1-day rental
1 × $19.6
$19.6 due
AmdMetal · AI Development Systems
Rent a DGX Spark by the day
GB10 Grace Blackwell with 128GB of unified memory.
Choose a 1–365 day term on the order page. Use this page to review the specs and usage limits.
One complete system for one rental term
An AmdMetal NVIDIA DGX Spark is assigned to one customer for the rental term. It is suited to local LLM inference, fine-tuning experiments, CUDA development, and AI application prototyping—not a shared inference API.
Workloads run in NVIDIA DGX OS on the system. Code, dependencies, containers, and model weights must be prepared independently. Before choosing a rental term, check the program architecture, runtime environment, and data export plan. Access to the complete system does not guarantee that every model or software package will run; required components must have ARM64-compatible versions.
Device label
GB10 platform, fixed hardware and OS
The system is supplied with the specs below. AmdMetal DGX Spark does not offer hardware or OS options.
- Superchip
- NVIDIA GB10 Grace Blackwell
- CPU
- 20-core Arm: 10 × Cortex-X925 + 10 × Cortex-A725
- Unified memory
- 128GB LPDDR5x, shared by the CPU and GPU; not dedicated VRAM
- Local storage
- 4TB NVMe, not expandable
- Program architecture
- ARM64
- Network and address
- 1 Gbps, unmetered traffic, with 1 static IPv4 address
- Operating system
- NVIDIA DGX OS, based on Ubuntu 24.04, with CUDA and the NVIDIA AI software stack preinstalled
- Models and management interface
- No models preinstalled; no BMC / IPMI
Unified memory is shared by the CPU and GPU, and is also used by the OS and running processes. Whether a model will fit depends on factors such as weight size, precision, context length, and framework overhead—not just the size of the model files.
Rental terms in whole days
The rate is $19.6 per day, for terms of 1–365 days. AmdMetal charges the same daily rate multiplied by the selected number of days, with no term-based discounts.
Amount due = days × daily rate
7-day rental
7 × $19.6
$137.2 due
30-day rental
30 × $19.6
$588.0 due
Amounts above are calculated by multiplying the number of days by the daily rate. The 30-day example is not a monthly plan. There is no contract; confirm the exact term and amount due on the AmdMetal order page.
Check runtime requirements before planning experiments
Best suited to single-system projects with existing code, data, and evaluation goals. AmdMetal provides the system environment; project dependencies and model deployment are handled independently.
Local model inference
Load model weights on the system to test inference workflows, APIs, and application integration. Check ARM64 support in the inference framework, then assess memory requirements.
LoRA fine-tuning experiments
Test adaptation workflows and training scripts. Budget storage for datasets, checkpoints, and experiment outputs, and plan to export them before the term ends.
CUDA development
Develop with the preinstalled CUDA software stack. Check that build tools and third-party dependencies support Arm rather than relying on x86 binaries.
AI application prototyping
Deploy the model, service code, and test data on one system to validate the end-to-end workflow. Save experiment results and application data separately.
No extensions beyond the fixed specs
Bandwidth changes, additional IPv4 addresses, memory upgrades, extra drives, other operating systems, multi-system clusters, and preinstalled model services are not available. Monthly billing is not offered. ARM64 cannot run x86 programs or Windows, so projects that depend on these environments are not suitable.
Allow time for model downloads
AmdMetal DGX Spark comes without preinstalled models; model weights must be downloaded separately. The 1 Gbps network spec does not guarantee that every download source will sustain that speed.
Estimated download times at 1 Gbps
| Weight size | Estimated time |
|---|---|
| 40GB | About 5.5 minutes |
| 100GB | About 14 minutes |
| 140GB | About 19 minutes |
Actual download times depend on the source, network route, protocol overhead, and disk write speed. Estimates exclude extraction, environment setup, weight conversion, and model loading; they do not guarantee that a given model will run on the system.
SSH access; export data before the term ends
After delivery, AmdMetal provides connection credentials through the console. The development environment and experiment data are managed independently.
-
Connect
Get credentials in the console
Sign in to the console, find the connection address and login credentials in the machine details, then connect to the system over SSH. Credentials are not shown on this page.
-
Assistance
Submit a ticket for hard reboot or reinstallation
The system has no BMC / IPMI. To request a hard reboot or reinstall NVIDIA DGX OS, submit a ticket through the console with the machine ID and requested action.
-
At expiry
Export data before it is deleted
Data on the system is deleted when the rental term ends. Export code, configuration, model checkpoints, and experiment results beforehand, and verify that exported files are complete and usable.
Confirm these three usage limits before ordering
The AmdMetal order page requires confirmation of each item. Confirm software architecture, OS selection, and data lifecycle before payment.
ARM64 architecture
x86 programs and Windows cannot run. Project dependencies and container images must be ARM64-compatible.
NVIDIA DGX OS only
Based on Ubuntu 24.04, with CUDA and the NVIDIA AI software stack. No other OS options are available.
Data deleted at expiry
Export data before the rental term ends. Directories on the system are not retained as archives after expiry.
Choose a rental term to match the project
The specs are fixed. Next, choose 1–365 days on the AmdMetal order page, confirm the three usage limits, and review the amount due.