This question came up in Chinese AI developer circles recently, sparked by a talk Wang Jian gave at the Wuzhen Summit. Wang is an academician of the Chinese Academy of Engineering and the architect behind Alibaba Cloud. His argument was blunt: large enterprises are structurally incapable of treating AI as anything other than a maintenance tool. Individuals and tiny teams, because they own nothing worth protecting, are the ones most likely to actually build something new.
I want to take that claim seriously rather than either cheering it or dismissing it, because I have watched both failure modes play out, and the details matter.
The failure mode nobody talks about loudly enough
Here is the concrete thing that happens. A well-resourced team at a mid-to-large company gets access to a capable model. They have three months of runway on the project, two senior engineers, a product manager, and a data pipeline that took fourteen months to build. They ship an AI feature. It improves a metric by 6%. Leadership calls it a win. The project closes.
What actually happened? They attached intelligence to an existing workflow. The workflow did not change. The mental model behind the workflow did not change. The org chart did not change. The 6% improvement is real, but the ceiling of what was possible was set on day one, when the team implicitly agreed that the existing product was the thing being improved.
That 6% cost something less visible: the opportunity to ask what the product would look like if it were built today, from scratch, with current tooling. Nobody asked that question because the answer would have been inconvenient. Sunk costs in infrastructure, in headcount, in customer commitments all made the question feel irresponsible.
Wang's framing names this precisely. He calls it the difference between a revolution in tools and a revolution. Large organizations get the former by default, not because their engineers are uncreative, but because the incentive structure punishes anyone who seriously proposes the latter.
What zero legacy actually buys you
The counterargument from the skeptics in that same discussion thread was sharp: sure, you have no legacy, but you also have no data, no distribution, no trust, and no infrastructure. A small team trying to build an AI-native product from scratch hits the data wall fast. Enterprise systems accumulated years of structured interaction logs. A solo builder has a laptop and a credit card charged to an API.
This is true and it is also incomplete.
The data wall is real for applications where proprietary historical data is the actual moat. Think fraud detection, medical imaging, industrial predictive maintenance. In those domains, the incumbent's data advantage is not a legacy burden, it is the product. No amount of zero-overhead agility closes that gap.
But a surprisingly large share of what people are actually building with LLMs right now does not live in that category. Workflow tools, domain-specific interfaces, niche content pipelines, micro-SaaS that automates a specific professional tedium: these do not require proprietary training data. They require understanding a user's specific problem well enough to wire the right model behavior to it. That understanding is a research problem, not a data problem. And research requires the ability to throw away your last three weeks of work without a meeting.
The cost of throwing away three weeks of work at a large company is not three weeks. It is three weeks plus the political capital spent getting the project approved, plus the morale cost of telling the people who built the discarded version, plus the delay on the roadmap item that was deprioritized to make room. The real cost is closer to three months of organizational friction. A two-person team throws away three weeks and starts over on a Tuesday.
Where the incumbent advantage reasserts itself
I want to be honest about where this argument breaks down, because it breaks down in ways that are predictable and that have already burned people.
Distribution is not legacy. A company with 40 million active users can ship a mediocre AI feature and get 400,000 people using it within a week. Those 400,000 users generate feedback, edge cases, and usage patterns that no amount of clever prompting by a solo builder can replicate. The builder learns in theory; the incumbent learns in production at scale. After six months, the incumbent's mediocre feature has been iterated against real usage data forty times. The builder's clever feature has been used by 800 people.
This is the dynamic that the skeptical commenter in the original thread pointed at when they wrote that small businesses will end up being the large platforms' cattle. It is not wrong. The platform captures the distribution, the data flywheel, and ultimately the margin. The small builder captures the first-mover learning and, if they are lucky, gets acquired or gets copied.
The question is whether that is a problem or just a description of how markets work. Builders who went into AI short-form video production in China's market last year, with no prior video production experience, made real money before the platforms commoditized the tooling. The window was maybe eighteen months. That window is a real opportunity even if it closes.
The structural bet Wang is actually making
What Wang is describing is not a permanent state of affairs. It is a window condition. In the early phase of a major platform shift, the cost of experimentation drops dramatically while the value of institutional knowledge temporarily declines. Knowing how to run a profitable video production house in 2021 was not an asset in 2024 if you were trying to build AI-native video tools. It was a prior commitment.
This window does not stay open. It closes when the new paradigm stabilizes, when the infrastructure matures, when the incumbents finish their defensive integrations and start using their distribution advantages again. At that point, having nothing becomes a disadvantage again.
So the practical read on Wang's argument is not "individuals will beat corporations at AI forever." It is closer to: right now, in this specific window, the cost of your institutional history is higher than it has been in a long time, and the cost of starting from zero is lower than it has ever been. Both of those conditions are temporary.
Practitioners who have shipped products during this window and hit the ceiling of what their team could sustain know this intuitively. The freedom to experiment is real. The distribution problem reasserts itself the moment you need to grow past the first cohort of early adopters.
What this means for how you structure the work
The teams that have navigated this most effectively, from what I have observed in both Western and Chinese developer communities, tend to do one specific thing: they treat the zero-legacy advantage as a finite resource to be spent rather than a permanent condition to be celebrated.
They move fast on product definition precisely because they know the window closes. They do not spend the freedom on process. They do not build infrastructure they do not need yet. And critically, they pick target users whose problem is specific enough that distribution through a large platform would actually hurt rather than help, at least initially. A tool that is exactly right for neurologists is not a tool that benefits from being surfaced to 40 million general users.
The failure mode is the inverse: treating the agility as evidence that the data and distribution problems do not apply to you, building for a generic user, and discovering eighteen months later that the incumbent shipped a version of your product as a free feature.
Wang's observation is correct as a structural analysis. The question it leaves open is the one I actually want answered: for the people who have shipped AI products in the last two years with small teams, what was the specific moment when you realized the zero-legacy window was closing for your particular product, and what did you do differently because of it?
