Why the AI-token label has stopped meaning anything
Search any crypto launch feed in 2025 and the words AI, agent, and intelligence appear in roughly one in three new tickers. Some of those tokens route real compute. Most of them do not. The category has become a branding shortcut the same way DeFi did in 2020 and metaverse did in 2021, and the result is that a buyer who types AI token into a search bar gets a wildly mixed basket of risks dressed in similar language.
The honest move is to stop asking whether a token is an AI token and start asking what kind of token it is underneath the label. Four buckets cover almost everything that launches today, and once you can place a project in one of them, the risk profile, the expected returns, and the questions you should be asking all change dramatically.
The four honest categories of AI tokens
Think of these as four distinct business models that happen to share the AI label. They have different sources of demand, different token mechanics, and different failure modes.
1. Infrastructure and utility tokens
These tokens pay for something the network actually provides. A user sends the token, the network performs a service, and the fee either burns the token or routes to validators or stakers. The classic examples in the seeded list are FET from Fetch.ai, which pays for agent-to-agent messaging and task coordination, and RENDER, which pays for GPU rendering work on the Render Network.
The test is straightforward. Can you point to a transaction where the token was the thing being paid, and can you show where that fee went? If yes, you are looking at infrastructure. If the whitepaper says the token will later be used for compute but no one is paying in it today, the token is closer to a claim on future utility, which is a different and riskier thing.
2. Governance tokens
Governance tokens give holders voting power over a protocol. They may also accrue fees through staking or buybacks, but the defining feature is that they steer the rules of the system. TAO from Bittensor is the clearest seeded example. TAO holders vote on which subnet emissions get allocated, which directly determines how new TAO is distributed across the network.
The honest read on governance tokens is that voting power is only valuable when the protocol controls something important and contested. A governance token for a system nobody uses is mostly an artifact, and the price reflects anticipated future usage rather than current cashflow.
3. Launchpad and meta tokens
Launchpad tokens sit one layer above individual projects. They entitle holders to allocations in new token sales, sometimes at a discount, and they often govern which launches the platform supports. VIRTUAL from Virtuals Protocol and its sister token VVV from the Venice ecosystem are seeded examples that sit in or near this category, since they coordinate the launch of AI-agent projects and gate access to those launches.
The economic logic of a launchpad token is the logic of a venture fund. It makes money when the portfolio launches appreciate, and it loses money when launches flop or fail to find buyers. Holders are effectively running a tiny index of speculative AI bets, with the extra twist that the token itself is part of that index.
4. Memecoins wearing AI branding
This is the largest category by token count and the riskiest by a wide margin. A memecoin borrows an AI meme, often an agent character or a chatbot persona, and rides attention cycles without a functioning product. The category is so large that several real AI-agent projects began as memecoins before shipping anything, which makes the boundary genuinely fuzzy.
The honest framing is that a memecoin is not inherently a scam. It is a token whose price is driven by attention, not by fees or governance rights. Calling it out is not a moral judgment. It is a clarification of what kind of bet you are making.
Risks that scale with each category
Every category has its own failure modes, and the failure modes matter more than the upside because they decide whether you can lose everything or just some.
Infrastructure tokens: the slow-bleed risk
Infrastructure tokens can fail quietly. If a competitor ships a better model or a cheaper API, network usage migrates and the token's fee flow dries up while the price drifts for months. There is no dramatic rug because the project is still running. The risk is that you hold a working product with a token that no one needs to use.
The historical pattern is brutal. Many 2021 DeFi infrastructure tokens continued to operate after their tokens lost 90% or more of their value, because the protocol kept working even as the speculative premium evaporated.
Governance tokens: the empty-vote risk
A governance token whose only voters are the team and a few large holders functions more like a corporate board than a community. Retail holders vote on proposals that have already been decided. Worse, the proposals themselves may not be meaningful if the protocol's treasury is small or its parameter space is narrow.
The classic tell is that governance volume is low even when the token price is high. If nobody is voting and the token is still trading, governance is a label, not a feature.
Launchpad tokens: the bag-of-launches risk
Launchpad tokens concentrate risk in exactly the projects you would not have bought yourself. The platform needs to launch tokens to keep users engaged, which creates pressure to list weaker projects, which dilutes the platform's reputation, which eventually bleeds into the parent token's price.
The 2024 to 2025 launchpad cycle on several chains showed this dynamic clearly. Platforms that listed every hyped memecoin saw their native tokens compress even as launch revenue stayed high, because the market read the listings as desperation.
Memecoins: the standard wipeout risk
Memecoins can go to zero in days. Liquidity is shallow, unlock schedules are often hidden, and the social signal that drove the rally can reverse overnight. The honest statistic is that the majority of memecoins launched in any given cycle lose most of their value within weeks, and a meaningful share go to effectively zero.
For any memecoin, the question is not whether it will moon. It is whether you can tell when the insiders have stopped buying.
Tests you can apply to any AI token in five minutes
The category matters less than the questions you bring to the asset. Here are four tests that work across the whole taxonomy.
Does the token capture fees?
Pull the protocol's documentation and look for a fee flow diagram. Where do the dollars (or stablecoins) go when someone uses the product? If the path goes through the token, you have a fee-capturing asset. If the path bypasses the token entirely, you have a governance or coordination token whose value is indirect.
Does issuance slow over time?
Look at the supply schedule. A token whose emissions fall each cycle is more likely to accrue value than one with a flat or rising inflation rate, all else equal. Many AI tokens in the seeded list have emissions that taper, but several newer launches have uncapped or even inflationary designs that dilute holders indefinitely.
Who is selling the token to you?
If the address selling into the rally is a multisig labeled team, treasury, or early investor, that is information. If it is a series of fresh wallets funded by the same source, that is also information. On-chain transparency makes this question answerable in ways that traditional finance still struggles with.
Is the AI part real or a wrapper?
Read the technical docs. Does the project train models, run inference on a distributed GPU market, coordinate agents, or do something else that genuinely uses AI compute? Or does the AI element consist of a chatbot interface, a logo, and a marketing story? The distinction is the difference between a token that may have product-market fit and one that has narrative-market fit.
How revenue-share and buyback claims actually work
Many AI tokens now advertise that they share revenue with holders or that they buy back the token from open market. Both mechanisms can be legitimate, but both are also easy to fake in marketing copy.
Revenue share, verified and unverified
A real revenue share pays holders in stablecoins or in the token itself from a verifiable income source. You can check the protocol's fee revenue on a dashboard, you can check the distribution contract, and you can confirm the dollars actually arrived. Unverified revenue share is a claim in a Medium post with no on-chain evidence.
The seeded tokens provide a useful spread here. RENDER has historically captured fees from GPU work and routed portions to stakers. TAO captures value through subnet emissions rather than direct fee distribution. VIRTUAL and VVV coordinate launches, with revenue mechanisms that have varied by quarter. Always check the most recent documentation because these designs evolve quickly.
Buyback mechanics as a proxy for value capture
Buybacks reduce circulating supply, which is bullish for price all else equal. The catch is that buybacks funded by inflationary emissions are not value capture. They are circular. The token is buying itself with newly minted tokens, which inflates the buyback number without changing the holder's real claim.
The honest test is to compare treasury inflows (real revenue) with treasury outflows (buybacks and operations). If buybacks are roughly covered by genuine revenue, you have a working mechanism. If buybacks are funded by emissions, you have marketing.
How to tell an AI-agent launch from a memecoin
This is the question most readers actually want answered, because the AI-agent narrative drove a large share of 2024 and 2025 token launches. The honest answer is that the boundary is blurry by design, and you have to look at specific signals.
Signal 1: does the agent do work?
An AI agent that actually performs a task, like executing trades, managing a portfolio, responding to on-chain events, or coordinating other agents, generates demand for the token that pays for its compute. An AI agent that is mostly a Twitter persona or a chatbot wrapper is closer to a memecoin.
Signal 2: is there a usage chart?
Look for active addresses, API calls, task completions, or any metric that shows the system is being used beyond token transfers. If the only chart going up is price, you are looking at attention, not utility.
Signal 3: who launched it and how is the supply distributed?
Agent launches with locked team allocations, clear vesting, and modest insider unlocks look more like real projects. Agent launches where a single deployer wallet holds a large share of supply, or where the team can mint more tokens, look more like memecoins regardless of the pitch deck.
The seeded examples illustrate the spread. FET supports a functioning agent economy with measurable activity. RENDER supports a real GPU market. TAO coordinates a working subnet ecosystem. VIRTUAL and VVV operate launch infrastructure with on-chain launches. Many other tokens in the same news cycle have far thinner evidence behind their agent branding.
Putting the taxonomy to work in your own research
The point of the taxonomy is not to label tokens from the outside. It is to give you a checklist you can run on any project in a few minutes. Start by placing the token in one of the four categories. Then run the four tests: fee capture, supply schedule, who is selling, and whether the AI is real. If you cannot answer all four, treat the position as speculative and size it accordingly.
A practical workflow looks like this. Read the project's main page and place it in a category. Pull the token contract and check supply, holders, and recent transfers. Find the protocol's documentation and look for a fee flow. Search the team's wallet history for unlocks and sales. Cross-check any revenue claim against on-chain data.
This is not glamorous work. It is what separates buyers who occasionally catch a 10x from buyers who consistently lose money to the same story told with different mascots.
Stay ahead of AI-token launches with better signal
AI tokens move fast and so does the news around them. Tracking which launches have real fee capture, which are governance plays, and which are memecoins wearing agent branding manually is a losing game. Zippfeed surfaces AI-token headlines with sentiment scoring (bullish, neutral, or bearish) and an importance rating, so you can apply the taxonomy above to the projects that actually matter instead of chasing every ticker that trends.
CoinDesk
