From the abstract, this seems to be a scheduling mechanism for datacenters that serve multiple models. I have no idea whether this applies to Anthropic.