Storage for AI workloads

Scalable storage solutions for building and using generative AI on Nebius AI Cloud.

High-speed dataset streaming

Feed datasets to your GPU cluster at maximum speed, ensuring faster training cycles and low-latency model inference.

Rapid checkpoints

Achieve the highest AI infrastructure goodput by employing high-speed shared storage for writing and reading checkpoints during multi-host training.

Ready for multi-modality

Store any volume of unstructured data — from text to videos — to streamline the operations for multi-modal AI training.

Nebius Object Storage

Fully S3-compatible object storage capable of storing static data, serving it to external consumers and providing high-performance data streaming for MLOps scenarios. You can seamlessly transfer your data between storage classes within the same service, to meet your data strategy requirements.

Standard class

  • Capacity-focused object storage buckets
  • Great for storing large volumes of unstructured data
  • Cost-efficient for static data
  • Unlimited scalability

Enhanced class

  • Performance-focused object storage buckets
  • Great for streaming data to GPU and checkpointing
  • Up to 2 GiB/s write throughput per GPU*
  • Unlimited scalability.

Nebius Shared Filesystem

A high-speed shared filesystem designed for AI workloads, delivering scalable performance for parallel AI computation and unlimited volume scalability. The primary choice for customers running training and inference workloads, it offers cost efficiency, ease of use and a rich feature set.

  • High-performance class based on all-flash NVMe storage.
  • Delivers over 500 GB/s** aggregate read performance.
  • Fully integrated with the Nebius cloud platform.

Third-party storage solutions

Strategic partnerships with leading vendors expand our storage lineup, making it easier for customers to choose the best option for their AI infrastructure.

WEKA and Nebius partner to catalyze AI innovation with an ultra-high-performance cloud infrastructure solution.

VAST Data and Nebius partner to accelerate global enterprise AI adoption with an on-demand AI cloud.

Block network storage

Block network volumes for boot and run virtual machines. You can choose one of three options of network disks that differ by performance, reliability and pricing:

  • Block volumes with 3x mirroring
  • Block volumes with erasure coding
  • Block volumes without data replication

Custom-designed storage hardware

To ensure enterprise-level reliability and exceptional power efficiency, we use proprietary hardware designs for most AI infrastructure components, including NVMe-based storage servers.

* Depends on the structure of data stored in the bucket, write concurrency and configuration of upload process parameters.

** Burst performance from a cluster with 254 GPU and 254 CPU hosts, measured during multi-modal training by a real customer