Store 2.5x more AI training data for the same hardware budget.
Leil runs a distributed file system built natively for high-density SMR drives. POSIX-compatible, linearly scalable, and ~60% lower cost per TB than all-flash clusters.
AI training clusters default to flash. They should not.
All-flash storage was built for latency-critical, random-access workloads. But AI training reads are sequential and throughput-bound. The latency premium of NVMe buys nothing when your training loop is reading tokenized text in order, front to back.
SMR hard drives deliver 2.5x the capacity density at one-third the cost per TB. The engineering problem is write alignment: SMR's shingled tracks require sequential write ordering that generic filesystems break constantly.
Leil is the filesystem layer that solves this natively, in the kernel path, without application changes. Your training code sees standard POSIX semantics. The drives run at their rated throughput.
Read the technical designThree steps from bare drives to POSIX mount.
Add SMR Nodes
Any Linux host with high-density SMR drives (8 to 20TB per drive). Leil handles drive detection and band geometry automatically.
Band-Aligned Writes
Leil's write scheduler batches concurrent writes into SMR-band-sized sequential streams. No application modification required.
POSIX Mount Point
Your training cluster mounts a standard POSIX filesystem. PyTorch, JAX, HuggingFace DataLoader: no changes required.
Measured in production, not estimated.
| Metric | Leil Storage | NFS on NVMe | Object Storage (S3) |
|---|---|---|---|
| Cost / TB / month | ~$22 all-in | $60 to $80 | $15 (POSIX unavailable) |
| POSIX compatible | Yes | Yes | No |
| Sequential write throughput | 1.4 GB/s per node | 2.0 GB/s per node | N/A |
| Checkpoint resume (100GB) | ~72 seconds | ~35 seconds | Hours via SDK |
| Setup complexity | apt install + mount | Requires NFS server config | Requires SDK rewrite |
What early-access teams are seeing.
We were paying $18,000 a month for NVMe storage on our pre-training pipeline. After running the Leil pilot on the same cluster with SMR drives, we cut that to under $7,000 while keeping the same dataset throughput.
The POSIX compatibility was the blocker for us. We could not use object storage because our checkpoint libraries expect a real filesystem. Leil just worked. Mounted it in an afternoon.
Get started with early access.
Leil is in design-partner stage. We work closely with the first 20 clusters to tune performance for your workload.
Full pricing details, FAQ, and early-access sign-up on the pricing page.