On-premise DGX Spark
Try it in our racks.
Then put it in yours.
The sane way to buy AI hardware is not to guess. Test a rented DGX Spark with public or synthetic inputs first, then decide whether to own one.
$0.79/hour, billed per minute · pay $10, get $20 in credit on your first top-up
Want to own one without keeping it in your office? Buy a GB10 machine and have GPUwerk host it.
Evaluation to ownership
The path from evaluation to your server room.
Test
Rent a Spark by the minute. First top-up: pay $10, get $20 in credit, about 25 hours of compute. Use public or synthetic data for the evaluation; no procurement required.
Validate
GPUwerk goes through your test results with you and writes down what fit and what it costs to run. Where a Spark is the wrong machine, you hear that too.
Own
Want hardware you own? You can choose an NVIDIA, ASUS or Lenovo GB10 system and send an order request, or GPUwerk quotes one on request, with delivery and remote hand-over scoped in writing. Want it to stay in a rack instead? Buy a GB10 machine and have GPUwerk host it.
Managed, on request
Optional, scoped in writing per customer: remote monitoring, security patches and model updates. Support terms, including how hardware issues are handled, are agreed in writing, not sold as a standing SLA.
Managed service
What managed means, concretely.
On request, scoped in writing.
Monitoring and updates
GPUwerk watches node health, applies DGX OS and security patches in agreed windows, and keeps your serving stack current as models and frameworks move.
Support, agreed in writing
Support scope and terms are agreed per customer in writing rather than sold as a standing SLA, and there's no blanket hardware-replacement guarantee: what applies to your setup is written down before you sign.
Team training
Onboarding for your engineers: serving stacks, fine-tuning, capacity planning. The goal is your independence, not your dependence.
Who it's for
Who goes on-premise.
Predictable heavy usage
Running 24/7? Compare a month of rental ($576.70 for 730 hours) with the purchase price; GPUwerk will run that crossover on your numbers before you sign anything.
Board-mandated sovereignty
"The server is in the basement" is an argument that wins meetings, with works councils, boards, and customers' procurement teams alike.
First engagement
First engagement support.
| Engagement step | What we do | Output |
|---|---|---|
| Intake | Review your workload and constraints | Scope statement, metrics, acceptance criteria |
| Workshop | Run public or synthetic sample data through candidate models | Written test results and pilot recommendation |
| Handoff | Plan operating model for your team | Runbook, training plan, support contacts |
Buying questions.
What does it cost?
Hardware at market price plus delivery, quoted in writing, and an optional managed service scoped per customer. Request a written quote with your country and use case; you get a written answer, not a sales call.
Do I have to use the managed service?
No. GPUwerk can deliver the machine, help you set it up remotely and hand it over. The managed service is optional, scoped in writing, and not a lock-in.
Can we start in the cloud and move later?
Yes, that is the recommended path. Everything you build on a rented DGX Spark (images, models and configs) moves to your own GB10 machine as-is, since it is the same platform.
Where do you deliver?
Ask with your country. Delivery with remote-guided setup and hand-over is quoted per country; on-site installation is possible in Czechia and neighbouring countries by arrangement. Every quote states lead time and terms in writing.
Buy the machine
you've already tested.
Start with a rented Spark today. Every hour of testing de-risks the purchase.