“Inference workloads represent "well in excess of 50%" of the total infrastructure utilization on CoreWeave's platform.”