AI tokens: what is actually under the hood of the hypest sector

AI tokens are the leading narrative of recent cycles: any ticker with the words agent, GPU, or neural can climb dozens of percent in a day. Under the shared label live three very different businesses.

The sector's three floors

• Compute: decentralized GPU networks (Render, Akash, io.net) rent out graphics cards. The only segment with measurable revenue: rental hours are visible on-chain. The question is always the same - what share of the capacity is actually loaded with paying clients.
• Models and data: Bittensor and similar projects build markets where models compete for token emission. Beautiful cryptoeconomics, but the usefulness of subnets to the outside world is hard to measure - emission often rewards gaming the metrics.
• Agents: the most speculative floor - tokens of "autonomous AI agents". Inside there is usually a bot with an LLM subscription and a Twitter account. Market caps of such tokens regularly exceed the value of the entire technology stack inside them.

How to filter

The question for any AI token: who pays and for what? If the answer is "the token is needed so the token goes up" - it is a meme with an AI prefix, and it should be valued like a meme (see the degen chronicle). If there is revenue from external clients - it is a business, and then the usual analysis applies: market share, node economics, unlocks.