How they use Nebius AI Cloud
Cedana uses Nebius for large GPU tests in its nightly Kubernetes test workflow. Its Nebius configuration specifies CUDA, amd64 and local checkpoint storage. A recorded nightly job reached the test step and failed after 22 minutes and 1 second.
The Kubernetes workflow restricts the Nebius test selection to large GPU workloads. The nightly configuration enables cloud Kubernetes testing and identifies it as the place where Nebius tests run. This documents software testing on Nebius rather than a production deployment of Cedana's customers.
Commercial use. Uses Nebius for large GPU workloads in its nightly Kubernetes software tests.
Sources
3 sourcesCedana Kubernetes test workflow
Listed 28 Oct 2026 · GitHub · Web page
Cedana configures Nebius Kubernetes tests for CUDA on amd64 with local checkpoint storage. The workflow selects large GPU tests for Nebius.
Open original sourceCedana nightly test workflow
Listed 28 Oct 2026 · GitHub · Web page
Cedana's nightly workflow enables cloud Kubernetes tests and identifies this workflow as where its large Nebius tests run.
Open original sourceCompany context
cedana.comCedana, operated by Cedana Systems, Inc., builds software to checkpoint, migrate and resume running CPU and GPU workloads. It serves AI inference providers, GPU cloud operators, research computing teams and pharmaceutical and biotech organizations. Its infrastructure captures running process state so workloads can resume from saved progress after interruptions or move between compatible machines. Cedana offers an open-source daemon and managed orchestration capabilities, with Kubernetes and Slurm integrations for inference, training and high-performance computing workloads.
Recent updates
2 updatesCedana adds DeepInfer deployment support with NVIDIA Dynamo
Cedana announced a ready configuration for running DeepInfer on NVIDIA Dynamo. It reports restoring DeepSeek-V4.1-Flash on eight B200 GPUs in 36 seconds, while noting that prefill/decode disaggregation and the required KV cache system remain incomplete in this configuration.
CedanaCedana and NVIDIA integrate tools for serving coding models on one GPU node
Cedana and NVIDIA presented an integrated system combining Cedana, NVIDIA Dynamo and NeMo Switchyard to serve multiple coding models on a single eight-B200 node. Cedana reports model restores of 57–70 seconds while preserving saved inference sessions, with deployment through a single Helm chart.
Cedana