The team behind vLLM's continuous batching technique is launching a platform to monetize idle neocloud GPUs — and claims 2–3x token throughput vs. vLLM.