We're Not at Yahoo Yet. We're Waiting for the Browser.

Notes from Zeng Ming on the Zhang Xiaojun Podcast, Sep 2026

Zeng Ming (曾鸣) was Alibaba’s chief strategy officer and co-founded CKGSB and Hupan. He spent 2.5 hours laying out a stage model for AI industrialization that is worth stealing. Here are the four parts that matter.

My view: AI could take society in countless directions. Zeng Ming’s framework rests on a crucial assumption: that AI will follow the adoption pattern of earlier general-purpose technologies rather than trigger a technological singularity. That assumption may prove wrong. If AI does produce a singularity, historical analogies will offer little guidance, and much of what follows will be impossible to predict.

1. The framework: every general-purpose technology runs the same three stages

Stage 1 — the technology becomes infrastructure. Raw capability gets standardized until anyone can call it without understanding it.

Stage 2 — application explosion. A hundred flowers bloom. Most of them die. This stage is not a waiting room; it is where the direction of the next technology round gets discovered.

Stage 3 — native applications. The era-defining products, built by people who never knew the constraints of the previous era.

The analogy he leans on is mobile: BlackBerry defined “smartphone” and still held 40%+ of the US market in 2010 (stage 1) → iPhone 4 and the App Store (stage 2) → Douyin and Pinduoduo (stage 3). ByteDance didn’t start with Douyin. It started with Toutiao, because in stage 1 only text was mature. Recommendation algorithms plus short video — two things that had to arrive separately — combined into a national application.

His second analogy is electricity, and it’s the more useful one. Edison lit Manhattan in 1882. The grid covered US cities around 1910. The washing machine and refrigerator also arrive around 1910; the air conditioner in 1925; television in 1930. Roughly thirty years from “the technology works” to “the applications people can’t live without.” And the company that got rich wasn’t the one selling generation equipment — GE’s fortune came from the appliances.

You cannot skip stage 2. Stage 3 is built on the accumulation of stage 2, which means the current “applications are dead, models eat everything” mood is a category error about timing.

2. Where we are now: post-token, pre-browser

Zeng dates stage 1 as essentially complete in 2026, and his marker is the token.

His reasoning is nice: a technology becomes infrastructure the moment it has a standard unit of measurement. A kilowatt-hour, a ton of water, a gigabyte. Once Jensen calls NVIDIA a token factory and Sam says OpenAI sells tokens, everyone knows how many tokens a job takes and what it costs. The frontier has become a utility.

Stage 2 opened at roughly the same moment, with OpenClaw over Spring Festival — the first time a general audience saw an agent complete a task on its own. Agents are the applications of the AI era.

But here’s the uncomfortable placement. His first instinct two years ago was “Yahoo is coming.” He has since moved it earlier:

we’re still waiting for a browser to come out

Why the browser? Because the browser did two things at once: HTML gave website-building a standard so anyone could publish, and clicking made consumption free. Supply exploded, then demand exploded, then Yahoo appeared to reduce the entropy — a directory, then categories, then structure.

We have none of that for agents. There is no standard for what an agent is, no zero-effort way for a normal person to invoke one, and — critically — no trust layer. Would you hand an agent your wallet? Your private data? How do you know its fees are real? A website you can evaluate in three seconds and abandon for free; a bad agent has consequences. That gap is a two-sided market waiting for someone to build it.

If he’s right, the largest near-term opportunity isn’t a better model or a vertical agent. It’s the agent browser: standards on the supply side, one-click invocation on the demand side. He guesses ~3 years for an entry point to emerge, ~5 before it’s obvious, and 8–10 years for the era’s Douyin.

3. Where OpenAI and Anthropic end up: excellent infrastructure

This is the “brutal truth” the episode title advertises.

Model companies, in his read, are AI cloud companies. Infrastructure. And infrastructure has exactly one endgame: oligopoly plus heavy regulation, because society depends on it, the capex is enormous, and no one is allowed to be the single supplier. Differentiated enough that everyone earns; homogeneous enough that everyone is substitutable.

That is a very good mature business. It is not the era-defining one.

The historical rhyme he uses to defuse “this time is different”: Yahoo went from founding in 1995 to a $120B market cap in 2000 — inflation- and index-adjusted, roughly a trillion-dollar company today. AOL peaked above $220B. Both were as shocking then as today’s valuations are now. Both are gone. Meanwhile Google was founded in 1998 and was, for a while, a B2B algorithm supplier to Yahoo.

Then the analogy that does the real work: refineries versus chemical companies. Refining is the base layer; chemicals became far more valuable. Or: a car cannot run without oil, and no oil company has ever built a car.

So his call is that OpenAI and Anthropic will keep doing extremely well for years and most probably will not be the big winners of the native era. Not a strategy error — a structural one. He notes the same pattern one level up: companies of stage 1 rarely survive into stage 2, and stage 2 into stage 3, because stage 1 rewards original technical breakthrough while stage 2 rewards product insight, user empathy, and a completely different stack (context management, memory, new interaction models).

And era-to-era? Asked whether any company has ever crossed and stayed great, his answer is that it isn’t a question of how many — it’s whether there are any at all. IBM once. Microsoft arguably. That’s the list.

4. Who wins the native era

Not, in his view, anyone currently obvious.

Google probably survives as a great AI cloud — TPUs, models, distribution. Whether it holds the consumer entry point is a different question.
ByteDance likely becomes an AI cloud company too. On Doubao he is blunt: it’s a first-generation chatbot with no ability to act, and the fact that ByteDance stopped spending unlimited money on it is the tell.
Tencent and Alibaba face restructuring of the two relationships they own — people-to-people and people-to-goods. Being visibly nervous is the correct response.
xAI and Meta he treats as already out of the frontier race, and interestingly he blames organization over compute. His anecdote about xAI engineers being the most visibly burned-out people in the Valley leads to a real claim: a wound-tight, top-down org cannot produce a breakthrough when the leader can’t supply the answer himself and won’t leave room for anyone else to.
Anthropic and OpenAI he rates as genuinely strong organizations — Anthropic for focus and for using models to train models (“intelligence compounding”), OpenAI for depth that lets it chase fast. Which is precisely why he thinks their ceiling is structural, not managerial.

The winner is a company nobody can currently describe. By his own framework that’s required: “it must be a company that no one knows what it looks like,” or it isn’t a big enough innovation to deserve the value.

Three concrete things he’d look for:

A 10x entry point. Whoever wins is not making slide decks or reports — too small to grow into. It has to be a complex job, because complexity is what lets you build a real intelligence loop instead of scaffolding around a model’s current weakness.
The black hole effect, his upgrade to network effects: the more scenarios you absorb, the more your intelligence compounds, the faster you swallow adjacent scenarios.
An AI-native organization. His most radical claim is that the company itself — hierarchical, position-based, industrial-era — is on the way out, replaced by task-based, partnership-shaped organizations. The earliest native CEOs, he notes, have already personally rewritten how their company runs. You don’t get a native strategy out of a non-native org.

The optimistic reading: if the agent-browser entry point is about three years away and truly AI-native applications are eight to ten years away, then the “AI application winter” of 2026 is a stage-2 trough, not a final verdict. The people building agents today are early—not wrong.

The pessimistic reading: almost everyone reading this works at a company that his framework says cannot make the jump.

His own resolution is worth keeping: companies are born, age, and die, and nothing is more natural. That’s exactly why new companies get a chance.

Source: Zeng Ming interviewed by Zhang Xiaojun, Zhang Xiaojun Podcast, September 2, 2026 (2:34:18).