How they use Nebius AI Cloud

SemiAnalysis uses Nebius-supplied compute for InferenceX, formerly InferenceMAX. The project tests inference frameworks, models and hardware, and credits Nebius among its compute supporters. Nebius also confirms this compute arrangement in its partner profile. Its Blackwell Microbenchmarks project credits Nebius and Verda for sponsoring compute for tests of GPU memory throughput, latency, tensor-core performance and HBM-resident elementwise throughput. Nebius and Verda provided B200 nodes with hardware counters enabled for NCU profiling.

For ClusterMAX 2.0, SemiAnalysis tested Nebius’s managed Soperator Slurm-on-Kubernetes service. The cluster included preinstalled drivers and Docker, passwordless SSH, and expected performance on nccl-tests and TorchTitan pretraining jobs. Its tests confirmed bare-metal-class performance on Nebius’s KubeVirt virtual machines. The test cluster exposed a 197 TB shared root filesystem through virtio-fs. A 22.1 GB PyTorch container pull took 45–60 seconds. Initial monitoring used a Grafana dashboard accessed through an SSH port forward, followed by dashboard and health-check updates. Nebius reports that ClusterMAX 3.0 testing ran over several months on NVIDIA GB300 and NVIDIA HGX B300 clusters, covering training, inference, concurrent storage workloads, and node and full-rack recovery. SemiAnalysis tested storage scaling across two racks. Nebius retuned the filesystem after reported errors, and SemiAnalysis verified their resolution. A synthetic error injection on its Nebius GB300 rack also tested automatic node recovery. The node returned to service after 8 hours and 40 minutes.

SemiAnalysis also rates providers through ClusterMAX and has produced research commissioned by Nebius.