2026: AI’s Breakneck Run, Capital Drain and Our “Engels Pause” — A Crossover with Jinbo Finance
Summary
庄明浩’s central judgment for mid-2026 is that models have cleared the intelligence threshold for most white-collar work, making Agent adoption unstoppable. The competitive center is shifting from pure Chat products such as 豆包, 千问, 元宝 and ChatGPT to OpenClaw, Coding Agent and general-purpose Agent products. Products such as 腾讯云’s WorkBuddy Agent are already reaching beyond technical users to HR teams, finance professionals and small-shop owners. By year-end, he expects China’s task-oriented AI to move from millions of tech enthusiasts and coders to tens of millions or more. “The Chatbot battle is basically no longer worth watching.”
Agents are turning Token demand from linear growth into geometric growth and pushing the raw materials fed to the silicon-based world toward physical limits. A single task may run 100 agents in parallel, while advanced models can work continuously for 10-plus or 20 hours. Compared with the 20-30-second conversations of the past, concurrency and duration alone could produce several orders of magnitude more demand.庄明浩 believes that, on a linear extrapolation, this “may only just be getting started.”
The market’s question for AI has moved from when model companies will turn profitable to what ROI and GDP Token consumption is actually creating. As leading Chinese and US model companies scale geometrically, their path to profitability may already be measurable. The next question is what applications, implementation work and To B SaaS products deliver after consuming vast quantities of Token. 沈帅波 says silicon investment has yet to visibly benefit the carbon-based world. 庄明浩 concedes there is still no perfect answer, but says that each layer of progress lifts “everyone with the tide.”
AI assets remain in an uptrend with expanding volatility, but 庄明浩 rejects equating “unstoppable” with a recommendation to chase prices. Nvidia spent last year rangebound around $180-$200 despite beating expectations for 3 straight quarters. After a Q1 pullback, new factors lifted it to a higher plateau, with the 2026 average price now holding around $220-$240. 沈帅波 says only “buy and never leave” avoids being thrown off the train. 庄明浩 says he is both lazy and fearful: “It’s like surfing—you’ve already been thrown off several times, and if you chase it again, you’ll get killed.”
Hong Kong-listed AI companies face a three-way squeeze from lockup expiries, valuation and global liquidity; positive product milestones may offset the risks but cannot eliminate them. The key dates are the first wave of cornerstone-investor lockups expiring in early July and full circulation in January of the following year. More than 50%, and possibly around 60%, of Hong Kong-listed companies have market caps below HK$500M, while new listings are creating another concentrated wave of unlocks. By one comparison cited on the show, OpenAI may already be a trillion-dollar company with ARR of $40B-$50B or more, implying roughly 20x sales; some Chinese Hong Kong-listed model companies trade at triple-digit ARR multiples—“a 90% drop would only bring them back to double digits.”
The real risk is debt and capital siphoning, not simply high equity valuations. 庄明浩 cites roughly $80B of financing for SpaceX, potentially several tens of billions to $100B absorbed by OpenAI, and debt issuance by Google, Nvidia and Meta. Those giants alone could pull “several hundred billion dollars” from the market. Nvidia’s first bond issue in years was around $20B, while Google also issued equity for the first time in years. His conclusion is blunt: “The stock may still be fine…but the debt is a problem,” because debt carries hard obligations and transmits stress through the system.
The technology agenda for the second half has narrowed to “autonomous evolution” and “world models,” while the hardware agenda remains a push to break through supply and physical constraints. Harness is like putting a chassis, steering wheel and drivetrain on an F1 engine. Loop goes further, laying roads, building gas stations, traffic lights and traffic rules so models can assign tasks, call tools, evaluate results and replan on their own. World models should in turn be split into renderers for human-facing output, planners that follow physical rules and simulators that understand composite world knowledge—three layers corresponding to different paths across video, games, robotics and autonomous driving.
AI may give this generation a new “Engels Pause”: productivity and GDP can rise without a synchronized increase in income, occupational identity or a sense of meaning. 庄明浩 uses the roughly 40-year stagnation in textile workers’ wages during the steam-engine era as an analogy, and argues that professional skills in presentations, analysis and content production are rapidly becoming commodities. Handcrafted essays, PPTs and even code may eventually become niche arts, much like vinyl records. What remains valuable will not be stronger standardized expertise but emotion, taste and courage. A friend’s answer to him was: “AI isn’t as interesting as you are.”
Deep dive
1. Models Clear the White-Collar Benchmark as Agents Enter the “Unstoppable” Phase
庄明浩 used the title of his Spring Festival PPT to sum up the first half of the year: “It can’t be stopped.” Models appear to have crossed the intelligence threshold required for most white-collar work; once that threshold is cleared, displacement and employment anxiety accelerate alongside capability.
The battlefield has shifted from pure Chat products such as 豆包, 千问, 元宝 and ChatGPT to OpenClaw at the start of the year, fast-growing Coding Agent products and general-purpose Agents. The key change is not another product launch, but Agents breaking out of the tech-enthusiast niche.
He pointed to a 腾讯云 conference as an example. After the company’s负责人 Dawson spoke with 姚舜宇, the event immediately followed with a demonstration of WorkBuddy Agent. Its actual scale may be far larger than outsiders assume, with users expanding from technical staff to HR, finance and small-shop owners.
沈帅波’s observation is that Agent was still a novel, niche term last year but has become broadly adopted in some circles this year. 庄明浩 takes the forecast further, targeting tens of millions of mainstream users by year-end rather than another fight for the Chatbot leaderboard.
2. Token Consumption Shifts from Linear to Geometric Growth as the Physical World Sets the Ceiling
In the pure-Chat phase, Token growth came mainly from larger models, longer conversations and the shift from pretraining toward post-training and reinforcement learning. Agents add task decomposition, multi-Agent clusters and other layers, creating an entirely different growth structure.
沈帅波 gives the example of a complex task running 100 agents simultaneously. A single thread becomes multi-Agent concurrency, while advanced models can work on their own for 10-plus or 20 hours. Compared with conversations that ended after 20-30 seconds, duration alone creates a multiple-of-dozens increase.
That is why every raw material “fed to the silicon-based world” is rising in demand while running into the limits of capital, the physical world, production lines, time and even the Earth itself. 庄明浩 keeps the claim conditional: “Purely from the standpoint of linear extrapolation,” demand is entering geometric growth, and it “looks like it has only just begun.”
3. Token Economics Moves the Question from Model Profitability to Real-World ROI
沈帅波 identifies an unresolved contradiction: silicon investment has become so large, yet has not directly translated into broad improvements in the carbon-based world. “How does this cycle keep going?”
庄明浩 believes the market has moved on to a new question. It used to focus on model companies’ revenue and profitability. Now, as leading Chinese and US vendors scale geometrically, their path to profitability may already be measurable. The question has moved forward: “You use this many Token every day and complete all these tasks—what have you actually produced?”
Token economics therefore comes down to applications, implementation and To B SaaS: does consumption produce higher ROI and GDP? He does not claim the problem is solved. He says that once part of the case is established, the market will ask the next-layer question—but with every layer of progress, “everyone rises with the tide.”
4. The Trend Has Not Slowed, but Pricing Has Entered the “Killing Zone”
Last year, the market already believed that capital, the physical world, production lines, time and even “the limits of the Earth” would constrain AI. Yet 2026 has brought no visible slowdown, and many indicators are accelerating. 庄明浩 says that once people cross the threshold of imagination, it becomes difficult to model what the world will look like after another 2 or 3 years of consecutive gains.
沈帅波 summarizes the trade as follows: “There is only one kind of person who can make money—the one who buys and never leaves.” Any exit risks being left behind; when investors look back, the slope is still rising and the explanation is usually something found after the fact.
Nvidia traded around $180-$200 last year. Its earnings beat expectations for 3 straight quarters without moving the stock. After a Q1 correction, new factors pushed the market to a higher level, and the 2026 average price has settled around $220-$240. The old reasons for a pullback have not disappeared; they have simply been overwhelmed by new weights.
沈帅波 questions why 庄明浩 says the trend cannot be stopped while refusing to participate. His answer: he is lazy, timid and worried that volatility will distort his other judgments. A one-day move of 3%-5% in an index, 7%-8% volatility in an ETF and a roughly 30% move in one stock within 30 minutes all add up, in his view, to “a killing line.”
5. Lockup Expiries and Giant Fundraising Make Liquidity the Primary Variable for Hong Kong AI
庄明浩 has previously flagged 2 dates: the first wave of cornerstone-investor lockups expiring in early July and full circulation in January of the following year. A linear extrapolation from history points to a major negative catalyst, but sufficiently strong positive milestones from representative companies could offset part of the lockup pressure.
Hong Kong saw a large number of new listings last year and will naturally face concentrated unlocks this year. At the same time, more than 50%, and possibly around 60%, of companies have market caps below the HK$500M minimum line; most are so-called penny stocks. If US inflation or interest rates deteriorate again, a global liquidity contraction would amplify the shock.
The bigger variable is giant-company financing: SpaceX has raised around $80B, OpenAI could absorb several tens of billions to $100B, and Google, Nvidia and Meta have all followed with bond issuance. Together, they could pull “several hundred billion dollars” from the market.
Debt is what particularly concerns 庄明浩: “The stock may still be fine—at least in equities, everyone accepts the wager—but debt is a problem.” Nvidia issued around $20B in bonds for the first time in years, while Google also issued equity for the first time in years. The amount may not be fatal; the shift in posture is what warrants concern.
6. Hong Kong AI Valuations Lack a Firewall; A-Share Listings Seek a Different Narrative
The show cites a comparison from a friend: OpenAI may already be a trillion-dollar company, with ARR of around $40B-$50B or more, implying roughly 20x sales. Traditional industries would call that expensive, but it is not outrageous for a technology company.
By contrast, some Chinese Hong Kong-listed model companies may trade at triple-digit ARR multiples, with slower revenue growth than Anthropic and revenue bases several zeros smaller. The sharp conclusion is that “a 90% drop would still be reasonable,” because only a 90% fall would bring the multiple back into double digits.
Hong Kong and US valuations are more likely to converge, which helps explain the urgency among these companies to seek A-share listings: the A-share market may provide a narrative separation from US equities. 庄明浩 offers no one-way trading conclusion and instead asks investors to assign their own weights to macro conditions, the industry, lockups and company milestones. “Accept the wager.”
7. Chinese Model Companies Can “Take a Sip,” but 豆包 Is Being Forced to Charge by Its Scale
The US and Chinese monetization paths may continue to diverge. The US is more comfortable with To B and subscriptions, while China has long relied on advertising or cross-subsidization—“the wool comes from the pig.” But after OpenClaw’s rise, Chinese users are becoming more willing to buy Token, and cloud and model vendors are seeing demand for their newest models outstrip supply.
For independent model companies, revenue of only a few hundred million yuan last year could grow into the low single-digit billions of yuan this year without requiring extremely high penetration. “The biggest guy eats the meat; I’ll have a sip of the soup.” The market is large and the soup is thick enough that smaller companies may meet their growth needs simply by executing well on existing products.
豆包 is different. Its user base is already so large that not charging could become a problem. Even a “money-printing machine” like 字节 must consider sustainability. 庄明浩 believes 豆包 has to be the first to test the waters, but the market remains skeptical that it will formally charge in the second half.
Of the 2 pricing scenarios, 豆包 appears more inclined toward a professional subscription: more PPTs, images and deeper Deep Research, rather than companionship, emotional support and other “interesting but useless” use cases. 沈帅波’s practical summary is: “The things you want to pay for, it won’t let you pay for; the things you don’t want to pay for, it insists on charging you for.” 沈帅波 also mentioned that the latest Claude 5 has been banned in the US.
8. Autonomous Evolution Is the Core Theme; Harness and Loop Are Engineering That Gets Models Running
From AI for Science and AI for Medicine to Harness, Loop, new labs, new teams and multi-Agent clusters, leading teams are converging on the same objective: have models assign and allocate tasks, call tools, evaluate interim results and then revise their earlier plans, creating an autonomous loop.
庄明浩 compares it with the tractor engine used in the popcorn machines of his childhood in Northeast China. A large model has gone from a broken engine to an F1 engine, but a bare engine remains of limited use. Harness supplies the steering wheel, brakes, wheels, driveshaft, cockpit and chassis—turning the engine into a complete car.
Loop is the external system: roads, gas stations, streetlights, traffic lights, overpasses and traffic controls. Once the base capability is ready, the model still needs the surrounding identity, processes, division of labor, environment and rules. The ultimate objective remains “to get it running on its own.”
9. A World Model Is Not One Product but 3 Different Businesses
The world model has become an umbrella term. Video, AI 3D, games, autonomous driving, automotive and robotics companies all claim to be building one. The common premise is that language alone cannot cover visual and physical knowledge such as terrain, collisions, breakage and fluid motion.
庄明浩 borrows 李飞飞’s taxonomy. The first category is the renderer: a system that can generate and extend an unlimited stream of images for the eye, primarily serving video, games and 3D. The second is the planner: it knows that a table pushed backward will topple according to physical rules, making it more relevant to robotics and autonomous-driving decisions.
The third is the simulator. It must know not only that a cup will fall, but when glass will break, how water will spill and how fast the liquid will flow, along with other composite knowledge. 李飞飞 defines her own work at this third layer. 庄明浩 believes the classification clarifies that different companies may cover only 1 or 2 of the layers.
10. Compute Demand Is Repricing “Old IT,” and History Has Become a Liability
The hardware story is uncomplicated: expand supply and keep pushing down toward the underlying constraints. Names that performed particularly well over the past 1 or 2 quarters include SanDisk, Micron, Intel, AMD, Dell and HP. The market’s reaction: “What year is this?”
For people who lived through the PC era, these companies still carry the old valuation imprint of computers, racks and USB drives. Their business models, gross margins and demand profiles spent years moving lower. When they suddenly reverse, “having relatively more knowledge and experience becomes a huge liability.”
Intel is said to have broken above its peak from the 2000 internet bubble, after spending the past 26 years in a deep trough before suddenly taking off this year. 沈帅波 jokes that less than a year ago the show was discussing how Intel was being left behind by Nvidia and how 黄仁勋’s personal stake was worth more than Intel. The story has now flipped.
11. The US-China AI Ecosystems Diverge on Acceleration vs. Pushback, Openness vs. Control
庄明浩’s experience in the US is sharply bifurcated. Silicon Valley is still “charging ahead,” while mainstream society may be strongly resistant. He says a poll could show opposition among ordinary people reaching 70%; games, media and consulting reports can all trigger backlash simply by using AI.
Data-center delays, employment and other nontechnical conflicts cannot be solved solely with stronger models or commercial tactics. Several giants are “running red lights by every possible means,” and Claude 5 being banned is, in his view, an extreme example of the conflict: repeatedly stressing that the model is powerful and should be controlled eventually became a self-fulfilling prophecy.
Chinese users are more proactive about learning, while open-source model companies have found a comfortable position between capital, users and enterprise customers. When 智谱 released GLM-5.2, it emphasized openness and contributing to “the upper limit of intelligence for all humanity.” 庄明浩 believes the roles of openness and compulsory control have emerged in a way that reverses the traditional assumption.
12. Capital Prefers AI Native Because Good Businesses Inside Conglomerates Cannot Easily Move the Group Valuation
沈帅波 points out that even solid incremental innovation from a major company rarely moves public opinion or the stock. A small move by an unlisted or AI Native company, by contrast, gets described as a journey to the stars and beyond. 庄明浩’s explanation: “One side is poetry and distant horizons; the other is the grim reality in front of you.” Capital naturally prefers the new.
A new company has only one thing to point to, so its performance can define the entire story. The same business inside a giant group may contribute so little in revenue, users and influence that it does not even reach the tail end of 1%, making a re-rating impossible.
The deeper market belief is that existing organizations, successful experience, inertia and path dependence all become burdens on new-technology development. The balance therefore tilts decisively toward new organizations. 庄明浩 retains some skepticism: “Whether it must ultimately work this way, I’m not sure.”
The discount even applies to previous-generation AI companies such as 商汤 and 百度. Their cloud revenue and model iterations may not be poor, but the market still refuses to pay. Talent leaving these companies to start new ventures only reinforces the loop: the market concludes that only new companies can execute new technology well.
13. Visual Attention Is Fully Allocated; AI Hardware Can Only Experiment at the Margins
沈帅波 challenges the narrative around no-interface AI prototypes. Internet products have spent nearly 20 years being culled and have already captured the highest-frequency needs. Sometimes people simply want to scroll, compare and browse rather than receive the optimal answer immediately. After testing one device, he at one point had 3 or 4 receivers hanging around his neck, like a “human data collector.”
庄明浩 responds with a rough time budget. If 1 billion people in China each look at or use a screen for 10 hours a day, that is 10 billion hours. WeChat and Douyin each take roughly 20%, Hongguo takes 5%-10%, and Kuaishou takes 3-4 percentage points. The visual battlefield has “no gaps left.”
LUI may capture some time in task and work settings, but it cannot broadly beat Douyin and short video. Podcasts and voice recorders can grow precisely because they use hearing, occupying a “second battlefield” outside the visual front.
AI therefore has to explore new hardware and new forms of interaction, perhaps even designing products for AI and the silicon-based world itself. At this stage, each product is simply “choosing one path and one implementation,” which requires a large number of casualties. 庄明浩 does not rule out looking back 20 years from now and concluding that “this entire wave was just cannon fodder.”
14. Technology Will Crush Sentiment, but Handcraft May Become Art in the New Equilibrium
沈帅波 asks whether insisting on doing things by hand is merely “preciousness” and meaningless resistance. 庄明浩 first accepts the technological determinist case: even if OpenAI, Anthropic, Google and Chinese private companies stopped, another group would push forward. As the saying goes, “Technology crushes sentiment every time. Every time.”
After AI models, Agents and the various wrappers, a typical PPT may leave only roughly 5% for humans to handle. Most situations only require a 90-point result. Further tweaking of fonts and layouts has declining practical value; what remains is “a tiny difference in beauty, a tiny bit of fun.”
He describes film and speech as “the three deaths of the content industry and the creator’s rebirth”: the material dies, the process dies, and the business model and copyright die as well. But death does not mean expression stops. What becomes scarce is what one wants to express, whose recognition one seeks and whether it can create an emotional response.
Vinyl records rose against the trend in 2025 after a long decline, showing how craft can move from a mainstream tool to a niche demand. Handcrafted essays, PPTs and even code may become forms of art in the future rather than efficient means of production.
15. After the “Engels Pause,” Expertise Depreciates and Interest and Courage Become the Residual Assets
庄明浩 uses the “Engels Pause” of the steam-engine era to describe the present. National GDP rose rapidly while textile workers’ wages failed to increase for roughly 40 years. This generation may face its own “Engels Pause.” Society may find new occupations and a new equilibrium decades from now, but today’s workers are the ones spending the core decades of their lives absorbing the transition.
He offers another angle: once machines replace many jobs, those activities become entertainment. Large-scale farming and greenhouse cultivation displaced some manual planting, while growing vegetables in the yard became a hobby. People once complained that tools had turned hobbies into work. Now “de-instrumentalization” may mean losing the livelihood, but it can also restore voluntary expression.
Professional investing, research and podcasting are being flattened in the same way. Models can already organize the latest arguments for and against storage into a male-female dialogue that is “80%-90%” comparable to what a professional analyst would deliver. 庄明浩 is therefore no longer trying to be “the most professional investor,” but “the most interesting investor.” His friend’s line—“AI isn’t as interesting as you are”—has become the new positioning.
By year-end, he expects China’s task-oriented AI to expand from millions of users to tens of millions or more, including parents and other nontechnical groups. In the US, the autonomous-evolution path should produce companies with concrete use cases, process-level results or interim engineering milestones. Faced with the destruction of professional identity, his final answer is not a fixed plan but “a brave heart”: “Otherwise, won’t you just break right there?”