Million IOPS
NVMe-backed volumes deliver over 1 million IOPS per instance, with sub-100µs latencies. Train against your dataset without I/O becoming the bottleneck.
/ STORAGE
High-performance storage for AI data
Tscale Storage delivers the data layer AI workloads actually need — NVMe block, S3-compatible object, Lustre parallel filesystem, and dataset caching. Built for the terabyte-to-petabyte scale modern training and inference demand.
Unified Namespace
One API across block, object, and parallel filesystem. Move data between tiers without changing application code or rewriting pipelines.
Encrypted By Default
AES-256 encryption at rest, in transit, and during replication. Customer-managed keys via KMS. SOC 2, ISO 27001, and HIPAA-aligned out of the box.
/ STORAGE TIERS
The right tier for every workload
AI workloads have wildly different storage profiles. Tscale ships with four purpose-built tiers — pick the one that matches your access pattern, or compose them in a single pipeline.
- Local NVMe — the fastest tier. Per-instance scratch, ephemeral, lost on termination.
- Block storage — persistent NVMe volumes, attach to any instance, snapshot & restore in seconds.
- Object storage — S3-compatible, petabyte-scale, lifecycle policies, 11 nines of durability.
- Lustre parallel FS — HPC-grade throughput for tightly-coupled distributed training.
/ DATA PIPELINE
Sustain 1+ TB/s end-to-end
Modern multi-modal training reads datasets at terabytes per second. Tscale’s storage stack is built to keep up — RDMA from NVMe to GPU, parallel Lustre for shuffle, S3 for cold archive. No stalls, no waiting on disk.
-
Direct RDMA to GPU
GPUDirect Storage bypasses the CPU entirely — 8-10x higher throughput on the same hardware.
-
Dataset caching
Region-local hot caches mean your second training run starts 10x faster than the first.
-
Tier-aware orchestration
Datasets flow between tiers automatically — hot working set on NVMe, warm on Lustre, cold on object.
/ PLATFORM
The full data platform
Storage isn’t a box — it’s a complete data platform. Snapshots, replication, lifecycle policies, dataset lineage, and integration with the rest of Tscale come standard.
Data Protection
- Snapshots (point-in-time)
- Cross-region replication
- Object versioning
- Cross-region failover
- 11 nines durability
Lifecycle
- Tier transition policies
- Automatic archival to cold
- Scheduled deletion
- Cost analytics
- Quota enforcement
Dataset Tools
- Dataset versioning (DVC)
- HuggingFace integration
- Streaming ingestion
- Deduplication & PII redaction
- Lineage tracking
Access Protocols
- S3 API (all object storage)
- POSIX / NFS (block + Lustre)
- NVMe-oF (RDMA)
- CSI driver (Kubernetes)
- SFTP / rsync
Security
- AES-256 at rest
- TLS 1.3 in transit
- Customer-managed KMS
- VPC isolation
- Audit logging
Access
- Web Console
- Tscale CLI
- REST API (Radar)
- Terraform provider
- Python SDK
Performance
Per block volume
NVMe-backed, sub-100µs latencies. Saturate your GPUs without I/O bottlenecks.
Per Lustre FS
HPC-grade aggregate throughput across thousands of OSTs and MDTs.
99.999999999% object durability
Erasure coding across 3+ availability zones. Your data survives anything.
No per-bucket limits
Store a billion objects or a single 100 TB checkpoint. Tscale scales with you.
For teams with data gravity
Sovereign Storage
Data stays in your jurisdiction — Lagos today, expanding to additional African and European regions. Full data-residency compliance built in.
Learn MoreCross-Region Replication
Replicate critical datasets across regions in milliseconds. Sub-10ms RTT for active-active topologies.
Learn MoreThe data feeds the compute
Storage is where the AI lifecycle starts and ends. Pair it with compute, instances, and managed Slurm to run the full training pipeline on one platform.