So The GPUs Never Wait
Parallel file and object storage, provisioned and tuned with compute, networking, Kubernetes, and Slurm as one environment. Built for training and inference, with residency and isolation enforced at the storage layer.
Three storage services
Parallel File
High-throughput, for training and checkpointing
Object
S3-compatible, for datasets and artifacts
Block
Persistent volumes for K8s workloads, endpoints, notebooks
Provisioned through partner platforms and integrated into the operating layer, so each workload gets the right storage pattern without hand-built paths per cluster.
.avif)
Residency, isolation, governance
Residency
Data held within defined jurisdictions.
Isolation
Separated by tenant, project, and workload.
Access
Governed by policy.
Encryption
Protecting data.
Audit
Operational trails across the operating domain.
Checkpointing
Storage paths are tuned for frequent checkpointing and rapid recovery, so GPU capacity isn't idle waiting on the storage layer during long runs.
Artifact lifecycle
Lifecycle policies across active datasets, retained artifacts, and long-lived assets:
Retention
How long datasets, artifacts, and outputs are held.
Replication
Copies maintained across systems and locations.
Backup and recovery
Recoverable state for active and retained data.
Performance visibility
Monitor throughput, latency, IOPS, saturation, errors, and capacity pressure across tenants, clusters, and datasets - in the context of the workloads consuming the storage.
