RTX PRO 6000 · Blackwell Server Edition
Most open weights fit on one 96 GB GPU.
Rent NVIDIA RTX PRO 6000 at $2.25 / GPU-hr, billed per second — 1–8 GPUs per node behind a private, OpenAI-compatible /v1 endpoint with one API key. Terminate anytime.
# one GPU · 96 GB GDDR7 · billed per second
$ curl https://amber-crest-8841.skyforgecompute.com/v1/chat/completions \
-H "Authorization: Bearer $SKYFORGE_API_KEY" \
-d '{"model":"qwen3-6-35b-a3b","messages":[…]}'
{"choices":[{"message":{"content":"Served from 96 GB of GDDR7."}}]}01Hardware
The spec sheet, not the pitch.
The numbers that decide whether your model fits on one GPU — verified hardware, nothing else.
| Spec | Value | Notes |
|---|---|---|
| Memory | 96 GB GDDR7 | per GPU — most open-weights models fit without sharding |
| Memory bandwidth | 1,597 GB/s | per GPU, feeding weights and KV cache |
| GPUs per node | 1–8 | pick a count at launch; one instance, one endpoint |
| Aggregate memory | up to 768 GB | eight GPUs in a single node for larger serving |
$2.25 / GPU-hr · billed per second from boot to terminate
Full pricing →02Which models fit
Tier-S: one template, one GPU.
Five curated templates from the catalog run on a single RTX PRO 6000. Each link opens the console with the template preselected.
| Model | Architecture | Served | Max context | Min VRAM | |
|---|---|---|---|---|---|
| GLM-4 9B | 9B dense | bf16 | 128K | 24 GB | Launch → |
| Qwen3.6 27B | 27B dense | bf16 | 256K | 80 GB | Launch → |
| Gemma 3 27Bgated | 27B dense | bf16 | 128K | 80 GB | Launch → |
| DeepSeek-R1-Distill 32B | 32B dense | bf16 | 128K | 80 GB | Launch → |
| Qwen3.6 35B-A3B | 35B MoE · 3B act | bf16 | 256K | 96 GB | Launch → |
Every row fits a single RTX PRO 6000 (96 GB) — served with vLLM behind your private, OpenAI-compatible /v1.
Full catalog →03Workloads
What teams run on it.
Inference serving
Serve open-weights chat and completion models behind a stable /v1 endpoint.
Fine-tuning & LoRA
Adapt open models to your data with LoRA or full fine-tunes on local NVMe.
Embeddings & RAG
Generate embeddings and back retrieval pipelines with the same endpoint.
Agents & tool use
Run agent harnesses with high request fan-out against your own models.
Batch & offline jobs
Spin up for a batch run, pay per second, and tear down when it finishes.
Notebooks & dev
Prototype in JupyterLab or VS Code Server with the full CUDA toolkit.
04Pricing
One rate. Per second. No gates.
On-demand is self-serve. Committed capacity is quoted by our enterprise team.
On-demand · self-serve
$2.25/ GPU-hr
Managed RTX PRO 6000 GPU-hours behind ready-to-run templates. 1–8 GPUs, billed per second, no commitment.
- Per-second billing
- OpenAI-compatible /v1 API
- One API key across instances
- Ready-to-run launch templates, or your own image
- 1–8 GPUs, no commitment, no tier gates
- Terminate anytime
05FAQ
Questions, answered.
Billing, compatibility, and scaling — the short version.
06Beyond on-demand
Steady workload? We'll quote it.
Committed-use pricing
Reserve capacity at committed-use rates when your workload is steady enough to plan around.
Private networking
Connect your fleet with private networking for production traffic.
Named support
Work with a named contact for onboarding, scaling, and incident response.