NVIDIA DGX Spark · GB10 Blackwell
128 GB of unified memory, from $0.59 an hour.
Rent a DGX Spark to prototype and serve open-weights models — full root SSH and Docker, CUDA 12.8, billed per second. Cluster up to four units for 512 GB unified, then step up to the RTX PRO 6000 fleet when you scale.
# full root · Docker · CUDA 12.8
$ nvidia-smi --query-gpu=memory.total --format=csv,noheader
131072 MiB
$ docker run --gpus all -p 8000:8000 vllm/vllm-openai
INFO serving on :8000 (OpenAI-compatible)
$ curl https://calm-node-2214.skyforgecompute.com/v1/models \
-H "Authorization: Bearer $SKYFORGE_API_KEY"01Hardware
One box, sheet of record.
Unified memory big enough for serious models — run up to roughly 200B parameters without sharding, with full root.
| Spec | Value | Notes |
|---|---|---|
| Compute | GB10 Blackwell | SM 10.0 · NVFP4 · Transformer Engine 2.0 |
| Unified memory | 128 GB @ 273 GB/s | coherent CPU+GPU memory on one box |
| Storage | 4 TB NVMe | local scratch for datasets and checkpoints |
| Access | root SSH + Docker | CUDA 12.8 — your stack runs as-is |
| Model headroom | ≤ ~200B params | on a single box, without sharding |
| Clustering | up to 4 units | 512 GB unified across the cluster |
From $0.59 / hr per instance · billed per second
Shapes & pricing →02Shapes & pricing
One Spark, or a 512 GB cluster.
Priced per instance, per hour, billed per second. No commitment, no tier gates — launch from the console.
| Shape | Unified memory | Price | Notes | |
|---|---|---|---|---|
| 1× Spark | 128 GB unified | $0.59/ hr | one GB10 box — the cheapest way onto the fleet | Launch → |
| 2× Spark | 256 GB unified | $1.10/ hr | two linked units — a step before the full cluster | Launch → |
| 4× Spark cluster | 512 GB unified | $2.10/ hr | four linked units — bigger models, parallel jobs | Launch → |
Priced per instance, per hour — billed per second from boot to terminate. Root SSH + Docker, CUDA 12.8 on every shape.
Own a DG1 workstation →base_url = https://<your-slug>.skyforgecompute.com/v1 · api_key = one SkyForge key, every instance
03How it compares
The cheapest way onto Blackwell.
A Spark is a dev box, not a training cluster — for bigger serving or training, step up the fleet.
| Option | Memory | Price | Best for |
|---|---|---|---|
| DGX Spark | 128 GB unified | $0.59 / hr | dev, prototyping, ≤ ~200B models |
| RTX PRO 6000 | 96 GB / GPU | $2.25 / GPU-hr | open-weights inference |
| B200 / B300 | 192 GB-class HBM | built to order | large-scale training |
A Spark is where you prototype — step up when throughput matters.
RTX PRO 6000 →04What runs on it
Ready-to-run, from the first minute.
Open frontier models
GLM, DeepSeek, Qwen 3, MiniMax, and Llama — served behind your private, OpenAI-compatible endpoint.
Inference server
vLLM, SGLang, TGI, or Ollama behind an OpenAI-compatible endpoint.
Fine-tuning
Axolotl, Unsloth, or LLaMA-Factory with datasets staged on local NVMe.
Agent harness
Tooling and orchestration for agentic workloads and high request fan-out.
Dev environment
JupyterLab or VS Code Server with the CUDA toolkit, ready to code.
Embeddings & RAG
Generate embeddings and back retrieval pipelines from the same endpoint.
05FAQ
Questions, answered.
Billing, model headroom, and when to step up — the short version.
06 — When you outgrow it
Start on a Spark. Scale on the fleet.
Prototype at $0.59 / hr, then move the same endpoint workflow to RTX PRO 6000 at $2.25 / GPU-hr. Want to own one? A DG1 workstation on a one-year term is an enterprise order.