DGX Spark ×3
NVIDIA's Grace Blackwell desktop system for running, fine-tuning and serving large models.
384 GB across three units: room for ~600B-class models at FP4.
Exclusive bare metal. Your hardware is never shared, and you get full root access.
Hardware, one-time
$19,500
Management & rack
3 units × $99 = $297 per month
How paying works
Place a pre-order. We confirm availability and send an invoice. Nothing is charged online.
Racked and managed by us in Asia or the US. Prices in USD, excluding taxes and import duties.
How it's connected
Three units in a ConnectX-7 triangle: each unit is cabled to the other two, using both of its ports. No switch needed.
Cabled and validated by us before handover. Switched configurations for larger clusters are available on request.
Total unified memory: 384 GB
Specifications
Per unit
- Superchip
- NVIDIA GB10 Grace Blackwell Superchip
- CPU
- 20-core Arm (10× Cortex-X925 + 10× Cortex-A725)
- Memory
- 128 GB LPDDR5x unified · 273 GB/s
- AI compute
- Up to 1 PFLOP FP4 (with sparsity)
- Storage
- 4 TB NVMe, self-encrypting
- Networking
- ConnectX-7 · 2× QSFP, up to 200 Gb/s each
- Operating system
- DGX OS (Ubuntu-based) · ARM64
This configuration
- Units
- 3
- Total memory
- 384 GB
- Total storage
- 12 TB NVMe
- AI compute, FP4
- Up to 3 PFLOP (with sparsity)
- Management & rack
- $297/mo
What you can run
The largest model classes that fit in memory, by configuration. Larger models are split across units over ConnectX-7, so speed depends on the interconnect and software.
Approximate, depends on quantization and context length.
| Configuration | Unified memory | FP8 (~1 byte per parameter) | FP4 (~0.5 byte per parameter) | Source |
|---|---|---|---|---|
| 1× DGX Spark | 128 GB | ~70B comfortably | Up to ~200B | NVIDIA |
| 2× DGX Spark | 256 GB | ~200B | ~405B class (e.g. Llama 3.1 405B) | NVIDIA |
| 3× DGX Spark This configuration | 384 GB | ~300B | ~600B class | Our estimate |
| 4× DGX Spark | 512 GB | ~400B | Up to ~700B (e.g. 671B MoE models) | NVIDIA |
What else it's good for
Fine-tuning and LoRA
NVIDIA rates one unit for fine-tuning models up to 70B with memory-saving methods such as LoRA and QLoRA. Smaller models can be fully fine-tuned.
Multi-agent development
Run several models and agents side by side on hardware you control, with no per-token bill and no data leaving your machine.
Local inference serving
Serve models to your team over a private endpoint with tools such as vLLM, TensorRT-LLM or Ollama.
Management & rack
What $99 per unit per month includes
Every machine is exclusive bare metal: never shared, with full root access. We look after everything below the operating system.
Access
- Full root access
- Private remote access
- Your SSH keys installed for root
- Your choice of OS image, or your own
Network and power
- Shared 1 Gbps port with 2 TB outbound a month
- Power within the agreed envelope, and cooling
Hands and eyes
- Monitoring of temperature, disks and reachability
- Reboots on request
- One approved OS reimage a month
- 15 minutes of remote hands a month
- A four-business-hour response target
Not included unless agreed
- Application uptime guarantees
- Model tuning or application support
- Backups of your data
Volume pricing for larger configurations
Need more than four Sparks, several dedicated CI machines, or a mix? Tell us the configuration and we'll quote it.
DGX Spark ×3
Place a pre-order. We confirm availability and send an invoice. Nothing is charged online.