Platform / GPU Cloud

Train faster when it matters most

Blackwell B200 through RTX 4090 — nine NVIDIA SKUs across four regions, SSH-ready in under 90 seconds, billed by the second from $0.49/hr.

Launch a GPU Browse the fleet

Per-second billing · No quotas or capacity requests · SOC 2 Type II · 99.9% uptime SLA

Launch instance

US East (Ashburn)

B200 SXM

Newest silicon

180 GB VRAM · 28 vCPU · 3 TB NVMe

$5.49/hr

H200 SXM

141 GB VRAM · 24 vCPU · 2 TB NVMe

$3.69/hr

H100 SXM

Nine SKUs, zero waitlists

Every row below is live capacity in Reviosa-owned data centers — not resold spot. From RTX 4090 dev boxes at $0.49/hr to full HGX nodes at $2.29 per GPU-hour effective.

Compare with CPU, metal, and storage pricing

GPU	VRAM	vCPUs	RAM	Storage	Network	On-demand	Deploy
B200 SXM Newest silicon 1× NVIDIA B200	180 GB	28 vCPU	360 GB	3 TB NVMe	100 Gbps Ethernet	$5.49/hr $2,599/mo reserved	Deploy
H200 SXM 1× NVIDIA H200	141 GB	24 vCPU	256 GB	2 TB NVMe	100 Gbps Ethernet	$3.69/hr $1,749/mo reserved	Deploy
H100 SXM Most popular 1× NVIDIA H100	80 GB	26 vCPU	224 GB	1.5 TB NVMe	100 Gbps Ethernet	$2.49/hr $1,179/mo reserved	Deploy
H100 SXM ×8 Node Best for training 8× NVIDIA H100	640 GB	208 vCPU	1,843 GB	22 TB NVMe	3.2 Tbps InfiniBand + NVLink	$18.32/hr $8,689/mo reserved	Deploy
H100 PCIe 1× NVIDIA H100	80 GB	26 vCPU	200 GB	1 TB NVMe	25 Gbps Ethernet	$2.09/hr $989/mo reserved	Deploy
A100 80GB SXM 1× NVIDIA A100	80 GB	24 vCPU	192 GB	1 TB NVMe	25 Gbps Ethernet	$1.29/hr $609/mo reserved	Deploy
RTX 5090 1× NVIDIA RTX 5090	32 GB	16 vCPU	64 GB	500 GB NVMe	10 Gbps Ethernet	$0.79/hr $379/mo reserved	Deploy
L40S 1× NVIDIA L40S	48 GB	16 vCPU	96 GB	750 GB NVMe	25 Gbps Ethernet	$0.69/hr $329/mo reserved	Deploy
RTX 4090 Starter 1× NVIDIA RTX 4090	24 GB	12 vCPU	64 GB	400 GB NVMe	10 Gbps Ethernet	$0.49/hr $229/mo reserved	Deploy

B200 SXM

Newest silicon

1× NVIDIA B200

$5.49/hr

$2,599/mo

VRAM: 180 GB
vCPUs: 28
RAM: 360 GB
Storage: 3 TB NVMe
Network: 100 Gbps Ethernet

Deploy

H200 SXM

1× NVIDIA H200

$3.69/hr

$1,749/mo

VRAM: 141 GB
vCPUs: 24
RAM: 256 GB
Storage: 2 TB NVMe
Network: 100 Gbps Ethernet

Deploy

H100 SXM

Built to keep every GPU busy

The details that decide real training throughput — provisioning speed, interconnect, billing granularity, and data locality — engineered as one system.

See the whole platform

Instant provisioning

SSH-ready in under 90 seconds

Warm capacity pools in all four regions mean there is no queue between you and a running node. One command, one coffee-free minute, one SSH prompt.

Prebaked images — PyTorch 2.7, CUDA 12.8, Ubuntu 24.04 — flashed straight to local NVMe.
No quota requests, ever — every account can launch every SKU on day one.
74-second median cold boot — 90 seconds is the ceiling we publish, not the average.

reviosa cli

$ reviosa launch h100-sxm --region us-east --image pytorch-2.7
✓ capacity reserved · us-east-1a .......... 0.6 s
✓ pytorch-2.7-cuda12.8 flashed to NVMe ... 38.1 s
✓ vpc + volume attached ................. 9.4 s

ready — ssh root@203.0.113.87 · total 71 s

Cold boot

74 s

median, all regions

Quota tickets

0

required, forever

Interconnect

3.2 Tbps between every pair of GPUs

Every H100 SXM ×8 Node ships on a rail-optimized InfiniBand fabric with NVLink inside the chassis — so gradient sync never becomes the slowest layer in your training stack.

3.2 Tbps InfiniBand + NVLink east-west on every 8× node — no shared oversubscribed uplinks.
900 GB/s NVLink intra-node — tensor parallelism without the bandwidth penalty.
NCCL-tuned images by default — 98% scaling efficiency measured at 64 GPUs.

hgx-a04 · 8× NVIDIA H100

Training

GPU0

99%

GPU1

98%

GPU2

99%

GPU3

97%

GPU4

99%

GPU5

98%

GPU6

96%

GPU7

99%

NVLink · 900 GB/s

Fabric

3.2 Tbps InfiniBand

All-reduce bus BW

480 GB/s measured

Per-second billing

Billed by the second, not the hour

Stop an instance mid-epoch and the meter stops the same second. On a H100 SXM that means experiments cost exactly what they run — $0.0007 per second, nothing rounded up.

No minimums, no rounding — a 47-second smoke test bills 47 seconds.
Live spend per instance in the console, CLI, and API — never a surprise invoice.
Reserved rates when the workload is steady — switch modes without redeploying.

Live meter

Running

train-h100-a · H100 SXM · US East (Ashburn)

$6.69this session

2 h 41 m at $2.49/hr · $0.0007 per second

metered stopped — $0.00

Stopped instances accrue $0.00. Volumes and IPs persist until you say otherwise.

Persistent storage

Volumes that outlive the instance

Checkpoints, datasets, and environments live on persistent NVMe volumes. Terminate a dev box tonight, attach the same volume to an 8× node tomorrow, and resume from the exact step you left.

Triple-replicated NVMe block volumes, resizable online without a reboot.
Attach to any SKU — detach from an RTX 4090, reattach to a H100 SXM ×8 Node.
Hourly snapshots with cross-region restore to any of our four regions.

Volumes

New volume

ckpt-70b-sft

4 TB NVMe block · US East (Ashburn)

Attached · train-h100-a

datasets-web-2t

12 TB NVMe block · US East (Ashburn)

Attached · prep-cpu-1

ckpt-70b-dpo

2 TB NVMe block · last snapshot 38 min ago

Available

Triple-replicated · resize online · hourly snapshots

Scale path

Start on one card. Scale to eight.

Same image, same volumes, same API. The only thing that changes is the number after the ×.

vs

Single GPU

H100 SXM

H100 SXM ×8 Node

Best for training

Distributed training on NVLink and InfiniBand — pretraining runs, 70B+ fine-tunes, and anything with a deadline.

$18.32 /hr = $2.29 per GPU-hour effective

GPU: 8× NVIDIA H100
VRAM: 640 GB HBM3 (8× 80 GB)
vCPUs: 208 vCPU
System RAM: 1,843 GB
Local NVMe: 22 TB NVMe
Network: 3.2 Tbps InfiniBand + NVLink

Deploy H100 SXM ×8 Node

Out of VRAM? When model plus optimizer state passes 80 GB, shard across the node instead of your weekend.

Batch stalling? InfiniBand keeps gradient sync off the critical path — 98% scaling efficiency at 64 GPUs.

Deadline math: eight GPUs for one day beats one GPU for eight — same spend, per-second billed, results a week sooner.

AO

Dr. Amara Okafor

Head of ML @ Loomline AI

“We moved our fine-tuning fleet to Reviosa in a weekend. The 8× H100 nodes hold 97% scaling efficiency on our 70B runs, and per-second billing changed how we schedule experiments — we stop nodes between sweeps and simply stop paying.”

3.4×

faster iteration on 70B fine-tunes after moving from single cards to H100 SXM ×8 Node clusters.

Read customer stories

Keep building

The rest of the platform

GPUs rarely work alone. Metal for the databases, storage for the checkpoints, serverless endpoints for the finished model — all on one private network.

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.

Create free account View pricing

No minimum commitment · Cancel anytime · $10 free credit for new accounts

Train faster when it matters most

Nine SKUs, zero waitlists

Built to keep every GPU busy

SSH-ready in under 90 seconds

3.2 Tbps between every pair of GPUs

Billed by the second, not the hour

Volumes that outlive the instance

Start on one card. Scale to eight.

H100 SXM

H100 SXM ×8 Node

The rest of the platform

Bare Metal & VMs

Storage & Networking

Serverless AI

Start training in minutes