My guess is that they are dynamically changing the quality of the model to always keep the speed above some floor. So once it gets below that they switch to a worse quant or reduce reasoning level, or some combination of both.