all the inference providers including big clouds want customers to take the risk for idle GPUs by making them reserve throughput this isn't surprising remember before AWS you could not at scale rent servers by the minute but it is a bit jarring since we're so used to that