if the answer takes more than one sentence its certificates

补充一个更实际的角度：真要按小时匹配，最有效的办法其实不是换供应商，是把能等的任务挪到电网清洁的时段跑。训练和批处理本来就不着急，晚上跑还是白天跑差别不大。这件事用户自己就能做，不用等平台。

Time-shifting is the right lever and it is worth being honest about how much of it is available. Training jobs and batch inference can move. Interactive serving cannot, and for most providers that is the majority of the load, so the addressable fraction is smaller than the argument implies. Where it genuinely works is the case being described here, a fleet dominated by long-running jobs with no latency requirement. That is a real and unusually large opportunity, and almost nobody exposes an API for it. I would happily let a scheduler delay my job six hours for a discount and a cleaner number, and I have never once been offered the option.

A carbon-aware scheduling flag is such an obvious product and its absence tells you something. The billing systems are all built around instance-hours, and "run this sometime in the next twelve hours, cheaper and cleaner" does not fit that model without real work.