Compare monthly totals for AMD servers by memory tier, storage and port speed. DGX Spark pricing is based on fixed hardware and rental days. Each product line is priced separately.
Data center: Tokyo, Japan.
Both OnuCloud tiers are dedicated servers for a single customer. The CPU and included NVMe drives are not shared. Choose a tier based on memory requirements, then budget for backup capacity and transfer speed. Memory cannot be upgraded on the same server after deployment; changing tiers requires a new server.
A good fit for file storage, backup targets or a single database service. With a tighter memory budget, consider allocating more to separate backup drives and recovery bandwidth.
A good fit for multi-VM hosts, larger database caches or plans to expand ZFS ARC. Only memory capacity changes; the CPU, included drives and base network specifications remain the same.
| Server | Memory | Included storage | Network and IP | Monthly price |
|---|---|---|---|---|
| AMD EPYC 7543P · 128GB | 128GB DDR4-3200 ECC | 2 × 1TB NVMe, direct-attached | 250 Mbps unmetered, 1 IPv4 address | $472/month |
| AMD EPYC 7543P · 256GB | 256GB DDR4-3200 ECC | 2 × 1TB NVMe, direct-attached | 250 Mbps unmetered, 1 IPv4 address | $624/month |
The two included 1TB NVMe drives are detected independently; they are not delivered as a RAID-1 array. Add up to two 2TB NVMe SSDs at $85/month each. These are also direct-attached and not part of a hardware RAID array. Total raw capacity is not the same as usable capacity after redundancy is configured.
Keep the 2 × 1TB NVMe drives for manageable hot data and standalone storage services. Mirroring the two drives provides about 1TB of nominal usable capacity, before filesystem overhead.
Trade-off: No added monthly drive cost, but less room to divide between the system, workloads and backups. Mirroring in the same server is not a substitute for off-site backups.
Additional drive cost: $0/monthAdd 1 × 2TB NVMe alongside the 2 × 1TB NVMe. Use the new drive as a separate backup target to keep workload and backup directories apart.
Trade-off: The added drive has no mirror redundancy. It suits deployments with another backup copy, but not as the sole location for data.
Additional drive cost: $85/monthWith 2 × 1TB and 2 × 2TB drives, plan separate mirrored sets for hot data and backups. Mirroring each pair of same-capacity drives provides about 3TB of nominal usable capacity.
Trade-off: Drives have clearer roles, but all remain in one server. Plan separately for fault isolation and off-site copies.
Additional drive cost: $170/monthZFS, software RAID, filesystems, snapshot retention and backup policies are configured by the customer. OnuCloud provides the hardware and network, not data redundancy setup.
All three tiers—250 Mbps, 1 Gbps and 2 Gbps—are unmetered. Pricing varies by port speed, with no additional charge based on monthly transfer volume in TB.
For incremental backups, estimate daily change volume first. For full-server recovery, consider the acceptable recovery window. Actual throughput also depends on the remote endpoint, network path, protocol and disk I/O; port speed is not a guaranteed end-to-end benchmark.
A good fit for incremental syncs with low change volume that can run over time. Keep the base tier if recovery can use a longer transfer window.
Trade-off: Lower port speed can limit recovery rates when transferring large files continuously.
UnmeteredA good fit for transferring large backup sets, full-server recovery and bulk data transfers. Confirm that the remote endpoint has suitable network capacity too.
Trade-off: Adds a fixed monthly cost, but does not remove remote rate limits or cross-network bottlenecks.
UnmeteredA good fit for aggregating backups from multiple sources or allowing more bandwidth for off-site replication. Estimate capacity based on concurrency and backup windows.
Trade-off: Only the port tier increases; application concurrency and storage throughput still need to be tested.
UnmeteredEach AMD dedicated server includes 1 dedicated IPv4 address. Additional addresses cost $2/month each, up to 255 more, for a maximum of 256 total. Addresses are charged individually and are not bundled with bandwidth options.
Add addresses as needed when VMs require separate public services, services need distinct allowlisted IPs, or fixed egress addresses need separate planning. Choose the number based on the services and how addresses will be assigned in the operating system.
Use the first IPv4 address for SSH and RDP access. Configure any additional addresses within the operating system for VMs or services. Multiple addresses do not automatically configure bridging, routing or service bindings.
AMD servers offer 14 free operating systems, plus Windows Server with a monthly license fee. Connect to Linux, FreeBSD and Proxmox VE via SSH, or Windows Server via RDP. Find login credentials in the control panel after deployment.
Standard
+$17/monthDatacenter
+$27/monthVersions: 2019 / 2022 / 2025. Languages: English, Traditional Chinese and Simplified Chinese.
The license fee is added to the monthly total, then multiplied by the selected billing term. Check software requirements when choosing a version; the operating system name does not mean business applications are installed.
Monthly, quarterly, semiannual and annual billing cover 1, 3, 6 and 12 months, respectively, with no discounts. The table shows base server costs only. Add drives, bandwidth, IPv4 addresses or Windows licenses to the monthly total before multiplying by the selected number of months.
| Server | Monthly × 1 | Quarterly × 3 | Semiannual × 6 | Annual × 12 |
|---|---|---|---|---|
| AMD EPYC 7543P · 128GB | $472 | $1416 | $2832 | $5664 |
| AMD EPYC 7543P · 256GB | $624 | $1872 | $3744 | $7488 |
Monthly server price + drive charges + bandwidth upgrade + additional IPv4 addresses × $2 + Windows license fee.
With a free OS, included bandwidth and no extra drives or IP addresses, the monthly total is the server price. The configuration page lists each charge and the amount due. Current availability is shown in real time in the control panel.
No setup fees or contracts. All charges are in USD; available payment methods are listed on the configuration page.
OnuCloud NVIDIA DGX Spark is rented as a complete server, with one device per customer. It suits model evaluation, CUDA development and AI application prototyping. Pricing is based on rental days, not AMD memory tiers or billing-term options.
Check the server specifications first, then confirm that the project runs in an ARM64 and NVIDIA DGX OS environment. Hardware is fixed: additional charges cannot be used to change the OS, expand storage or upgrade network speed.
Fixed hardware and OS, billed by rental days. The daily rate stays the same for longer rentals. No setup fees or contracts.
The CPU and GPU share 128GB of LPDDR5x memory, which is also used by the OS, runtime environment and models. When assessing whether a model will fit, account for weight precision, context, cache and application processes. Not all of the reported capacity is available for model weights.
NVIDIA DGX OS is based on Ubuntu 24.04 and includes CUDA and the NVIDIA AI software stack; models are not preinstalled. Model weights and application dependencies must be prepared separately. If existing software relies on x86 binaries or Windows, choose a compatible technical approach before renting.
Unmetered 1 Gbps is suitable for downloading model weights and exporting results, but actual speeds depend on the model host, network path and application. Budget for preparation, execution and export—not just inference or training time.
| Component | Included specification | Limitations |
|---|---|---|
| Superchip and CPU | NVIDIA GB10 Grace Blackwell; 20-core Arm, 10 × Cortex-X925 + 10 × Cortex-A725 | ARM64 architecture; x86 programs and Windows are unsupported |
| Memory | 128GB LPDDR5x unified memory, shared by CPU and GPU | Memory cannot be expanded; models and the runtime environment share this capacity |
| Storage | 4TB NVMe | Storage cannot be expanded; allocate space for models and results |
| Network | 1 Gbps unmetered, 1 fixed IPv4 address | Bandwidth cannot be changed; additional IP addresses are unavailable |
| Operating system | NVIDIA DGX OS, based on Ubuntu 24.04, with CUDA stack | Windows, Proxmox VE and other distributions are unsupported; models are not preinstalled |
| Access and support | SSH; find credentials in the control panel after deployment | No BMC / IPMI; hard resets and reinstallation are handled through support tickets |
Rent for 1–365 days. Amount due = rental days × daily rate of $15. The examples below use the same fixed server specifications, with no discounts or hardware changes, and are not converted to AMD monthly pricing.
1 day × $15 = $15
A good fit when inputs are ready and the goal is clear. Include preparation time in the rental period if large model weights need downloading or dependencies need migrating.
Fixed server specifications7 days × $15 = $105
A good fit for experiments that include environment setup, parameter tuning, repeat measurements and results review. Plan evaluation data and output formats in advance so setup does not consume the rental period.
Fixed server specifications30 days × $15 = $450
A good fit for a development cycle with multiple validation rounds. This example covers 30 consecutive days; it is not a monthly plan, and the daily rate does not change with rental length.
Fixed server specificationsThe DGX checkout page asks for confirmation of each requirement. They determine whether applications can run and what must be completed before the rental ends; they are not optional add-on services.
x86 programs and Windows are unsupported. Check ARM64 compatibility for dependencies that include native extensions, specialized binaries or container images; containers do not automatically convert between architectures.
The OS is fixed to NVIDIA DGX OS and cannot be changed to another distribution. Plan application deployment for the included CUDA environment rather than treating an OS change as a fallback after the project starts.
Transfer models, evaluation records and results before the rental ends. Exporting data takes time too; do not schedule every rental day for compute and leave backups until the end.
Short-term model comparisons, LoRA / QLoRA fine-tuning experiments, CUDA debugging and prototype validation can be budgeted by project stage. Check model size and software dependencies separately; not every model or training workload will fit in unified memory.
Deployments that need x86 programs, Windows, multi-node training, additional storage or memory, or more IP addresses should not use this fixed server as an expandable configuration. Check specifications before checkout; architecture limits cannot be solved by adding rental days.
The OnuCloud DGX checkout page brings together rental days, the three usage requirements and the amount due. No setup fees or contracts. Billed in USD; available payment methods are listed at checkout.