One engine on your cloud's object storage — a parallel storage engine with Lustre and pNFS to feed GPUs at scale, and enterprise NAS with NFS, SMB, iSCSI and NVMe/TCP for everything else. Capacity costs what your buckets cost.
One namespace for hundreds of GPU clients — training data, checkpoints and inference caches.
Shared file and block storage for everything around AI — and everything else.
One namespace, hundreds of compute clients, striped throughput that scales with the cluster — the parallel filesystem large GPU clusters train on, without the operations burden. Stock in-kernel Lustre client, nothing to build; on GKE, Google's own Lustre CSI driver mounts it.
Parallel NFS with the client already in Linux — metadata routes through an open metadata server while data flows directly to MayaNAS data servers.
Checkpoints write at tens of gigabytes per second straight into object storage — 14.43 GiB/s from a single client and 32.42 GiB/s from two in MLPerf® Storage v3.0 — so training spends its time training.
Linux, macOS and Windows clients on the same shares — home directories, tools and notebooks, media projects. Full POSIX, nothing to install.
Block volumes for applications that want a disk rather than a share — iSCSI for broad compatibility, NVMe over TCP for low latency — from the same engine and the same pool.
Instant point-in-time snapshots for recovery, and replication for data protection and disaster recovery.
Active-active server pairs with automatic failover. Clients reconnect and carry on — no manual recovery steps.
End-to-end checksums detect and repair silent corruption, and transparent compression stretches capacity further.
Data written through MayaNAS lives in your bucket — no cache tier to size and warm, and capacity grows at bucket prices.
MayaNAS deploys into your own account and network. Your datasets, checkpoints and models stay in your buckets — no third-party service in the data path.
Choose the region, keep your own encryption keys on your buckets, and meet data-residency requirements without waiting for a managed service to reach your geography.
Bring the same filesystem up next to whichever GPUs you get — one source of truth in object storage, one way to deploy, on Google Cloud and Azure.
In-kernel Lustre, NFS and SMB clients. Nothing proprietary to install on your GPU nodes; Kubernetes mounts it with the drivers you already use.
Pick your cloud to see what's available on its marketplace — deployed into your own account and billed through the cloud agreement you already have. Prefer a guided first run? Talk to us.
Evaluating first? The open-source Community Edition deploys the same Lustre with one command on Google Cloud and Azure.
Self-serve with terraform from your cloud marketplace — ready for your most demanding training runs.
MLPerf® Storage v3.0, Closed division, submitted by ZettaLane Systems on Google Cloud. Entries 3.0-0137 (Llama 3 8B checkpointing) and 3.0-0139 (Llama 3 70B checkpointing). Result verified by MLCommons Association. Retrieved from mlcommons.org/benchmarks/storage/ on 7 September 2026. The MLPerf name and logo are registered and unregistered trademarks of MLCommons Association in the United States and other countries. All rights reserved. Unauthorized use strictly prohibited. See www.mlcommons.org for more information.