Fleet management
Define fleets to provision and manage clusters across GPU clouds, Kubernetes, and on-prem environments. Fleets can be used to control how instances are provisioned and reused.
dstack
Rivendicadstack is an open-source control plane for AI workloads across GPU clouds, Kubernetes, and on-prem clusters, from a single YAML and CLI workflow.
What dstack is
dstack is an open-source orchestration layer for heterogeneous AI compute. It acts as a unified control plane for provisioning GPUs and running workloads across GPU clouds, Kubernetes, and on-prem clusters.
The product is designed for containerized AI workloads and covers development, training, and inference. Users define fleets, dev environments, tasks, services, and volumes in YAML, then apply those configurations through the CLI or API while dstack handles infrastructure provisioning and workload scheduling.
Core capabilities
Dev environments
Create interactive dev environments that agents or IDEs can access, so developers can work close to the target compute without manual infrastructure setup.
Tasks
Schedule single-node or distributed tasks for training and batch jobs. The docs also note that tasks can run web apps.
Services
Deploy model inference and other web-facing workloads as services, with dstack handling the orchestration around them.
Volumes
Manage network volumes to persist data across workloads and runs.
CLI, UI, and API
Use a CLI, UI, and API to manage workloads, with YAML-based configuration files stored in a repo.
Common use cases
Multi-cloud GPU operations
Provision clusters across multiple GPU clouds or mixed infrastructure, then manage them through one control plane instead of separate vendor tools.
AI development environments
Create reproducible development environments that can be opened by a developer or an agent from an IDE, reducing local setup work.
Model training and batch workloads
Run training, batch jobs, and distributed experiments on single-node or multi-node compute with the same configuration model.
Model serving
Deploy inference endpoints or other services that need a secure, scalable runtime on top of your existing compute.
Bring-your-own infrastructure
Connect existing Kubernetes clusters or SSH-accessible servers when you already have infrastructure in place and want dstack to orchestrate workloads on it.
Pros and Cons
Pros
- Works across GPU clouds, Kubernetes, and on-prem clusters from a single control plane.
- Supports multiple accelerator types, including NVIDIA, AMD, TPU, and Tenstorrent.
- Covers the full ML workflow from dev environments to tasks and services.
- Provides both CLI and UI access, plus a programmatic API.
- Includes support for existing Kubernetes clusters and SSH-based on-prem fleets.
Cons
- The public pricing details are not clearly laid out on the fetched pages, so buyers may need to contact sales or review hosted options to understand commercial terms.
- The main workflow is configuration-driven, so teams that want a fully managed experience may need to adopt dstack's YAML and CLI model.
FAQ
How do I get started with dstack?
You can install dstack locally with `uv` or deploy it with the `dstackai/dstack` Docker image. The docs also note that you can sign up for dstack Sky instead of self-hosting the server.
How does dstack fit into a workflow?
dstack works by defining configurations in YAML files in your repository, then applying them with `dstack apply` or through the programmatic API. It provisions infrastructure and schedules workloads while handling autoscaling, port-forwarding, and ingress.
Can dstack be used with existing Kubernetes or on-prem infrastructure?
Yes. The docs say you can connect existing Kubernetes clusters through the Kubernetes backend and run dev environments, tasks, and services on them. dstack also supports SSH fleets for on-prem servers.
What is dstack's pricing model?
The source pages do not show a public pricing table for the core product. They do show hosted options such as dstack Sky and dstack Enterprise, with Sky described as hosted by dstack and Enterprise as self-hosted with SSO, air-gapped setup, and dedicated support.
Who is dstack for?
dstack is built for development, training, and inference on heterogeneous AI compute. The docs emphasize fleets, dev environments, tasks, services, and volumes as the main configuration types.
Quick Facts
- Category
- AI infrastructure orchestration
- Source domain
- dstack.ai
- Deployment model
- Open-source server with self-hosted and hosted options
- Primary interfaces
- CLI, UI, and API
- Supported backends
- GPU clouds, Kubernetes, SSH fleets
- Accelerators mentioned
- NVIDIA, AMD, TPU, Tenstorrent
Alternative a dstack
Firebase Studio
Firebase Studio is a web-based workspace for full-stack app development with Gemini-assisted coding, app previews, cloud emulators, collaboration, and browser deployment.
Kastra
Kastra authorization infrastructure for AI systems checks prompts, tool calls, shell commands, API requests, and browser actions before execution.
Termo
Termo is an AI agent platform with dedicated VMs for browsing, code execution, memory, and scheduled automation for research and build tasks.
Orca
Orca is an Agent Development Environment for shipping with coding agents, running multiple CLI agents in parallel across isolated worktrees, with desktop and mobile workflows.
EZsite AI
EZsite AI is an AI website builder that turns a URL into a fullstack React or Vue.js app with hosting, custom domains, code export, and backend features.
AI Magicx
AI Magicx is a unified AI workspace for chat, image, video, voice, music, email and developer tasks, helping teams and creators manage multiple models in one place.