ZettaLane's MayaNAS posted 32.42 GB/s checkpoint write bandwidth on a Llama 3 70B model in MLPerf Storage v3.0, the only Lustre parallel filesystem to run directly on cloud object storage in the benchmark round.

That figure, recorded from two clients, came without local SSD in the data path. The system relies on regional Google Cloud Storage buckets as bulk storage, with a small NVMe device holding only pool metadata and small blocks.

Santa Clara-based ZettaLane was incorporated in June 2018 and has remained self-funded. Founder and CTO Supramani Sammandam returned to the company full-time in 2026 after a Technical Staff stint at Dell's Office of the CTO, where he watched VAST and WEKA win AI infrastructure deals.

What does the architecture actually change?

The system treats object storage as primary, not as backup. NVMe holds metadata only. There is no local data-cache tier in the path. The MLPerf read workload ran cold from object at GPU scale, a claim ZettaLane says is verifiable in Google Cloud Monitoring metrics.

On a RetinaNet workload, 72 NVIDIA B200 GPUs drew data from a single 200 Gb/s MayaScale client. That ratio is the headline number: one network endpoint feeding seven dozen accelerators, with the working set living in object buckets rather than in flash.

How does ZettaLane differ from Nasuni?

Sammandam drew the comparison himself. Both pair a filesystem with object storage underneath. The split comes in caching philosophy.

"Like Nasuni: a file system with object storage as its foundation," Sammandam said. "Nasuni is cache-forward — an edge filer serves from a local cache with the authoritative copy in object. MayaNAS reads and writes directly to object, with no local data-cache tier in the path."

The deeper difference is workload target. Nasuni sells distributed enterprise file. MayaNAS serves enterprise NAS and extends into HPC and AI parallel throughput on object. The MLPerf submission validated that path on stock in-kernel Linux clients, with cloud-native pNFS built on an open-source metadata server originally developed at Los Alamos National Laboratory.

The technical premise sits on OpenZFS. Each Object Storage Target in the Lustre setup is an OpenZFS dataset whose vdevs are regional, Standard-class GCS buckets. Bulk reads and writes go straight to object storage through a layer the company calls objbacker.

What does this mean for AI storage economics?

MLPerf submissions typically focus on raw throughput. ZettaLane's bet is that price-per-gigabyte-seconds matters as much as GiB/s alone, since training runs are growing longer and datasets larger. Running Lustre against GCS Standard class inherits the per-gigabyte pricing of object rather than the per-gigabyte pricing of flash.

The company is not alone in chasing that economics curve. Pure-play AI storage vendors have spent three years writing custom filesystems on top of object or block primitives. ZettaLane's route is to skip the custom filesystem and lean on OpenZFS and Lustre, both already in the Linux kernel.

No major customer references exist yet. Sammandam acknowledged that gap and said the MLPerf verification is the proof point he is taking to market. The next milestone is a production deployment, and the founder is hoping a major object-storage vendor will partner with the company.

For the semiconductor industry, the relevant data point is the ratio: 32.42 GB/s checkpoint write on two clients, 72 B200 GPUs fed from one 200 Gb/s endpoint, and not a single SSD in the data path. Storage may finally be moving from a flash-priced line item to an object-priced line item, assuming ZettaLane can convert MLPerf scores into purchase orders.