull back, rebalance and negotiate before moving forward again. It is difficult even to define when autonomous evolution is complete; the threshold may be crossed before society has given it a name.
For entrepreneurship and investment, he offers principles rather than price targets. He believes the primary market can only be approached from the long side, while the secondary market must match an individual’s temperament and risk tolerance. First understand “what kind of game you are playing,” watch how strong participants make decisions, and recalibrate between “repeatedly stepping away” and “repeatedly stepping in.” The AI theme still appears likely to continue, but nobody can remain untouched and nobody can mark the cycle’s position in real time.