Here’s an interesting idea; Zettalane has built MayaNAS, a Lustre parallel filesystem, with NFS, pNFS and SMB support, based on cloud object storage.
It says MayaNAS can stream AI training at GPU speed, checkpoints land at tens of gigabytes per second, and capacity costs what object buckets cost. MayaNAS has been validated by MLPerf Storage v3 benchmarks, recording 32.42 GB/s Llama 3 70B checkpoint write bandwidth, using two clients. On RetinaNet, 72 B200 GPUs were fed from one 200 Gb/s MayaScale client.
Santa Clara, CA-based Zettalane was started up in June 2018 by CTO Supramani (Sam) Sammandam, a storage vet with more than 25 years in systems software at Lucent Technologies, IBM, HP, Dell and others. No VC funding has been recorded
MayaNAS runs Lustre and OpenZFS with a design that changes storage economics: each OST (Object Storage Target) is an OpenZFS dataset whose vdevs are regional, Standard-class Google Cloud Storage (GCS) buckets. Bulk data is read and written directly and concurrently to object storage through our objbacker layer, while a small NVMe special vdev holds only pool metadata and small blocks. There is no local SSD or ephemeral scratch tier in the data path
Sammandam tells us: “In the MLPerf Storage v3.0 results: Zettalane’s MayaNAS was the only submission to run a standard Lustre parallel file system directly on cloud object storage — regional GCS as the data path, on standard cloud VMs, no local SSD, in the customer's own account.
“Verified by MLCommons, it's a different tier from the flash results (Everpure et al.): parallel-filesystem throughput at object-storage economics, and not a GiB/s race. And it covered the full pipeline without staging data to NVMe — including the POSIX-only KV cache that S3-native stacks can't serve.”
We asked: “In a way you are like Nasuni, with a file system using object storage as a foundation?”
Sammandam replied: “Yes — at a high level, like Nasuni: a file system with object storage as its foundation. Others have taken that route by writing their own file system and then having to work hard to prove POSIX compliance. We took a different approach — building on battle-tested OpenZFS and its native separation of metadata (on a small NVMe device) from bulk data (on object). We presented that at the OpenZFS Developer Summit 2025, and the Lustre design at LUG 2026.
“Nasuni is cache-forward — an edge filer serves from a local cache with the authoritative copy in object, so performance tracks how much of the working set fits the cache. MayaNAS reads and writes directly to object, with no local data-cache tier in the path; in MLPerf the reads were served cold from object at GPU scale — verifiable in Google's own Cloud Monitoring metrics.
“The bigger difference is what it is: MayaNAS is a standard Lustre and pNFS parallel file system (it serves NFS/SMB too), on the stock in-kernel Linux clients — built for parallel, GPU-scale throughput, which is why it fits a benchmark like MLPerf. Our cloud-native pNFS builds on LANL's open-source MDS (metadata server). Nasuni targets distributed enterprise file; MayaNAS does enterprise NAS as well and extends into HPC/AI parallel throughput straight on object.
“That's really the whole idea. Object storage is clearly where AI storage is heading — but adopting it usually means rewriting applications to speak S3. MayaNAS addresses that directly: standard Lustre or pNFS on object storage, so you get object economics with nothing to rewrite.
“There's more here … the OpenZFS-on-object internals, the pNFS and LANL MDS work, MayaScale NVMe-over-TCP block, and branchable Postgres for Agentic AI, all on the same foundation.”
Read more in a Zettalane blog about its MLPerf Storage v3 submission. Watch an iTWire TV video here - but set aside 70 minutes to do do.