We are seeing high contention on H100 hardware which is resulting in delays on predictions and scale-out.