GPU Instances

The card you asked for, booted in minutes.

H100s and A100s with bare-metal passthrough and NVMe scratch, billed by the second. No reservations, no sales call, no minimum.

See the fleet Price a configuration

$ hp gpu create --card h100 --region bom
✓ booted in 2m14s · nvidia-smi: H100 80GB HBM3 · CUDA 12.4

Mumbai · Singapore · Frankfurt · Virginia

Four cards. Live availability.

What's listed as available boots now. Everything else tells you the queue length instead of pretending.

H100 80GB

80GB HBM3

4 ready

$2.49/hr

≈ $1,818/mo at full use

Deploy

A100 80GB

80GB HBM2e

2 ready

$1.69/hr

≈ $1,234/mo at full use

Deploy

L40S

48GB GDDR6

3 in queue

$1.12/hr

≈ $818/mo at full use

Join queue

RTX 4090

24GB GDDR6X

6 ready

$0.64/hr

≈ $467/mo at full use

Deploy

Same H100 on a major cloud: $4.10/hr. Here: $2.49/hr. Same silicon, same driver.

Pick by workload

Training & fine-tuning

80 GB HBM3 and full-card throughput.

→ H100 80GB · $2.49/hr

Inference serving

Cost-per-token that scales down at night.

→ A100 80GB · $1.69/hr

Rendering & simulation

48 GB for scenes that don't fit elsewhere.

→ L40S · $1.12/hr

Prototyping

The cheapest way to find out if the idea works.

→ RTX 4090 · $0.64/hr

Bare-metal passthrough

The whole card, PCIe-attached to your VM. nvidia-smi shows the same numbers the datasheet does.

Per-second billing

Stop the instance and the meter stops with it. A 40-minute fine-tune costs 40 minutes.

NVMe scratch on every size

Local NVMe for datasets and checkpoints, so the GPU is never waiting on a network volume.

Images that just work

Ubuntu with the CUDA toolkit, drivers and container runtime pre-installed. Bring your own Dockerfile or don't.

Fast egress, flat price

Pull models and push artifacts without a surprise line item at the end of the month.

Snapshots & resize

Snapshot a finished environment and bring it back on a bigger card when the job outgrows it.

Questions

Is this a shared or a dedicated GPU?

Dedicated. Each instance gets a whole physical card via PCIe passthrough — no time-slicing, no MIG partitions unless you create them yourself.

How does per-second billing work on an hourly price?

The hourly figure is the rate; usage is metered to the second and capped at the monthly equivalent. A stopped instance only pays for the disk it keeps.

What does “in queue” mean?

Every card in that row is currently allocated. Queue a request and you'll be billed only from the moment the instance actually boots.

Can I run multiple GPUs in one instance?

Multi-GPU shapes are available on request for H100 and A100. Talk to us with the node count and we'll confirm availability in the region you need.

Boot a GPU. Pay for the minutes you use.

From $0.64/hr · Per-second billing · Mumbai · Singapore · Frankfurt · Virginia

See the fleet Need CPU-only compute?