dstack.ai

What It Is

dstack is an orchestration stack for AI infrastructure, positioned as a unified control plane for GPU supply and workload orchestration. It can connect GPU clouds, Kubernetes, and on-prem clusters, using abstractions such as fleets, dev environments, tasks, services, volumes, and gateways to manage AI workloads. It is not a chatbot or model platform, but a lower-level engineering tool for training, inference, and development environments.

Core Capabilities

Based on the available text, dstack’s core value is cross-backend GPU orchestration: it can provision GPU VMs directly via cloud APIs, connect to existing Kubernetes clusters, or use SSH fleets to manage bare-metal servers and pre-provisioned VMs. On the development side, it supports creating remote GPU-enabled development environments via YAML and connecting with VS Code or SSH. On the runtime side, it supports tasks, services, logs, metrics, events, volumes, gateways, and more. The documentation also includes training and inference examples for SGLang, vLLM, NIM, TensorRT-LLM, TRL, Axolotl, and others, indicating that it is more of an AI infra/MLOps toolchain.

Pricing and Integrations

The collected content does not disclose commercial pricing, SaaS plans, or payment methods, but the homepage clearly offers an open-source installation path, and the server can be self-hosted via uv, pip, or Docker. Actual costs mainly come from the GPU cloud services or self-owned resources you use. Its integrations are relatively strong: the documentation lists CLI, HTTP API, and Python/REST API support, and it works with backends including AWS, Azure, GCP, Lambda, Nebius, Crusoe, Runpod, Vast.ai, Kubernetes, and more. It can also install agent skills to work with Claude, Codex, and Cursor for editing configurations and invoking the CLI.

Pros and Cons

The main advantages are unified management of multi-cloud and on-prem GPUs, reduced cloud vendor lock-in, and no requirement to use Kubernetes or Slurm. Its YAML-, CLI-, and API-centric approach makes it suitable for automation and platform teams. Case studies show Graphsignal using it for inference benchmarking, Toffee using it for multi-cloud inference while reducing GPU costs, and EA using it to accelerate AI project development. The limitations are also clear: it requires familiarity with Docker, SSH, cloud permissions, GPU specifications, and network configuration, so it is not ideal for individual users without infrastructure experience. The collected text does not specify Chinese-language support, commercial service SLAs, privacy/compliance details, or exact pricing.

Who It’s For and Access from China

dstack is suitable for AI platform teams, ML engineers, research teams, and companies that need to schedule GPUs across clouds—especially organizations that already have cloud accounts or their own GPU servers and want to unify development, training, and inference workflows. The text does not provide information about accessibility from China, so this remains unknown. If access to GitHub, Docker Hub, overseas GPU clouds, or documentation is affected by local network conditions, a proxy may be needed. Domestic alternatives or related approaches could include Kubernetes/Slurm/Ray, SkyPilot, or cloud-provider-native GPU and container platforms.