AI companies are not issuing their own tokens, but they are all aggressively selling tokens.

Today I attended a discussion event about Web3 and AI at Fudan University. While we were chatting on-site, an interesting question suddenly popped into my head.

Now, these companies that build and serve large-model APIs—like Kimi and MiniMax—what exactly are they selling?

On the surface, they’re selling model capabilities: question answering, generation, reasoning, search, and tool calling. But if you set aside these packaging explanations and look only at the most straightforward business actions, you’ll find something that looks a lot like Web3:

Every time you call the model, the system deducts some portion of your tokens.

Then when you look back at public blockchains, you’ll notice that these two things are actually quite similar.

In the Web3 world, people say: I sent a transaction and consumed 0.01 ETH. In the AI world, people say: I called a model once and consumed 100,000 tokens.

The former sounds like blockchain-world language, while the latter sounds like cloud computing or SaaS billing. But if you look one layer deeper, what they do is actually very similar:

They both take a kind of underlying resource, divide it into the smallest units that can be computed, consumed, and settled, and then sell it to developers and users.

From this perspective, I think many things that today’s AI companies are doing—and many things that past public-chain projects did—are, in terms of business structure, fundamentally aligned.

They’re all selling tokens.

Of course, by “token” here, I don’t mean that AI companies really issue a coin that can be freely traded, listed on exchanges, and speculated on like public-chain projects. That’s not what I mean. I mean that they’re all selling a standardized unit of resource usage.

At its core, they’re selling the right to call

When you use Kimi, you’re not directly buying “an article” or “an answer.” You’re buying the model’s ability to process text—how much it consumes of the context window, the cost of the reasoning process, and the frequency and quota of API calls. The platform only takes these originally very abstract things, breaks them into tokens, and then charges you according to the amount consumed.

Doing things on a public chain is the same. You’re not directly buying the four words “transfer successful.” You’re paying for the resources consumed to get the network to complete one accounting, verification, ordering, and state update. It’s just that in the on-chain world, the unit used to measure this resource consumption is called gas—and then you pay for it using native tokens like ETH and SOL.

So, in a very deep sense, AI companies and public-chain projects really are alike: they’re not directly selling results; they’re selling “the right to access underlying compute resources.”

Looks similar, different in nature

But if an article only writes up to here, it’s not enough. Because while AI companies’ tokens look like public-chain tokens, they’re actually not the same thing.

The most core difference is that their underlying rights structures are different.

The tokens AI companies sell to you are, in essence, an internal billing unit within the platform. You top up, open an account, get an API key, and then consume quota according to the platform rules. What you have isn’t an asset that can circulate freely, be transferred freely, and exist independently of the platform; instead, it’s a usage right recognized by the platform.

You can think of it as what? More like game points, or a cloud provider’s backend call quota, or a balance in a membership system. This is valuable because it can be exchanged for services; but the boundaries of its value, the usage rules, and how pricing is adjusted are basically all controlled by the platform.

But public-chain tokens are different. Things like ETH and SOL are not only measurement units inside a system. They are native assets in the network itself. They can be held by you, transferred, traded, staked, and used as collateral—and they can also exist independently of any specific call action.

One is platform pricing, the other is network pricing

Today’s AI companies’ token prices are basically set by the company itself. What model you call, what you pay for input, what you pay for output, how much you pay for a long context, how much you pay for tool calls—these are all written in the platform’s backend. Whether users accept it or not, in essence, it’s still platform pricing.

But blockchain gas isn’t as simple as just the platform slapping on a price tag. On-chain fees are affected by network congestion, and also by the combined influence of market supply and demand, users’ bid amounts, and the protocol’s mechanisms.

The AI token billing logic is essentially corporate pricing; the public-chain token billing logic is more closely aligned with protocol-and-market joint pricing.

AI teaches Web3 a lesson in return

The most valuable part of this observation isn’t that “AI is similar to Web3,” but that it can help us re-understand an old question from the other direction: why do many Web3 projects ultimately fail, while token billing from AI companies feels naturally reasonable to people?

The reason is simple.

Because the tokens from AI companies have a very clear resource object behind them, and a very clear reason for paying.

When you use a model, you’re truly consuming compute. When you run long context, you’re truly occupying window resources. When you do search or tool calling, you’re truly increasing platform costs. Every additional call means the platform has to bear more marginal cost.

So the platform splits this kind of consumption into tokens and then charges by tokens—that logic fits very well. Users can understand it right away: the money I spend corresponds to the resources I actually consumed.

But the issues with many Web3 projects in the past weren’t because they were called tokens; it was because, behind them, there wasn’t a strong real consumption scenario. When many projects talked about their business model, they didn’t first think about why users would need to keep using and paying; they thought first about “how to build up tokens.”

First make the use case work, then talk about tokens

I think the biggest inspiration AI companies give to Web3 entrepreneurs might be right here.

First, think clearly about this: who is the user actually paying for?

Then, make it solid: why does the user keep paying repeatedly?

Then look further: in this business, is there truly some kind of resource consumption that can be split, measured, and settled?

Instead of asking right away, how to issue tokens, how to get listed, how to manage market cap.

Not everything is worth putting on-chain

Not everything is worth putting on-chain. Because many businesses are already more efficient with centralized platforms, with clearer contractual relationships—there’s simply no need to force them into an on-chain architecture.

And not every unit of measurement is worth trading. Because many measurement units only work well as internal settlement tools. Once you put them into the market for free trading, it can actually distort the original logic of use.

In many cases, the best token isn’t the one that spikes the hardest in price, but the one that you don’t need to constantly watch—yet is continuously consumed, continuously settled, and continuously repurchased in real business.

Quick summary

So coming back to today’s topic: “AI companies don’t issue coins, but they’re selling tokens like crazy.”

What AI companies sell isn’t a coin in the crypto-currency sense. They’re selling a token that’s been corporateized, productized, and contract-formalized.

In the Web3 world, what shouldn’t be just a coin that can go up and down. The truly valuable part should originally be the ability to natively price some resource, some network capability, or some right to change state.

So, let’s not keep thinking about issuing coins. First, figure out what you’re actually selling.

Are you selling a story, or are you selling a resource?

Are you selling imagination, or are you selling the right to call?

Are you selling a financial illusion, or are you selling an underlying capability that truly exists—something that can be consumed repeatedly, priced repeatedly, and repurchased repeatedly?

That might be the most valuable reminder AI brings to Web3.