Original video in Chinese.
Key Takeaway
- AI and the internet have completely different underlying logics: the internet has near-zero marginal cost and is consumption-driven, so it needs massive numbers of users; AI has real marginal costs (compute, electricity) and is production-driven, so it must first allocate its limited Tokens to heavy users who can maximize output value.
- Heavy users are the core of healthy AI growth: they are willing to pay for stronger capabilities, provide feedback from high-quality production scenarios, and expose real flaws, forming a positive loop of “high usage + high payment + high-quality feedback.” Anthropic is a typical example, focusing on high-intensity production scenarios rather than mass-market noise.
- Chasing massive numbers of users is dangerous: light users become “capacity black holes,” consuming large amounts of inference resources while contributing very little revenue and only shallow feedback, and they will also steer the product toward being “cheap and fun,” ultimately slowing down the model and the business model. I make it clear that I only serve heavy users; I don’t do education, I only do filtering.
AI does not need massive numbers of users. What AI needs most of all are heavy users. Because AI’s logic is completely different from the internet’s. A lot of domestic companies still haven’t realized this.
This is a viewpoint I posted inside Knowledge Planet. In this video, I’ll explain it in full, as a sort of Q&A.
Think about it: why does the internet need massive numbers of users?
First, the internet’s marginal cost declines, so it can support massive numbers of users.
Copying and distributing on the internet almost does not create additional costs. When users increase, server pressure can be handled through architectural optimization and cloud expansion.
The internet’s marginal cost approaches zero. So the logic of “one more user means one more source of revenue” is completely valid financially.
Second, the internet is consumption-driven, and it needs massive numbers of users.
The internet, especially consumer internet, corresponds to the attention economy. The core of its business is to get as many people as possible to see content, click ads, and complete purchases.
Consumption is broadly accessible. Everyone needs food, drink, entertainment, and social expression. The more users there are, the stronger the network effect and the higher the platform’s value.
Declining marginal costs and consumption-driven demand are the underlying logic of internet businesses. But if you use these two points to look at AI, you’ll find that this logic doesn’t work.
First, AI’s marginal costs are real.
Every conversation, every inference, consumes electricity, GPU compute, memory, and data center resources.
You need to understand that these resources are all hard constraints. Electricity is not something you can use as much as you want. GPUs and memory are not things you can buy in any quantity you want.
This means AI companies are no longer internet companies where “the more users, the better.” They have to think like factories:
Who should the limited Tokens be sold to first?
How can the output value of Tokens be increased?
How can low-value users be prevented from occupying all the capacity while contributing only extremely low revenue?
Second, AI is production-driven.
AI is a productivity tool, not a consumer product. This characteristic determines that it naturally does not need, and cannot be fully utilized by, everyone.
Ordinary users only need simple Q&A. The people who can truly turn AI into productivity are those with clear work goals, complex tasks, and professional scenarios. These people can push AI to its limits. Their output value per Token is extremely high.
So, under the premise of marginal costs, under real-world constraints, AI companies must prioritize allocating the best resources to the people who can maximize the use of this capacity.
That is why I say AI needs heavy users most.
Heavy users are willing to pay for faster responses, longer context, stronger tool calls, and more stable outputs. That directly covers the compute costs.
And only heavy use can expose AI’s real flaws in production scenarios and drive valuable iteration.
When an AI company has more heavy users, a business loop takes shape:
high-intensity use, plus high willingness to pay, plus high-quality feedback, equals higher output value per unit Token.
That is the kind of healthy growth an AI company truly needs.
Take Anthropic as a typical example. Unlike OpenAI, they are completely focused on high-intensity production scenarios. They know exactly what they are doing:
not building a flashy internet product, but serving the people who can squeeze the model’s capabilities the hardest and who are also most willing to pay for those capabilities.
By contrast, if an AI company still uses the internet playbook to do AI, it will definitely end up in a very uncomfortable place.
A large number of free users and light users may look lively, the data may look good, and DAU may look beautiful. But what they consume are inference resources, while what they contribute is extremely low revenue and very shallow feedback.
In the short term, that is growth.
In the long term, it is a capacity black hole.
What’s even more troublesome is that these users will also distort the product direction.
Because light users want something cheap, fun, barrier-free, and easy to use at will.
Heavy users want stability, power, control, and deliverability.
These two demands are completely different things.
If an AI company, in pursuit of massive numbers of users, throws all of its resources into light-use scenarios, its model, product, and business model will all be slowed down.
This is also why I don’t make content that is especially shallow.
It’s not because I can’t explain it; it’s because I don’t want to serve everyone.
In the AI era, what is truly valuable is not the bystanders, but the people who have already started using AI intensively to transform the way they work.
So I’ve always said: I don’t do education; I only do filtering.
I only face heavy users, people who can turn AI into their own production system.
OK, that’s all for this episode. If you want to understand AI, want to become a super individual, and want to find like-minded people, come join our newtype community. See you next time!