Walrus Protocol and the New Data Layer for AI on Sui
Programmable storage that AI teams can actually use, with verifiable provenance, low‑friction access, and chain‑agnostic tooling.
This is a sort of follow-up / part 2 of my prior post “Four Power Tokens Fueling Sui’s 2025 Flywheel” so feel free to go back and read this first if you are interested. So much has happened since then.
When the Walrus mainnet went live on March 27, 2025, it turned a big idea into an operating network: make large data “programmable,” not just parked somewhere off-chain. Walrus is built by the original Sui contributors at Mysten Labs, it stores “blobs” of arbitrary binary data across a decentralized set of operators, and it uses Sui as a control plane to make that data addressable, ownable, tradable, and automatable through Move smart contracts. That combination is why AI projects — from agent frameworks to decentralized compute — are picking Walrus as their storage backbone.
In late July, Walrus also introduced Quilt, a batch‑storage API that slashes overhead for many small files — think agent messages, logs, and dataset shards — while keeping the ability to fetch single files quickly. As of July 2025, Walrus reported 800+ TB encoded across 14M blobs, with Quilt expected to hit mainnet the week after.
1 | How Walrus works (and why it’s different)
Red Stuff encoding, self‑healing availability. Walrus’ two‑dimensional erasure coding (“Red Stuff”) spreads data across nodes so the network can rebuild missing pieces efficiently, with a target replication factor around 4–5x rather than 100x‑style full replication. The peer‑reviewed paper describes self‑healing recovery that only moves the lost data, plus availability challenges that work in asynchronous networks.
Sui as a programmable control plane. Walrus writes an on-chain Proof‑of‑Availability (PoA) to Sui, and represents storage resources as Move objects. That lets builders automate renewals, gate access with smart contracts, or compose data with other on-chain logic. Walrus is chain‑agnostic for apps; you can build on Solana, Ethereum, or your own L1 and still use Walrus via its APIs.
Economics with WAL. Payment, security, and governance are anchored by the WAL token. Max supply is 5,000,000,000 WAL, with an initial circulating supply of 1,250,000,000 WAL and a distribution that is majority community‑oriented via airdrops, subsidies, and a community reserve. The design aims to stabilize storage in fiat terms and introduces future burn mechanisms tied to slashing and stake‑shift penalties.
2 | Why AI builders care
AI systems aren’t just compute bound, they’re data‑bound: models, fine‑tuning artifacts, agent memories, conversation logs, and evaluation evidence all need durable, queryable storage with clear provenance. Walrus’ value props for AI teams keep showing up in partner case studies:
Programmable access control (e.g., with Mysten’s Seal for decentralized secrets), enabling privacy‑preserving model or data access.
No download fees and predictable economics for heavy read workloads, a big lever for decentralized inference and data generation networks.
On-chain verifiability for claims, reasoning artifacts, and proofs that underpin “verifiable AI.”
3 | The AI partnerships and projects building on Walrus
Below is a complete roundup of announced AI‑related integrations so far, with what each is actually doing on Walrus.
io.net: BYOM on decentralized GPUs, models stored on Walrus.
io.net, a large decentralized GPU network, is integrating Walrus to launch a Bring Your Own Model platform. Models can be stored privately on Walrus, then pulled into io.net GPU clusters for training or inference with pay‑as‑you‑go billing and private compute execution.
OpenGradient: “user‑owned AI,” 100+ models hosted on Walrus.
OpenGradient replaced its legacy IPFS setup with Walrus and now hosts 100+ models on a programmable, verifiable storage layer. The roadmap includes smart contract enforced access control via Seal, new monetization strategies, and support for larger models.
Gata: open execution infrastructure for AI and a decentralized data factory.
Gata uses Walrus to make decentralized AI economically viable at scale. Its DataAgent coordinates browser‑side contributors to generate synthetic datasets; Walrus removes download fees and provides predictable storage economics for the many small, frequently accessed files that training pipelines consume. Quilt (mentioned above) support further improves small‑file economics.
Talus: on-chain AI agents with Sui + Walrus.
Talus builds tokenized agents whose workflows execute on Sui, reading and writing larger data — training sets, context, memory — to Walrus. The Nexus framework orchestrates reads and writes so agents can fetch models and state from Walrus without centralized storage risk.
Swarm Network: “verifiable AI” for real‑time fact‑checking on Rollup.News.
Swarm stores claims data today, and will add agent logs, media evidence, and reasoning artifacts to Walrus, using Tusky as a Walrus file manager. The goal is an open, audit‑ready knowledge graph for agent outputs.
Itheum: tokenizing data for musicians and AI agents.
Itheum lets creators and agents tokenize high‑value assets like master audio tracks and AI models, using Walrus to handle large files with reliable performance and programmability.
Chainbase: omnichain data network for AI, migrating 300 TB to Walrus.
Chainbase is integrating Walrus to store raw data for 220+ blockchains and its ~300 TB dataset, enabling fully decentralized data pipelines for DeFi, AI, and Web3 via its Manuscript framework.
Veea — edge AI meets decentralized storage.
Veea is adopting Walrus as a VeeaHub STAX edge solution. The plan pairs Walrus’ programmable storage with NVMe‑powered edge nodes for low‑latency data access, hot‑content delivery, and AI workloads at the edge. The companies highlighted internet‑less transaction demos and position the stack for DePIN and enterprise scenarios.
Additional ecosystem signals: media partners such as Decrypt and Unchained are moving archives onto Walrus; Pudgy Penguins is migrating creative assets; Crossmint added Walrus as a first‑class storage option in its minting APIs. These aren’t AI per se, but they validate throughput, reliability, and tooling maturity.
4 | Developer’s view: building AI on Walrus
Data as a first‑class, on-chain resource. Because storage resources and blob references live as Sui objects, you can automate renewal, enforce access with Move, and compose data with other on-chain programs. That removes a lot of glue code teams build around object stores.
Architect for scale with Red Stuff. For large models and datasets, you get fast reads and self‑healing resilience without paying a central provider. For chatty agents and telemetry, I again remind you that Quilt lets you bundle thousands of small files into one blob, cutting overhead by orders of magnitude while preserving random access.
Chain‑agnostic, Sui‑powered. Use Walrus from any stack you already have, rely on Sui for verifiable availability, and opt into Seal when you need decentralized secrets and encrypted access.
Getting started: read “Bringing Programmability to Data Storage,” scan the Red Stuff explainer and paper, then hit the docs and SDKs. If you’re optimizing small‑file heavy workloads, pilot Quilt now.
5 | What to watch next
Quilt on mainnet (target July 30, 2025), which should improve cost curves for agentic systems, messaging, and NFT collections.
Programmatic access at scale, as partners lean into Seal‑backed permissioning and tokenized data flows.
Edge deployments, where Veea + Walrus could bring low‑latency data locality to AI inference and data‑heavy experiences outside centralized clouds.
6 | Final take
Walrus turns decentralized storage into a programmable substrate that AI builders can compose with their compute, their agents, and their economics. The architecture: Red Stuff for resilience and efficiency, Sui for verifiability and programmability, and WAL for aligned incentives, is now backed by a cadence of real AI integrations. If your AI roadmap needs verifiable provenance, predictable reads, and on-chain control, you should kick the tires.
Please follow me on X if you are not already: https://x.com/1111bigdog
Disclaimers
I am just a big, dumb animal and I’m totally open to feedback / corrections / questions so please comment with any thoughts.
I am not a financial advisor and this is not financial advice.
Don’t eat yellow snow.
Bigdog is on a roll.