dstack is an open-source orchestrator for AI workloads that provisions training jobs, inference services, and development environments onto GPU capacity across clouds. You describe what a workload needs in YAML, and dstack finds matching GPUs, launches them, and manages the lifecycle. It serves ML teams that rent GPUs from many providers but want one way to run everything.[1][2]
Quick verdict
AI teams that juggle GPU capacity across clouds gain the most from dstack. Its fleet, service, and gateway model maps multi-cloud GPUs into one deployment target with autoscaling and cost matching. Teams without infrastructure experience should look elsewhere, because every backend needs its own credentials and YAML configuration. If you want GPUs without touching infrastructure at all, choose a fully managed serverless platform instead.[1][2][3]
Key features
Four workload kinds from YAML Define dev environments, tasks, services, and fleets in configuration files, so experiments, batch jobs, APIs, and clusters share one workflow.[1]
Broad backend coverage Provision onto major clouds, specialist GPU providers, and marketplace capacity, plus Kubernetes clusters and on-prem machines over SSH without extra backend setup.[1][2]
Autoscaling inference services Serve models behind gateway endpoints that scale replicas with traffic, with rolling deployments for updates. Service autoscaling is experimental in 0.21.[1]
Fleet management Hold long-lived GPU pools with placement rules and reuse, so recurring training and serving skip repeated provisioning.[1]
Accelerator choice Target NVIDIA, AMD, TPU, and Tenstorrent hardware through the same definitions, with tooling that compares GPU offers by price and availability.[1]
Gateway routing Expose services and dev environments over HTTPS endpoints with centralized access control instead of per-machine tunnels.[1]
Three control surfaces Drive identical operations from the CLI, the Python API, or the REST API, with a dashboard and Prometheus-compatible metrics for inspection.[1]
Agent-ready workflow Install with a single package-manager command and generate fleet, service, and gateway definitions from editor assistants with documented skills.[1]
Who will love it
Multi-cloud training teams: one set of definitions spans reserved and marketplace GPUs, which matters when a single provider cannot cover every workload.[1][2]
Self-hosted inference operators: autoscaled services behind your own gateways keep traffic inside infrastructure you control, with the software itself free under MPL-2.0.[1][2]
Budget-watchful GPU renters: offer comparison across providers plus scale-to-zero services attacks the two largest cost drivers: wrong-region rentals and idle cards.[1][2]
Who should avoid it
Teams without infrastructure help: per-backend credentials, fleet YAML, and gateway setup assume somebody on the team already thinks in infrastructure terms.[1]
No-code builders: everything flows through configuration files, CLIs, and APIs; there is no visual builder or non-developer path.[1]
Compliance-gated buyers in a hurry: no SSO, data-region, or enterprise-support terms were documented at research time, so regulated teams must confirm these directly before committing.[1]
What the reports suggest
One inspected vendor-published deployment account: Toffee's infrastructure architect on replacing Terraform scripts with dstack services and gateways across GPU neoclouds (2026); no Reddit or G2 user report located.
1 cited community source. One vendor-published customer story (2026), not a user sample; no sentiment or prevalence inference possible. No Reddit threads surfaced and no G2 dstack page was located, so user sentiment is uninspected rather than negative.
A 2026 vendor-published case study reports Toffee's infrastructure architect replacing per-provider Terraform scripts with dstack services and gateways across GPU neoclouds, with faster deploys and materially lower GPU spend. This is one vendor-hosted customer account, not a user sample, and it is analyzed here rather than quoted.[1]
Alternatives and compatibility
Beam is the published catalogue alternative for open serverless GPUs with a self-hosting option: Python-first functions, autoscaling endpoints, and scale to zero for teams that prefer managed execution over orchestrating their own capacity. Choose dstack when multi-cloud provisioning inside your own accounts matters more than a managed service.[1]
Modal is the published catalogue alternative for fully managed serverless GPUs without infrastructure decisions. Choose dstack when you need one control plane across many GPU providers and can own the YAML-defined operations it requires.[1][2]
Questions worth asking
What is the best dstack alternative for non-technical users?
Modal is the safer pick for teams that want serverless GPUs without touching infrastructure: a fully managed Python platform with autoscaling endpoints and sandboxes for similar training and inference workloads. Choose it when you value a managed service over multi-cloud portability. This stays conditional because no head-to-head testing was done in this pass.[1]
Is dstack really no-code?
No. dstack is YAML-and-CLI-first: workloads, fleets, and gateways are declared in configuration and driven from the command line or APIs. It removes per-cloud scripting, not programming or infrastructure thinking.[1]
How much does dstack cost per month?
There is no flat monthly price: the software is free under MPL-2.0, and compute is billed by each cloud provider rather than by dstack. As checked on October 9, 2026, teams either bring their own cloud accounts or provision marketplace capacity through dstack Sky, so monthly spend equals whatever GPUs you consume plus any provider fees.[1][2][3]
Can you self-host dstack?
Sources and editorial review
Source research. Reviewed on October 9, 2026. Reviewed by Sanjib Sah. Written by Sanjib Sah. AI assisted; reviewed by a human editor.
We have not run or tested this product.
One inspected vendor-published deployment account: Toffee's infrastructure architect on replacing Terraform scripts with dstack services and gateways across GPU neoclouds (2026); no Reddit or G2 user report located. One vendor-published customer story (2026), not a user sample; no sentiment or prevalence inference possible. No Reddit threads surfaced and no G2 dstack page was located, so user sentiment is uninspected rather than negative.
- dstack documentation
official · Observed October 9, 2026
Docs intro: workload kinds, server/CLI/Python/REST/dashboard/gateways/volumes/metrics, accelerators, cost optimization, install, agent skills; What's new 0.21.0 and 0.20.x notes.
- dstackai/dstack on GitHub
official · Observed October 9, 2026
dstackai/dstack public repository, MPL-2.0, Python/Go/TypeScript; 0.21.0 tagged August 2026; gateway replica and metrics-scaling guides. No star count asserted (ledger/live disagreement).
- dstack backends
official · Observed October 9, 2026
Backend coverage: VM-based GPU clouds, Kubernetes clusters, SSH fleets for on-prem and marketplaces; SSH fleets need no backend configuration.
- dstack - orchestration for AI workloads
official · Observed October 3, 2026
Homepage positioning as open-source alternative to Kubernetes for AI (intake ledger, Oct 3).
- dstack Sky
official · Observed October 3, 2026
Managed GPU capacity marketplace usable without owning a cloud account (intake ledger, Oct 3; Nebius addition confirmed via search).
- How Toffee streamlines inference and cut GPU costs with dstack
community · Observed October 9, 2026
Vendor-published customer story quoting Toffee's infrastructure architect on replacing Terraform scripts with services/gateways; links onward to the original Toffee engineering post. Analyzed, not quoted (not Reddit/G2); vendor-hosted, so not an independent sample.