AI's Frontrunners Are Just Refineries, Not the Native-App Winners
AI has entered its agent phase: the first-stage leaders are only refineries, not the native-app winners; new platforms don't come from old platforms; organizations must shift from jobs to tasks; and in the end what people compete on is the creativity to conjure something from nothing.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Building agents today is like building websites in 1992
Zeng Ming divides the industrialization of a general-purpose technology into three stages: first it becomes social infrastructure, then applications explode, and finally native applications appear. In 2026, token as a unit of measure reaching consensus shows the first stage has matured; the agent events around Chinese New Year (commonly called "lobster") mark AI's formal entry into the agent stage. He draws the analogy to the 1992 website-building movement: people then built websites to share information, people today build agents to share capabilities. The browser hasn't appeared yet, and whoever can establish agent standards and seize the user entry point may hold the biggest opportunity of the next two years.
— Zeng MingA ten-trillion-dollar company will appear, but not these two
Zeng Ming says plainly that OpenAI and Anthropic are remarkable technological and commercial achievements, but business history shows the first wave of breakout companies is most likely not the big winners of the native-application stage. Yahoo reached a $120 billion market cap five years after founding, AOL was valued at $220 billion, and neither became the Google of the next era. Model companies are more like AI cloud companies, refineries that refine the base oil, while the real application value belongs to chemical plants and car companies — and cars were not built by oil companies. He believes a ten-trillion-dollar company will certainly exist, but not necessarily the two currently in the lead. Model companies will move toward a mature business of oligopoly plus heavy government regulation.
— Zeng MingThe unit of an AI-native organization is the task, not the job
The company corresponds to the industrial-age hierarchy; AI overturns the industrial age, and the corporate system will wither. The basic unit of an AI-native organization is not the job but the task: people are not organized by reporting lines but flow around "what needs to be solved," and humans and AI coordinate around tasks. So well-designed jobs and levels will disappear, and the traditional middle layer will die too; future founders must first define the organization's task network. Silicon Valley's NewLab is a prototype of the new organizational form, and OpenAI itself is more like a Lab than a traditional company. The culture of the new organization is more like a sports team, emphasizing transparency, sharing, and co-creation, no longer relying on commands and performance reviews.
— Zeng MingExcellence can only be judged afterward; being good is visible on the spot
Zeng Ming argues that "good ≠ great." Being good is growth that keeps getting positive feedback, while greatness requires overcoming negative feedback again and again, persisting when everyone disapproves, and finally proving through enormous success that you were the era's prophet — so "greatness is judged in hindsight." Good people easily fall into a loop of proving themselves, which makes their strategies conservative. Great people have a small ego, tend to attribute success to the times and their partners, and are driven by mission rather than external valuation. When identifying founders he asks, "Ten years from now, what kind of self would you be satisfied with?" If the answer is "a ten-billion-dollar company," that is mainly ego-driven, not an inner mission. Truly great founders tend to be altruistic and empathetic.
— Zeng MingStrategy in the AI era can't be planned, only generated
Strategy in the AI era is not planning but generation. Zeng Ming proposes a "strategy generation system": the organization must build an environment and network in which strategic insight emerges naturally. Because decision frequency is higher and quality demands are higher, a CEO cannot plan by linear extrapolation, nor can strategy be outsourced to McKinsey — McKinsey's explosion was a product of the slow maturation of the industrial age. His methodology is "look ten years out, think three years out, act one year out," with the core being "think three years out": you must at least see two or three milestones ahead, creating tension between short-term action and the mid-term picture. Organization-building must ensure enough context rather than control, letting the right people make the right decisions at the right time.
— Zeng MingThe right analogy for robots is appliances, not cars
Zeng Ming believes robotics is still in a strategic exploration phase, far from convergence. Of the two current paths — build a general brain first or close the loop on scenarios first — both are logically valid, and it's impossible to judge in advance which is faster. By historical analogy, the auto industry had thousands of companies from 1900 to 1920, and only entered scale when Ford built the assembly line in 1913; robots have yet to reach their Model T moment, and whoever truly sells ten thousand robots first is already remarkable. But the better analogy for robots is appliances rather than cars: after electricity was invented came countless categories — refrigerators, washing machines, air conditioners — and robots interacting with the physical world will have N kinds of big scenarios, with home, companionship, and industrial all potentially producing their own native giants.
— Zeng MingThe incumbents who feel safe are the least safe
Zeng Ming says that before a giant wave no one is safe, and feeling safe is the most unsafe thing of all. In business history, no era's companies have ever transitioned smoothly into the next era; IBM and Microsoft are the only two possible exceptions, and both are full of uncertainty. AI is a productivity revolution that overturns the Industrial Revolution, not a continuity innovation like mobile internet, so old giants' technical reserves and organizational culture are negative assets. Google can live on as an AI cloud company but may not capture the consumer entry point; ByteDance has an AI cloud opportunity, but Doubao is not the future application; Tencent's social relationships will be restructured; Alibaba is likewise not naturally safe. The new platform will most likely be created by a new company.
— Zeng MingHuman value lies in conjuring from nothing, not writing songs or painting
Borrowing Drucker's framework, Zeng Ming defines the AI era as the era of creativity: all knowledge work that can be structured will be taken over by AI, and people are on one hand forced and on the other finally liberated to develop new potential. That potential is called creativity — not writing songs or painting, but the ability to originally define complex problems and conjure something from nothing. AI has already taken away all existing knowledge, and human value lies in creating what does not yet exist. So he calls his own mission "to be a happy researcher," and believes civilization will next reorganize around creativity, with even the education system forced to shift from "knowledge indoctrination" to "open exploration," because young people already know the old path leads nowhere.
— Zeng MingIn their own words · checked verbatim
The most successful companies are always contrarian, but being contrarian doesn't guarantee success.
最成功的企业一定是反共识的,但反共识不一定成功。
Zeng Ming7:19
It may not survive to the end. It will probably do quite well going forward, but it is most likely not the big player, the big winner, of the native-application stage.
它不一定能活到最后。它很可能会往后活得不错,但它大概率不是原生应用阶段的大玩家,大赢家。
Zeng Ming27:49
First-stage companies rarely survive into the second stage, and second-stage companies rarely survive into the third stage.
第一阶段的企业很难活到第二阶段,第二阶段的企业很难活到第三阶段。
Zeng Ming29:50
So companies will disappear, companies will die out. There's nothing to regret about that. There will be more fun things in the future — how many people are happy going to work at a company now? Not happy.
所以公司会消失,公司会消亡。这没有什么可以遗憾的呀。将来有更好玩的,你说多少人现在去公司上班是开心的?不开心。
Zeng Ming1:15:33
My technology is so good, I've created so much value, why might my company be worth nothing? Because of the basic laws of economics: as long as there is homogeneous supply, you cannot capture high profits, and you are a company that isn't worth much.
我的技术这么牛,我创造了这么大的价值,为什么我公司可能不值钱?因为经济学的基本规律,只要有同质化的供给,你就不可能获取高额利润,你就是一家不怎么值钱的公司。
Zeng Ming1:28:01
Greatness is judged in hindsight, because you overcame negative feedback again and again, and when everyone disapproved of you, you finally proved through enormous success that you were the era's earliest prophet.
卓越是事后论定的,是因为你克服了一次一次的负反馈,让所有人都不认可你的时候,你最后会以巨大的成功证明,你是时代最早的那个先知
Zeng Ming1:31:32
I have always felt that before a giant wave no one is safe, and whoever feels safe is the least safe.
我从来觉得在巨浪面前没有人是安全的,谁觉得安全,谁就是最不安全的
Zeng Ming2:02:39
Figures
| K3 model parameter scale | 2.8T | 35:58 |
| Years since OpenAI's founding (as of recording) | 11 years | 2:17:47 |
Glossary
- Token factory
- Treating AI compute as infrastructure for producing tokens, symbolizing intelligence's entry into a stage of standardized pricing.
- Agentic OS
- An OS that can directly dispatch all agents to complete tasks by intent; Zeng Ming considers it a third-stage development that cannot yet be realized.
- New Lab
- A new organizational form for people dissatisfied with the corporate system, centered on researchers; OpenAI is seen as its earliest example.
- High agency
- A Silicon Valley buzzword meaning taking charge of one's own fate and actively seeking tasks and growth, replacing hierarchy-and-command-driven work.
- Context not control
- A management philosophy: give ample shared context and information so members can make the right decisions on their own.
How to listen
CEOs working on AI strategy, executives at traditional giants hunting for a second curve, AI investors, and technical founders torn between building models and building applications.
The rapid-fire Q&A and book list at the end can be skipped (after 2:29).