Every byte, one hop from the GPU
Triple-replicated NVMe block volumes from $0.08/GB-month and S3-compatible object storage at $8/TB-month — on the same private fabric as your compute, with free ingress and 20 TB of pooled egress.
11×9s designed durability · Free ingress · Attach in seconds · SOC 2 Type II
Storage · US East (Ashburn)
New volumeckpt-70b-sft
2.6 / 4 TB
NVMe block · attached to train-h100-a
datasets-web-2t
9.8 / 12 TB
NVMe block · attached to prep-cpu-1 · resize suggested
scratch-dev
0.3 / 1 TB
NVMe block · available
atlas-archive
Object bucket · 128.4 TB stored
Snapshots: last 38 min ago · daily 02:00 UTC · cross-region → EU Central
Trusted by ML teams shipping in production
Three primitives, zero egress surprises
Block for the hot path, objects for the archive, snapshots for the what-ifs. Each is metered per GB actually stored — nothing provisioned, nothing rounded up.
See full pricingckpt-70b-sft
Resize online4 TB · 3× replicated
2.6 TB used · 120k IOPS
NVMe Block Volumes
$0.08 /GB-mo
A 1 TB volume runs $80/mo — no IOPS tiers, no provisioned throughput.
- Volume size
- 10 GiB – 16 TiB per volume
- Attachment
- Attach to any instance
- Replication
- 3× synchronous
Buckets
Free ingressObject Storage
$8 /TB-mo
100 TB stored is $800/mo — with egress free up to 3× stored volume.
- Interface
- S3 API, unlimited buckets
- Transfer in
- Free ingress
- Durability
- 11×9s designed
snap-ckpt-70b
Daily 02:00Snapshots & Backups
$0.04 /GB-mo
Incremental — 500 GB of retained snapshots is $20/mo; deltas only after the baseline.
- Format
- Incremental, compressed
- Restore
- Cross-region copy
- Schedule
- Hourly to monthly, or API
The network is part of the product
Every instance, volume, and bucket you create lands on one private fabric. Traffic between them is never metered — the only line items are below.
VPC fabric
Private network spanning all four regions. GPUs read from volumes and buckets at line rate, unmetered.
Free
included with every account
Managed load balancer
L4 and L7 with TLS termination, health checks, and per-second connection metrics.
$12/mo
per node
Pooled egress allowance
Shared across every instance on the account, refreshed monthly. Ingress is always free.
20 TB
included per month
Additional IPv4 address
Static and portable — remap to any instance in the region via API in seconds.
$3/mo
per address
East-west traffic inside a VPC — instance to instance, instance to storage — is never metered, in any region.
Engineered for checkpoint-heavy work
Training jobs write terabytes and read them back under deadline. The details below are what make that boring — in the best possible way.
See the whole platformTriple replication
Three copies before we say “done”
Every block write commits synchronously to three isolation zones before the
acknowledgment reaches your instance. A checkpoint that returns from
fsync
is already in three places.
- Synchronous, not eventual — the ack means three durable replicas, not a promise of them.
- Self-healing — a failed drive triggers re-replication in under a second, no operator involved.
- Continuous scrubbing — every stored block is checksummed and verified on a 14-day cycle.
PUT ckpt-step-8000.pt · 4.2 GiB
Writingzone-a
1.1 ms
zone-b
1.3 ms
zone-c
1.2 ms
Re-replication start
< 1 s after failure
Scrub cycle
Every block, 14 days
S3-compatible
The S3 API you already script against
Point aws s3,
boto3, rclone, or s5cmd at a Reviosa endpoint and everything works —
multipart uploads, presigned URLs, lifecycle rules. Migration is a sync
command, and ingress is free, so it costs exactly nothing.
- Drop-in endpoint — change one URL, keep every tool and SDK you have.
- Free ingress, generous egress — reads are free up to 3× your stored volume each month.
- Unlimited buckets and objects — shard datasets however your loader likes them.
$ rclone sync ./shards reviosa:training-shards --transfers 64
✓ 2,048 objects · 214.6 GiB ........ 3 m 42 s
$ aws s3 ls s3://training-shards --endpoint-url \
https://obj.us-east.reviosa.com | tail -1
2026-07-28 09:14:52 104857600 shard-2047.tar
ingress billed: $0.00Ingress
$0.00
always, at any volume
Free egress
3×
stored volume, monthly
Portability
Detach, reattach, resume — in seconds
Volumes are first-class citizens with their own lifecycle. Prep a dataset on a cheap CPU box, detach, attach to an 8× H100 node, and train — the data never moves, only the compute underneath it.
- 2.1 s median detach, 3.4 s attach — one CLI command or API call each.
- Any SKU, same region — the volume doesn't care if it's under an RTX 4090 or an HGX node.
- Independent lifecycle — terminate the instance, keep the volume; billing follows the data, not the box.
ckpt-70b-sft · 4 TB NVMe block
Attached · train-h100-aDetached from dev-rtx4090-a
09:41:02 · volume quiesced, cache flushed
2.1 s
Attached to train-h100-a
09:41:05 · mounted as /dev/vdb
3.4 s
Training resumed from step 8,000
09:41:09 · same checkpoint, 8× the compute
5.5 s total · dev box tonight, HGX node tomorrow, same data
Designed for eleven nines
99.999999999%
Designed annual durability for object storage — the odds of losing an object work out to roughly one in 100 billion object-years.
3×
synchronous replicas on every write
14 days
full-fleet background scrub cycle
4 regions
for cross-region snapshot copies
99.9%
uptime SLA, credited automatically
Maya Lindqvist
Infrastructure Lead @ Atlas Weather Labs
“We keep our full satellite and radar archive in Reviosa object storage. Free ingress meant the migration cost us a weekend of rclone and nothing else — and because our GPU jobs read training shards over the VPC, there's no egress line item on the invoice. The storage bill is one number: terabytes stored.”
1.2 PB
of weather archive stored and served to training jobs over the private fabric — with zero metered egress.
Read customer storiesNow put the data to work
Storage is the quiet half of every training run. The loud half — GPUs, bare metal, and serverless inference — lives one hop away on the same fabric.
GPU Cloud
Nine NVIDIA SKUs, SSH-ready in under 90 seconds. Attach your volumes, read your buckets, and train — billed by the second.
Bare Metal & VMs
Single-tenant EPYC and Xeon, delivered in hours. Databases and data prep next to the volumes they feed from.
Serverless AI
Open-weight models behind an OpenAI-compatible endpoint. Serve the checkpoints you just trained — autoscale from zero, pay per token.
Start training in minutes
Create an account, add a card, and launch your first GPU instance. Per-second billing means you only pay for what you use.
No minimum commitment · Cancel anytime · $10 free credit for new accounts