SN46 @instantsubnet might be one of the more interesting approaches to decentralized inference Instead of trying to solve the whole "can we trust the miner's hardware" problem with TEEs, they kinda flipped it around, miners provide the raw compute, while Instant runs its own code inside the miner setup through a virtual network runtime And the speed is kinda crazy too, 300 to 500 tokens/sec on open-weight models like GPT-OSS 20B. That's way faster than what you'd normally get from API providers like OpenRouter And apparently this is only the first step, the longer-term goal is getting into analog chip architectures where the model weights are basically built into the hardware If they can pull that off, decentralized inference on $TAO could get really damn fast