The xAI cluster, equipped with around 500,000 NVIDIA GPUs, is currently utilized at only about 11%, significantly lower than the typical 40% utilization seen in other AI labs. The underuse is partly due to intermittent training processes and delays caused by bandwidth limitations and network latency when synchronizing tens of thousands of GPUs. To maintain the appearance of higher utilization, some teams reportedly rerun completed tests to report better usage metrics.

Source: theinformation.com