Why ICP’s AI narrative is the most underestimated and may bring the most market-differentiating value applications in 2025:

TL:DR
> Undervalued/value depression
> AI agent sovereignty (no TEE required)
> AI agent basic business logic is implemented on-chain
> AI agent sovereign economic ecology
> On-chain large model and GPU subnet
> The above is difficult to achieve in other public chains
> ai16z / ElizaOS on-chain deployment
> AI builds with humans (agent developers)

1/ Underestimation:#ICPhas not appeared in many crypto AI research reports, so more than 95% of people don’t know about it. But this problem is not difficult to solve. On the contrary, it is more important to appear at the right time.

2/ The current state of the crypto AI market is centralized AI agents issuing tokens on-chain. The agent itself has not had much more application aside from tweeting; the tokens on-chain are the only part related to web3. On the other hand, there is GPU computing power aggregation, like IO[dot]net, but currently the crypto AI does not have a huge demand for computing power; rather, the demand mainly comes from the tokens issued by the on-chain transaction agents, making computing power aggregation a bit awkward.

3/ What #ICP is doing, and what it can do in the future:

3.1/ On-chain AI. Currently, there is no on-chain AI on other public chains, which is why there has been recent discussion about TEE (How do you prove that the AI agent's tweet is generated by itself? How can you ensure it has its own tweet account password and that humans cannot interfere?). Here, the on-chain AI that #ICP can achieve is divided into several levels:

AI agents have their own sovereignty: meaning they can operate independently (humans cannot interfere or reset), and can prove this. #ICP can deploy agents directly within its canister smart contracts, and the agents can generate their own private keys, signing each tweet with this key. This way, agents can achieve and prove their sovereignty without the need for TEE. Of course, #ICP network node operators theoretically could (with significant effort) collude and obtain the private keys inside the canister; in this case, the threat to the agent's sovereignty could be recognized, but TEE can add a layer of protection.

The basic commercial logic of AI agents operates on-chain: this is a more programmatic part of the agent logic, requiring lower demand for large models. The birth of #ICP serves as a decentralized network for hosting software services, its canister smart contracts can store 400GB of data; therefore, AI agents and their basic commercial logic can fully run on-chain, which is true DeAI, and it will not be affected by cloud server downtime or attacks.

AI agent sovereignty economic ecology: Most of the agent's uses are to provide software services, thus it needs to be able to host services, pay hosting fees to data centers like AWS, and charge users fees, ultimately achieving a self-sustaining economic cycle. However, the KYC for data centers and banks currently does not support non-human users. But if an agent structures its software service on #ICP, it only needs to burn $ICP and $CYCLE as hosting fee payments and charge users in crypto. Each agent can have its own #ICP wallet (by the way, the #ICP chain integration technology supports multi-chain interoperability and payments).

Uploading large models to #ICP to achieve 100% on-chain AI: The difficulty of this implementation is relatively high, as large centralized data centers have the most powerful computing capabilities. However, within the current #ICP subnet environment (all CPU machines), it is already possible to upload certain large models like GPT2. Additionally, according to tests by DFINITY Foundation engineers, billion-parameter level large models are also possible. Theoretically, carrying super large models or a large number of different models can also be achieved by deploying GPU subnets. The value of on-chain computing lies in the massive burning of $ICP. This is why #ICP's network protocol revenue has consistently ranked in the top 5-10 in recent months, as its computing potential is at least a hundred times that of other public chains. Although the current speed is not perfect, there are already projects like @DecideAI_, @onicaiHQ, and @ICPandaDAO launching purely on-chain agents. Of course, to achieve a more scalable computing power, GPU sharing and aggregation are also meaningful, with related projects like @ngpu_ai.

An additional noteworthy case: The popular AI agent open-source framework ElizaOS initiated by the ai16z team has been deployed on #ICP. Shaw, Lucid, and Daniel have all been OG developers in the #ICP ecosystem and understand its unique value for on-chain AI. We will soon see Eliza living on #ICP.

3.2/ AI builds with human: Imagine an AI agent that can not only issue tokens but also build projects and applications on-chain. According to a recent survey by Menlo Ventures: in the Silicon Valley AI circle, the primary use case for agents is coding co-pilot. This is not achievable on other public chains because they can only issue tokens, not host software services; but it is possible on #ICP. Therefore, those who cannot write code can develop software on-chain. A specific case will be left as a surprise, and everyone can follow @caffeineai.