A dedicated system for one customer, hosted in Tokyo, Japan.
Built for local LLM inference, fine-tuning experiments, CUDA development and AI application prototyping. Check ARM64 software compatibility first, then allow time in the rental period for downloading model weights, running experiments and exporting results.
Delivered with a fixed configuration and no preinstalled models. Full specifications and usage limitations are listed below.
The CPU, unified memory, storage and networking are provided as one fixed DGX Spark configuration. It suits projects that can be reproduced within defined hardware limits, not those that rely on adding resources after rental.
| Component | Specifications and notes |
|---|---|
| Superchip | NVIDIA GB10 Grace Blackwell |
| CPU | 20-core Arm: 10 × Cortex-X925 + 10 × Cortex-A725 |
| Unified memory | 128GB LPDDR5x unified memory, shared by the CPU and GPU; it is not two separate pools of memory that can be added together. |
| Local storage | 4TB NVMe, fixed capacity; additional drives cannot be installed. |
| Network | 1 Gbps, unmetered traffic; bandwidth cannot be changed. |
| Public IP | 1 fixed IPv4 address; additional IPs are not available. |
| Operating system | NVIDIA DGX OS, based on Ubuntu 24.04, with CUDA and the NVIDIA AI software stack preinstalled. Models are not preinstalled. |
| Architecture | ARM64. Check that project dependencies, container images and binaries are compatible with this architecture. |
Only NVIDIA DGX OS is provided. It cannot be replaced with Windows, Proxmox VE or another distribution. The fixed configuration has no options for extra memory, storage, bandwidth or IPv4 addresses.
The rate is $15 per day, with rental periods from 1 to 365 days. Amount due = number of days × daily rate. There are no discounts or contracts. The examples below illustrate rental costs and do not promise how long a performance test will take.
$15Amount due
1 × $15 = $15
Best for focused testing when dependencies are confirmed and test scripts are ready. Allow time for downloading model weights and exporting results, too.
$105Amount due
7 × $15 = $105
Useful for comparing models across runs, tuning inference parameters or validating a fine-tuning approach. Allow extra time for downloads, troubleshooting dependencies and retesting.
$450Amount due
30 × $15 = $450
Suitable for prototype iterations with a clear end date. Pricing remains based on the number of days; the daily rate does not change with a longer rental.
Allow time for preparation, downloads, experiments and exports. Finishing a download does not mean the experiment is complete, and data still needs to be transferred before the rental ends.
According to NVIDIA, 128GB of unified memory can support quantized inference with around 200B parameters and LoRA / QLoRA fine-tuning with around 70B parameters. These workloads have different resource requirements, so parameter counts alone are not directly comparable.
Quantization makes model weights more compact, but runtime space is also needed for the context cache, framework and system. A model-size estimate does not guarantee support for any context length or concurrency level.
Before deployment, note the quantization format, context length, target concurrency and dependency versions. Then test memory use and response behavior with the actual workload.
This refers to parameter-efficient fine-tuning, not a promise of full-parameter training. Sequence length, batch size, optimizer and framework implementation all affect actual memory requirements.
Start with a small sample to validate the training pipeline, then scale up the experiment. Multi-node training and workloads that require a larger resource pool are outside the scope of this dedicated system.
DGX Spark is suited to short-term AI projects with known hardware requirements. Choose based on the software stack and workload constraints, rather than assuming unified memory alone makes it a substitute for other servers.
Compare quantized models, validate an inference API, test a retrieval-augmented workflow or run LoRA / QLoRA experiments. A rental period is easier to plan when the input data, evaluation criteria and destination for exported results are already in place.
CUDA development also requires checking ARM64 dependencies. A dedicated system provides exclusive access to the resources; it does not replace application compatibility testing.
x86 programs and Windows cannot run on this system. Multi-node clusters are not provided, and storage or memory cannot be expanded. Choose a different environment if the workload requires any of these capabilities.
Projects planned around long-term, month-to-month operation are also not suited to this rental model, which is charged by the day.
No models are preinstalled. Model weights and datasets must be downloaded separately. The 1 Gbps link rate is only a theoretical figure: actual download times also depend on the source, network path, concurrency and extraction process.
Estimates are based on file size and link rate, rounded to the nearest approximation. They exclude protocol overhead, throttling, verification and file processing.
| File size | Approx. theoretical time |
|---|---|
| 40GB | 5.5 minutes |
| 100GB | 14 minutes |
| 140GB | 19 minutes |
The 4TB NVMe drive holds the system, model weights, datasets, caches and experiment outputs; not all of its capacity is available for model storage. Include multiple model versions and intermediate checkpoints in capacity planning.
Data is deleted when the DGX Spark rental ends. Back up model weights, training checkpoints, evaluation records, code and final results before then. The rented system must not be the only copy.
Choose a backup destination before starting work, and save recoverable checkpoints throughout the experiment. After exporting, verify file integrity and keep a record of run parameters and the environment to make the experiment reproducible, not just the model files.
Transferring large files takes time. Leave enough time between the final run and the end of the rental for export and verification.
After delivery, find the SSH credentials in the console and connect to NVIDIA DGX OS using the provided IPv4 address. The system has no BMC / IPMI. Hard resets and OS reinstalls are handled through a console support ticket.
Include the task stage, symptoms and troubleshooting steps already taken in the ticket. Do not include passwords, keys or other credentials in public content or email subject lines.
All three requirements are listed here in full. Confirm each one on the checkout page before choosing the rental period.
x86 programs and Windows cannot run on this system. Check that project binaries, Python extensions and container images support ARM64. If dependencies cannot be migrated, do not order based on memory capacity alone.
The OS is fixed to NVIDIA DGX OS, based on Ubuntu 24.04, with CUDA and the NVIDIA AI software stack. No other distribution or Windows is available. Models must be prepared and downloaded separately.
Back up experiment results and models independently. Decide where data will be transferred and how it will be verified before choosing the rental period. After the rental ends, the system cannot be relied on to retain weights, checkpoints or project files.
OnuCloud DGX Spark is $15 per day, with rental periods from 1 to 365 days. Check ARM64 compatibility, NVIDIA DGX OS and your data export plan, then choose the number of days on the separate checkout page.