After messing around in the Web3 storage space for so long, my biggest takeaway is that projects generally have their shortcomings. Some only focus on the storage layer and have a superficial understanding of application scenarios; others hype up their solutions but are far from practical implementation. Truly connecting all the links in the chain is rare.

Walrus's series of actions in early 2026 changed this situation. Especially the deep collaboration with io.net, which truly opened my eyes — they connected storage and computing power in a real sense.

Previously, training AI models was a hassle. Datasets were stored on decentralized platforms, but model training had to be transferred to centralized GPU clusters. Data transfer back and forth was time-consuming and posed privacy risks. Now, things are different. Through the BYOM (Bring Your Own Model) platform, users can directly store custom AI models in Walrus and then utilize io.net’s distributed GPU clusters for training. The entire process keeps data within the storage layer, without leaving it.

A friend of mine working in AI image generation recently conducted a practical test. He stored a 5GB art style dataset in Walrus, used io.net’s GPU resources to fine-tune the model, and the results were astonishing — training costs were 60% cheaper than on AWS, and thanks to Walrus’s built-in privacy computing mechanisms, data security was ensured, while training efficiency increased by 30%.

This integrated solution of storage plus computing power is still quite rare in the Web3 field. The addition of Zark Lab further enhances the AI intelligence layer, and the combined power of this lineup in 2026 is worth continuous observation.