Every AI token I get asked about arrives with a story attached, and the story is almost always better than the token. Decentralized compute, an open marketplace for inference, permissionless data for training, agents paying each other in the background. All of it sounds inevitable when you read the deck. The problem is that a narrative that good will pull twenty tickers up together, and only three or four of those tickers have anything underneath them. Figuring out which ones is not glamorous work, but it is the only work that separates a name you can hold through a drawdown from a name you are just renting until the sector rolls over.
So the question I keep coming back to is boring and specific. When someone actually uses this network, does money move, and does the token have to be involved for it to move. If the answer to either half is no, I am looking at a beta play on the theme. That is fine as long as I know that is what I bought. What gets people hurt is holding a pure beta play and telling themselves it is a fundamental one.
Start with utilization, not price
The first thing I want is evidence that real jobs are getting paid for, not promised. A lot of AI infrastructure projects publish a big number for total capacity, GPUs registered, nodes online, models listed. Capacity is supply. It tells you nothing about whether anyone showed up to buy. What I want is the demand side, and it is usually harder to find because it is usually smaller.
Concretely, I go looking for a few things. Paid jobs settled on-chain or reported through a verifiable feed, ideally with a count and a value, not just a vanity dashboard. Revenue actually flowing to the people running nodes, because if operators are only earning inflationary token emissions and no customer payments, the network is subsidizing its own activity and calling it usage. And burn or fees that scale with real work, so that a busy month looks different on-chain from a quiet one. If usage doubles and nothing measurable changes in the token's fee flow or burn, the token is decorative.
The failure mode here is treating emissions as adoption. A protocol can hand out tokens to node operators for staying online, those operators sell into the market, and the whole thing produces a lot of on-chain motion that looks like a thriving economy. None of it requires a single paying customer. When I see high on-chain activity and near-zero external revenue, I assume I am looking at a closed loop until proven otherwise.
Ask whether the token is required or ornamental
This is the question that quietly decides most of it. Take the network's core action, renting a GPU, running an inference call, buying a dataset, and ask what happens if you try to pay for it in a stablecoin instead of the native token. If the answer is that you comfortably can, and plenty of well-designed systems let you, then the token is not in the payment path. It might still capture value through fees or staking, but you have to trace exactly how, because the default assumption should be that it does not.
A few patterns tell me the token is genuinely load-bearing rather than bolted on:
- Every unit of work must be paid in the native token, or the native token is burned as a mandatory step even when the user pays in something else.
- Node operators or providers must stake the token to participate, and that stake is slashed for bad behavior, so the token is doing real security work tied to usage volume.
- Fees are collected in the token and either burned or routed to holders through a mechanism you can actually point to in the contracts, not just in the marketing.
And a few that tell me it is ornamental. Governance as the only stated use, which usually means the team could not find a real one. A token you stake to earn more of the same token, with no external cash coming in. Or a payment flow where the native token is one accepted option among several and clearly the least convenient. If I have to squint to explain why the token has to exist, the market will eventually squint too.
Compare valuation per unit of actual usage
Once I have a name or two that pass the first two tests, I try to price them against their own usage rather than against the story. The cleanest way I have found is to build a rough ratio of fully diluted valuation to something that measures real work: annualized fees paid by customers, annualized revenue to node operators from paying users, or total value of settled jobs. It is crude, and every project defines its numbers slightly differently, so I never trust a single figure. But when I line up three or four networks in the same sub-sector this way, the spread is usually enormous, and the enormous spread is the useful part.
Two things matter as much as the ratio itself. First, use fully diluted valuation, not just circulating market cap, because AI tokens tend to have long emission schedules and the unlocks are real supply that will hit the market. A network that looks cheap on circulating supply can look absurd once you account for tokens that have not vested. Second, look at the trend, not the snapshot. A high multiple that is falling because usage is compounding is a very different animal from a high multiple that is flat because the number never moved. I would rather own an expensive-looking network with real quarter-over-quarter usage growth than a cheap-looking one that has been sideways since launch.
A short checklist before you size anything
When someone sends me an AI token, I run the same quick pass before I have any opinion on price. Is there evidence of paid jobs, or only capacity and promises. Do node operators earn customer money, or only emissions. Could a user pay in a stablecoin and skip the token entirely. What is fully diluted valuation against real usage, and is that number trending the right way. What does the unlock schedule dump onto the market over the next year or two.
Most tickers fail on the first or third question, and that is the point of asking them early. The handful that survive tend to be the same handful the sector quietly keeps coming back to, because they are the ones where usage would keep paying the network even if the narrative went cold for a while. Everything else is a way to express a view on the theme, which is a legitimate trade as long as you hold it like one and let go before the story does.