On-demand GPU compute

Train and serve models on idle silicon — billed by the second.

Rent H100, H200, A100 and RTX cards from vetted data centers at a fraction of hyperscaler list price. No reservation, no annual commitment, no egress surprises.

Deploy in under 60 seconds Per-second billing Cancel anytime

Hourly rate vs. market

NVIDIA H100
80 GB HBM3
$3.90
$5.20
NVIDIA H200
141 GB HBM3e
$6.85
$9.10
NVIDIA A100
80 GB HBM2e
$1.24
$1.60
RTX 4090
24 GB GDDR6X
$0.54
$0.74
RTX A6000
48 GB GDDR6
$0.71
$0.95

Rates shown are per GPU per hour in USD. The market column reflects published on-demand list pricing at major cloud providers.

The marketplace

Capacity that would otherwise sit idle

Operators list spare cards, you rent them by the hour. Every node is benchmarked before it goes live, and you only pay while the instance runs.

2,400+

GPUs online

Consumer RTX cards through to multi-node H200 clusters with InfiniBand interconnect.

18

Regions

North America, Europe and APAC. Pin workloads to a region when data residency matters.

42s

Median time to first token

From clicking deploy to a running container with your image pulled and CUDA ready.

99.4%

Job completion rate

Interrupted jobs are rescheduled onto equivalent hardware automatically, at no extra cost.

Bring your own image

Any OCI container. PyTorch, JAX and vLLM templates are one click away.

Persistent volumes

Attach NVMe storage that survives restarts and follows you across nodes.

SSH and Jupyter

Direct root SSH, port forwarding, and a hosted notebook on every instance.

API and Terraform

Script provisioning, autoscale on queue depth, tear down when the run finishes.

Managed inference

Open models from $0.02 per minute

Skip the provisioning entirely. Point an OpenAI-compatible endpoint at a dedicated model instance and pay only for the minutes it stays warm.

7B–8B class

Small chat models

from $0.02 / min

Single L4 or RTX 4090. Good for routing, classification and drafting.

30B–34B class

Mid-size reasoning

from $0.03 / min

A100 80 GB with speculative decoding enabled by default.

70B class

Large general models

from $0.11 / min

Dual H100 with tensor parallelism. Sustains long-context sessions.

MoE / 100B+

Frontier open weights

from $0.22 / min

H200 nodes with 141 GB per card for models that will not fit elsewhere.

Prepaid credits start at $10, minimum top-up $5. Unused credits do not expire.

Pricing

What you pay, next to what you would pay elsewhere

Switch between hourly on-demand and a reserved month. Reserved capacity is billed upfront and locks the rate for 30 days.

GPUMemoryOur rate / hrMarket / hr

All prices in USD, excluding applicable taxes. Storage and public IPv4 are billed separately.

For operators

Have cards sitting idle? Put them on the market.

List spare capacity, set your own floor price, and get paid for every second a renter is on your hardware. You keep control of what runs and when the node is available.

  1. Install the agentOne command on Ubuntu 22.04 or later. It reports health, utilization and thermals.
  2. Set your termsFloor price per hour, availability windows, and whether you accept long-running reservations.
  3. Get matchedJobs are routed by price, region and interconnect — not by who bid highest for placement.
  4. Get paidPayouts twice a month. The platform fee is 15% of gross, deducted before payout.
Questions

Before you deploy

How is usage billed?

Per second, from the moment the container starts to the moment it stops. Stopped instances are not billed for compute; attached storage keeps billing until you delete the volume.

Can I get a contiguous multi-GPU node?

Yes. Nodes with 2, 4 and 8 GPUs are listed as single units with NVLink or InfiniBand where the operator provides it. Interconnect bandwidth is shown on every listing.

What happens if a host goes offline mid-run?

The job is requeued onto equivalent hardware and you are not charged for the interrupted portion. Checkpoint to a persistent volume so a restart resumes rather than starts over.

Who can see my data?

Volumes are encrypted at rest and instances are isolated per tenant. ENOSIS does not inspect workload contents. For regulated data, restrict deployment to hosts carrying the compliance attestations you need.

Is there a minimum commitment?

No. On-demand instances can be destroyed at any time. Reserved monthly capacity is the only commitment, and it is optional.

Do you support spot-style pricing?

Interruptible instances run at a discount and can be reclaimed with two minutes of notice. They suit checkpointed training and batch inference, not live serving.

Talk to us

Tell us what you need to run

Card, memory, how long, and which region. We come back with a concrete quote, usually the same business day.

ENOSIS INC
2647 Gateway Road, Suite 240
Carlsbad, CA 92009
[email protected]

Sent. We'll reply to the address you gave us. By sending this you agree to us processing your details to answer the request.