ABCDE: Summarizes the AI+Crypto Track over the past year

Author: ABCDE Investment Partner Lao Bai

More than a year after the release of ChatGPT, the discussion about AI+Crypto has heated up again in the market recently. AI is considered one of the most important tracks for the bull market of 2024–2025, even Vitalik Buterin himself published an article “The promise and challenges of crypto + AI applications” to explore the potential directions for AI+Cryto in the future.

This article will not make too many subjective predictions but simply review the AI and Crypto startups observed over the past year from a primary market perspective, examining the angles from which entrepreneurs have entered the market, the achievements made so far, and areas still being explored.

1. The Cycle of AI+Crypto

Throughout 2023, we discussed dozens of AI+Crypto projects, among which a clear cycle could be observed.

Before the release of ChatGPT at the end of 2022, there were few blockchain projects related to AI in the secondary market, with mainly old projects like FET and AGIX coming to mind, and similarly few in the primary market.

January to May 2023 can be considered the first concentrated explosion of AI projects, as ChatGPT made a huge impact. Many old projects in the secondary market pivoted to the AI track, and almost every week, we could discuss AI+Crypto projects in the primary market. However, AI projects during this period felt relatively simple, many of which were “skin overlays” and “chain modifications” based on ChatGPT, with almost no technical core barriers. Our in-house development team could often replicate a project’s basic framework in a day or two. This led to many discussions about AI projects during this time, but ultimately, we did not invest in any.

From May to October, the secondary market began to turn bearish, interestingly, the number of AI projects in the primary market also significantly decreased during this period until the last one or two months when the quantity became active again. Discussions, articles, and other content about AI+Crypto also became richer. We once again entered a “boom” where we could meet AI projects every week. After half a year, the newly emerged batch of AI projects showed a clear improvement in understanding the AI track, the commercial application, and the integration of AI+Crypto compared to the first wave of AI Hype. Although the technical barriers are still not strong, the overall maturity has taken a step up. It wasn’t until 2024 that we finally made our first bet in the AI+Crypto track.

2. AI+Crypto Tracks

Vitalik Buterin, in his article “The promise and challenges of crypto + AI applications,” gives a forecast from several relatively abstract dimensions and perspectives: AI as a participant in games, AI as a game interface, AI as game rules, and AI as the game’s goal.

We summarize the AI projects currently seen in the primary market from a more specific and direct perspective.

Most AI+Crypto projects revolve around the core of Crypto, namely “technological (or political) decentralization + commercial assetization.”

Decentralization needs no explanation, as it’s intrinsic to Web3. Based on the type of assetization, it can be broadly divided into three main tracks:

● Assetization of computing power

● Assetization of models

● Assetization of data

2.1. Assetization of Computing Power

This is a relatively dense track, involving not just various new projects but also many old projects pivoting, such as Akash from Cosmos and Nosana from Solana. After pivoting, their tokens often surge, reflecting the market’s optimism about the AI track. RNDR, though primarily focusing on decentralized rendering, can also serve AI, hence many classify RNDR and similar computing power-related projects under the AI track.

The assetization of computing power can be further divided into two directions based on the use of computing power:

● Represented by Gensyn, “decentralized computing power used for AI training”

● Represented by most pivoted and new projects, “decentralized computing power used for AI inference”

An interesting phenomenon can be observed in this track, or rather a lack of enthusiasm for the chain:

Traditional AI → Decentralized Inference → Decentralized Training

Those with traditional AI backgrounds are skeptical about decentralized AI training or inference. Those focusing on decentralized inference doubt the feasibility of decentralized training.

The main reason is technical because AI training (specifically large models) involves massive data, and even more demanding than the data is the bandwidth requirement for high-speed communication of this data. In the current environment of Transformer large models, training these large models requires a power matrix equipped with numerous high-end 4090-level graphics cards/H100 professional AI cards + NVLink and professional fiber switches forming hundred G-level communication channels. The feasibility of decentralizing this, hmm…

AI inference demands far less computing power and communication bandwidth than AI training, making decentralized implementation much more plausible. This is why most computing power-related projects focus on inference, with training primarily undertaken by big players like Gensyn and Together, who have secured funding of over a hundred million. However, from the perspectives of cost-effectiveness and reliability, centralized computing power for inference remains far superior to decentralized at this stage.

This explains why decentralized inference skeptics view decentralized training as “impossible to achieve,” while traditional AI views decentralized training and inference as “technically unrealistic for training” and “commercially unreliable for inference.”

Some argue that when BTC/ETH first emerged, many thought the distributed node computation model was less reliable than cloud computing, yet it succeeded. The future of AI training and AI inference will depend on the demand for accuracy, immutability, redundancy, etc. Purely in terms of performance, reliability, and price, it’s currently unlikely to surpass centralized solutions.

2.2. Assetization of Models

This is another crowded track, and relatively easier to understand compared to the assetization of computing power, especially since ChatGPT’s popularity, one of its most famous applications has been Character.AI. You can seek knowledge from sages like Socrates and Confucius, chat casually with celebrities like Elon Musk and Sam Altman, or even express love to virtual idols like Hatsune Miku and Raiden Shogun, all thanks to the charm of large language models. The concept of AI Agent has deeply resonated with people through Character.AI.

What if agents like Confucius, Musk, and Raiden Shogun were NFTs?

Isn’t this AI X Crypto?!

So, it’s less about the assetization of models and more about the assetization of agents built on large models. After all, large models themselves cannot be put on the blockchain. It’s more about mapping agents based on models to NFTs to create a semblance of “model assetization” in AI X Crypto.

Now, there are agents that can teach you English or even be your romantic partner, among various others, including projects for agent search and marketplaces.

A common issue in this track is the lack of technical barriers, essentially NFTizing Character.AI. Our in-house tech experts can create an agent that speaks and sounds like BMAN using existing open-source tools and frameworks overnight. The integration with blockchain is also very light, similar to ETH’s Gamefi NFTs, where the Metadata may only store a URL or hash, with models/agents hosted on cloud servers, and the chain transaction is merely a transfer of ownership.

The assetization of models/agents will continue to be one of the main tracks in AI x Crypto in the foreseeable future. It’s hoped that projects with significant technical barriers and closer, more native integration with blockchain will emerge.

2.3. Assetization of Data

Logically, data assetization is the most fitting for AI+Crypto because traditional AI training mostly utilizes visible data from the internet, or more precisely, public domain traffic data, which might only account for 10–20%. Most data actually lies in private domain traffic (including personal data). If this data could be used for training or fine-tuning large models, we could have more professional agents/bots in various vertical fields.

Web3’s best slogan is Read, Write, Own!

Thus, through AI+Crypto, under decentralized incentives, releasing and assetizing personal and private traffic data to provide better and richer “feed” for large models sounds like a logical approach. Indeed, a few teams are deeply engaged in this field.

However, the biggest challenge in this track is — data is hard to standardize like computing power. For decentralized computing power, your GPU model directly translates to a certain amount of computing power, while private data’s quantity, quality, and purpose are hard to measure. If decentralized computing power is ERC20, then assetizing decentralized AI training data is like ERC721, mixed with many projects and traits like PunkAzuki, making liquidity and market creation much harder than ERC20. Therefore, projects working on AI data assetization are struggling.

Another notable aspect in the data track is decentralized labeling. Data assetization affects the “data collection” step, but collected data needs processing before feeding it to AI, which is the data labeling step. This step is currently a centralized, labor-intensive task. Transforming this labor work into decentralized work through token rewards, Label to Earn, or dispersing tasks like a crowdsourcing platform is a concept. A few teams are currently exploring this area.

3. The Missing Pieces in AI+Crypto

Here are the missing pieces in this track from our perspective:

3.1. Technical Barriers: As previously mentioned, the vast majority of AI+Crypto projects have almost no barriers compared to traditional Web2 AI projects, relying more on economic models and token incentives for front-end experience, marketing, and operations. While decentralization and value distribution are strengths of Web3, the lack of core barriers inevitably gives a sense of X to Earn. We are still hopeful for more teams with core technologies, like RNDR’s parent company OTOY, to make significant strides in the Crypto space.

3.2. Current State of Practitioners: From what we’ve observed, some teams in the AI X Crypto track are well-versed in AI but have a shallow understanding of Web3. Conversely, some teams are very Crypto Native but have a superficial knowledge of AI. This is very similar to the early days of the Gamefi track, where teams either understood games and thought about chain modifications for Web2 games or were proficient in Web3 and focused on innovating and optimizing play-to-earn models. Matr1x was the first team we encountered in the Gamefi track with a deep understanding of both games and Crypto, which is why I previously mentioned Matr1x as one of the three projects I decided on immediately after discussion in 2023. We hope to see teams with a deep understanding of both AI and Crypto in 2024.

3.3. Commercial Scenarios: AI X Crypto is at an extremely early stage of exploration. The various types of assetization mentioned above are just a few broad directions, each with potential niches that warrant detailed exploration and segmentation. The current projects in the market combining AI with Crypto often feel “forced” or “rough,” not fully leveraging the competitive strengths or composability of AI or Crypto. This is closely related to the second point mentioned above. For example, our in-house development team has conceived and designed a more optimal combination method. Unfortunately, after reviewing many projects in the AI track, we still haven’t seen any team entering this niche, so we continue to wait.