GPUStack AI features and pricing (October 2026)

GPUStack is an open-source GPU cluster manager for serving AI models. IT and platform teams use it to run LLMs and other models on their own GPU servers, Kubernetes clusters or cloud GPUs and to expose them through standard APIs, with an Enterprise Edition for governance.

The platform has two parts: Model as a Service for deploying, routing and monitoring models, and GPU as a Service for GPU instances with SSH or Jupyter access and S3 or NFS storage. It also offers a playground, a model catalog, API keys with quotas and Grafana and Prometheus dashboards. Models come from Hugging Face, ModelScope or local files, and the supported accelerators include NVIDIA, AMD, Ascend and several other vendors. Installation runs via Docker or Helm, including air-gapped setups. The operator is GPUStack Co., Ltd., and a team needs an engineer who runs Linux, Docker or Kubernetes.

• Serves open models from your own GPUsModels from Hugging Face or ModelScope are deployed as endpoints with an OpenAI- or Anthropic-compatible API. Internal apps, chat front ends and coding tools can then call models that run on company hardware, so prompts do not have to go to an outside provider.
• Picks and tunes the inference engineThe scheduler maps each model to a compatible engine version such as vLLM, SGLang or llama.cpp and checks that it fits the GPUs. Throughput, latency and standard modes preset engine parameters, so platform teams spend less time on manual tuning.
• Splits large models across nodesLarge models can be spread over several GPUs and nodes with tensor and pipeline parallelism, including Ray-based vLLM clusters. A team can then run models that no single server can hold, using the GPUs it already owns.
• Routes and meters model trafficModel routes balance load, retry and fall back between deployments, and can include public providers such as OpenAI or Claude. Usage is metered per user, key and model, with token quotas, which supports internal cost allocation.

• Supports many accelerator types in one platform, among them NVIDIA, AMD, Ascend, Hygon, MetaX, Moore Threads and Cambricon.
• Open source under the Apache License 2.0, with a separate Enterprise Edition that adds high availability, multi-tenancy, audit logs and white-label branding.
• An air-gapped installation is documented, and the privacy policy says users can disable telemetry before or during operation.
• Also provisions GPU instances with SSH or Jupyter access, so one control plane covers model serving and raw GPU rental.

The privacy policy names GPUStack Co., Ltd. as operator and says personal data collected in the People's Republic of China is stored there, which documents China as the home market. The software runs wherever customers install it.

GPUStack runs in your own environment, so prompts and model data stay on your hardware. Check the telemetry it sends, which the privacy policy says can be disabled. Clarify which governance features belong to the Enterprise Edition, since pricing is on request. The privacy policy refers to Chinese law for data collected in the PRC. No DPA is published, so ask for one if account or support data is involved.

No reviews yet. Be the first to share how your team uses it.

Every review shows its verification level. We read each one before it goes live. How we check reviews