I was reading through OpenGradient's architecture documentation today and found a detail marked "Coming Soon" that changes how the current live network actually operates. Data Nodes, the dedicated layer responsible for serving model weights to inference nodes, are not yet live. The whitepaper describes them as a planned role. The network runs without them today.

That matters because inference nodes are explicitly described as stateless. They do not store model weights locally. Every inference request requires streaming weights from somewhere first. Without dedicated Data Nodes, that somewhere is Walrus directly, the decentralized storage layer built on Sui using Red Stuff erasure coding at 4.5x replication factor.

What I find genuinely worth examining is what that means for inference latency today versus what the architecture intends once Data Nodes arrive. Streaming model weights from a decentralized storage network for every inference request is a different performance profile than streaming from a specialized node designed and optimized specifically for that job. The gap between those two arrangements grows larger as model sizes grow. A 7 billion parameter model requires roughly 14 gigabytes of weights. Serving that across a decentralized storage network on every cold inference request introduces a bandwidth demand the architecture acknowledges it has not fully solved yet.

Data Nodes exist in the documentation as the intended solution. They do not exist on the network yet as the actual one.ReplyForwardAdd reaction

#opg $OPG @OpenGradient