dstack
Open-source control plane for orchestrating GPU workloads anywhere.
About
dstack is an open-source container orchestrator and control plane for GPU workloads. It provisions compute across any GPU cloud, Kubernetes, or on-prem clusters and covers the full AI workflow: dev environments (GPU boxes pre-wired to VS Code or Cursor), tasks (single-node or distributed training), and services (model endpoints and Streamlit-style demo apps). It supports NVIDIA, AMD, Google TPU, Intel Gaudi, and Tenstorrent accelerators.
It is cloud-agnostic and actively developed (0.20.x releases through 2026 added fleet-first UX, InfiniBand support, and a dstack Sky managed marketplace with Nebius GPUs), and companies like EA use it to eliminate GPU-provisioning ticket queues. The honest caveat: dstack is still pre-1.0 in spirit — fast-moving, with UX and concepts evolving release to release — so teams should pin versions and expect some churn in exchange for a genuinely open alternative to managed GPU platforms.
What it can do
- GPU provisioning across any cloud, Kubernetes, or on-prem
- Dev environments, training tasks, and model services
- Multi-accelerator support (NVIDIA, AMD, TPU, Gaudi, Tenstorrent)
- SSH fleets and distributed training primitives
- dstack Sky managed marketplace with on-demand GPUs
Setup
Install via uv tool install dstack[all], run dstack server, configure backends, then dstack apply.
Prerequisites
Linux/macOS/Windows (WSL2), Git, OpenSSH, cloud credentials or on-prem servers.
Use cases
Works with
Trust score — 8/10
Scored 8/10 (verified tier, metrics-v1). Evidence: Verified via vendor site/docs/GitHub and reputable coverage, wave 77; Trending score 80/100 from researcher assessment.
Sources
Every fact above was checked against these sources. Where a claim comes from the provider and couldn't be independently verified, the text says so.