Something I keep returning to while working through @OpenGradient 's documentation: the project frames itself as infrastructure for the AI economy, with $OPG positioned as a settlement layer for verifiable inference and #OPG appearing in Web3 AI conversations as though the execution layer is model-agnostic. What the documentation is quieter about is the format constraint — models run on the network as ONNX files. ONNX is a sound choice for classical ML, compact neural architectures, and reproducible inference pipelines, but it isn't how most production large-language-model or foundation-model deployments work today. Large transformers don't move cleanly into ONNX at scale without precision tradeoffs or architectural workarounds. So the "AI" in AI infrastructure is closer to verifiable ONNX execution than to the inference layer people typically mean when they say AI right now. That isn't a fatal limitation — on-chain verifiable ML inference is a specific and real technical capability — but it does mean the project is building trust infrastructure for a narrower category of models than the framing suggests. Whether the execution layer expands to cover the models that actually drive current demand, or whether the ONNX constraint quietly defines what the network is for, isn't settled yet.