[ghughes]

But given the rumored architecture (MoE) it would make complete sense for them to dynamically scale down the number of models used in the mixture during periods of peak load.