AI Crypto Has a Data Problem: How Much of the Sector Is Actually Decentralized?

The decentralized nature of many proposed decentralized AI implementations relies mainly on the distribution of graphics processing units. However, distributed datasets, which neural networks need, may still be subject to centralized access control.

Why Decentralized AI Still Depends on Centralized Data

AI Models Need More Than Decentralized Compute

Training such models requires GPUs, and datasets and software to prepare and access data. A licensing market is developing for training AI models, including curated and higher-quality datasets, according to the U.S. Copyright Office.

Decentralized AI compute mainly decentralizes compute used in training rather than data: i.e., separate computers conducting the training, but one developer/rights holder controlling the corpus used to train the model.

Who Actually Owns the Data Used to Train AI?

There is no single answer to the question of who owns AI training data. Datasets may contain materials with different copyright, licensing, or access terms, and not all publicly accessible materials are in the public domain.

This creates a basic challenge for decentralized artificial intelligence: distributing the infrastructure is not sufficient to distribute the legal rights over the data used to train AI.