Hardware Pricing Docs FAQ

The DGX Spark cloud
for builders.

Deploy dedicated NVIDIA DGX Spark nodes in under a minute. GB10 Grace Blackwell, 128 GB unified memory, 1 PFLOP of FP4, billed by the hour, no commitments.

See pricing

EU-Central · checking availability · $50 free on your first $100 top-up

Dedicated bare metalnever shared or virtualized
SSH-key access onlypasswords disabled fleet-wide
EU data residencyhardware in EU-Central
Status published openlystatus.gpuwerk.com

The hardware

Every instance is one whole machine. Nothing is shared, virtualized, or oversubscribed.

NVIDIA DGX Spark · per node
Superchip
GB10 Grace Blackwell
AI performance
1 PFLOP (FP4)
Memory
128 GB unified LPDDR5x
CPU
20-core Arm
Storage
1 TB NVMe
Network
ConnectX-7 · 200 Gb/s
OS
DGX OS 7 · Ubuntu 24.04
Deploy time
Under a minute

What fits in 128 GB

Unified memory is the whole point, a single node holds models that would otherwise need several discrete GPUs.

Pricing

Per-minute billing, $0 setup fee, no egress charges. Stopped instances don't bill. Top up $100, get $50 free.

Recommended

On-Demand

$2.90/hour
Nodes
1× DGX Spark
Memory
128 GB
Storage
1 TB NVMe
Setup fee
$0
Commitment
None

A month, always on

$2,117/month
What it is
$2.90/h × 730 hours
Nodes
1× DGX Spark
Up front
Nothing
Setup fee
$0
Commitment
Stop any time

Docs: deploy and connect

Every instance is a clean box with a hostname and key-based SSH. Root on real hardware, bring your own stack. Managed model serving is coming soon.

# deploy from the API
curl -X POST https://api.gpuwerk.com/v1/instances \
  -H "Authorization: Bearer $GPUWERK_TOKEN" \
  -d '{"image":"dgx-os","ssh_public_key":"ssh-ed25519 AAAA..."}'

# your shell on the box
ssh -p 22001 root@ssh.gpuwerk.com

# the whole GB10 is yours
nvidia-smi && python -c "import torch; print(torch.cuda.get_device_name(0))"

FAQ

The short version of how billing and instances work.

How does billing work?

Per-minute, against a prepaid credit balance. You're only billed while an instance is running, stopped instances cost nothing. Your first $100 top-up is credited with an extra $50, and if your balance runs out your instances are stopped automatically, never silently overdrafted.

What happens to my data when I stop vs. terminate?

Stopping keeps everything on the node's NVMe, start it again and your files, models, and environment are exactly as you left them. Terminating releases the hardware and permanently erases local data, which is why it requires typing the instance name to confirm.

Is there an egress or ingress fee?

No. The hourly rate is the whole price, bandwidth in and out over the 200 Gb/s ConnectX-7 link is included.

How isolated is my instance?

Each instance is one whole physical DGX Spark, no virtualization, no shared GPUs, no noisy neighbors. Access is SSH-key only; password login is disabled across the fleet. Hardware is located in EU-Central.

Do you offer managed model serving?

Not yet, today every instance is a clean box you drive over SSH: install vLLM, SGLang, or whatever stack you prefer, and it all fits in unified memory. Preloaded model images that serve on boot are on the roadmap.

Start building on a Spark today.

$50 free with your first $100 top-up. Deployed in under a minute. $0 setup fee, no commitments, no egress charges.