dstack Review (2026): Pricing, Features & Honest Verdict

TLDR

dstack is an open-source orchestration platform that provisions and manages GPU workloads across clouds, Kubernetes, and on-prem servers from one declarative interface. It is the most approachable Slurm alternative we have tested, though its ecosystem is still young. Best for: multi-cloud ML teams without dedicated DevOps. Price: free open source, usage-based hosted tier. Rating: 7.3/10.

What is dstack?

dstack is a unified control plane for GPU provisioning and AI workload orchestration. In practice, that means you describe what you need in a short YAML file, a dev environment with an A100, a distributed training job across 16 GPUs, or a vLLM inference endpoint, and dstack provisions the hardware, runs the workload, and tears it down when finished. It works across AWS, GCP, Azure, Lambda, and other GPU clouds, plus Kubernetes clusters and your own bare-metal servers connected over SSH. The pitch is blunt: Slurm was built for HPC labs and Kubernetes was built for microservices, and neither was built for ML engineers who just want a GPU. dstack is also vendor-agnostic on the accelerator side, with support spanning NVIDIA, AMD, TPU, and Tenstorrent hardware, which is rarer than it should be in this category.

The company behind it, dstack Inc., was founded by Andrey Cheptsov, who previously ran product for PyCharm at JetBrains. That pedigree shows in the developer experience, which feels closer to a polished IDE workflow than a sysadmin tool. The project is fully open source, and the homepage claims engineers across 100+ AI-first companies use it, including Electronic Arts and Argilla. We dug into the docs and the GitHub repo, and the project ships releases at a steady clip, which matters when you are betting your training pipeline on it.

What Are dstack’s Key Features?

Dev Environments

One command spins up a remote GPU development environment you can open in VS Code or reach over SSH. You define the GPU type, disk, and Docker image in YAML, and dstack finds matching capacity in whichever backend you have configured. This replaces the familiar ritual of manually launching a cloud instance, installing CUDA, and forgetting to shut it down. Idle environments can be stopped automatically, which is where most teams quietly bleed money. Because the config lives in your repo, a new hire gets the exact same environment on day one instead of a wiki page of setup steps.

Fleets and SSH Fleets

Fleets let you provision and scale GPU clusters across providers from one config. The more interesting half is SSH fleets: point dstack at existing bare-metal boxes or pre-provisioned VMs, and it absorbs them into the same control plane as your cloud capacity. If you bought a few on-prem H100 servers and still burst to the cloud for big runs, this is the feature that justifies the tool on its own.

Tasks for Training

Tasks handle batch workloads and distributed training. dstack provisions the cluster, wires up the nodes, runs your job, and releases the hardware when it exits. It supports multi-node distributed runs and queueing, the core of what teams traditionally kept Slurm around for, without Slurm’s configuration archaeology. You lose some of Slurm’s deep scheduling controls, like elaborate fair-share policies, so very large research clusters may still want the old guard.

Services for Inference

Services deploy models as production endpoints with autoscaling and gateway support, and they play nicely with vLLM and SGLang serving frameworks. This rounds out the lifecycle: the same tool that gave you a dev box and ran your fine-tune can also serve the resulting model. Most competitors cover one or two of these stages; covering all three is dstack’s strongest argument. A fine-tune that finishes on Friday can be a load-balanced endpoint by Friday afternoon, using the same YAML dialect and the same credentials, with no handoff to a separate serving platform.

How Much Does dstack Cost?

The core product costs nothing. dstack is open source, and self-hosting the server gets you the full orchestration engine with every backend, no feature gates, no seat limits. You pay only your own cloud bills. For a tool in a category where NVIDIA’s Run:ai sells similar orchestration at enterprise contract prices, free is a serious number. The server itself is lightweight enough to run on a small instance or even a laptop while it orchestrates fleets elsewhere, so the hosting cost of the control plane rounds to zero.

dstack Sky is the hosted version for teams that do not want to run the control plane themselves. Pricing is usage-based: you can attach your own cloud accounts or buy GPU capacity through its marketplace, and new accounts get $5 in marketplace credits to test with. Our complaint is transparency. There is no public rate card on the site, so you cannot estimate marketplace markup before signing up. Enterprise, the self-hosted tier with SSO and dedicated support, is custom-priced through a sales call.

Against the field: SkyPilot is also free and open source. Modal, the managed serverless alternative, bills compute per second with roughly $30 in monthly free credits on its starter tier, but locks you into its runtime. Slurm is free but will cost you an engineer’s sanity to operate. dstack’s economics are hard to beat; the only real cost is migration effort.

Who is dstack Best For?

Use dstack if you are an ML team of two to fifty engineers running GPU workloads across more than one provider, or mixing on-prem hardware with cloud bursts. It collapses dev boxes, training, and inference into one workflow, and the open-source license means you can adopt it without a procurement cycle. Teams fleeing a hand-rolled Kubernetes GPU setup will feel the relief within a week.

It is a particularly good fit if:

• You bought on-prem GPU servers and still burst to cloud, and want both behind one interface
• Your researchers keep asking DevOps for instances, and DevOps keeps asking for headcount
• You run AMD or TPU hardware alongside NVIDIA and are tired of tooling that assumes CUDA only
• You want training and inference on the same orchestration layer instead of two stacks

Skip dstack if you are a solo developer with one cloud account and occasional GPU needs; a Runpod or Modal account is less machinery. Skip it too if you run a thousand-GPU research cluster with elaborate fair-share scheduling requirements, where Slurm’s depth still wins, or if your platform team is already fluent in Kubernetes and committed to Kueue. And if your top priority is always paying the absolute lowest spot price across clouds, SkyPilot’s cost-hunting scheduler does that job better.

Best dstack Alternatives

SkyPilot

SkyPilot is the other open-source heavyweight in multi-cloud GPU orchestration, and it is free. Its standout trick is cost arbitrage: at submission time it shops across your configured clouds and regions for the cheapest matching GPU, including spot capacity. dstack counters with a cleaner out-of-the-box experience and built-in inference services. If your cloud bill is the metric you are judged on, start with SkyPilot; if developer experience matters more, start with dstack.

Modal

Modal is the fully managed route: serverless GPU compute with per-second billing and around $30 in free monthly compute on its starter plan. There is no infrastructure to manage at all, which beats dstack for small teams shipping fast. The trade-off is lock-in to Modal’s Python runtime and pricing that runs above raw cloud rates at sustained scale. dstack keeps you on your own accounts and your own terms.

Slurm

Slurm is free, open source, and has scheduled HPC workloads since before deep learning was cool. Large research labs still trust its gang scheduling and fair-share policies for massive shared clusters, territory where dstack is thinner. But Slurm assumes static on-prem clusters and a dedicated admin; cloud elasticity is bolted on at best. For a cloud-native ML team, choosing Slurm in 2026 is choosing a second job.

Final Verdict: Is dstack Worth It?

dstack earns its 7.3/10 by solving a real, expensive problem with an honest license. GPU orchestration across mixed infrastructure is usually a choice between Kubernetes complexity, Slurm archaeology, or a managed platform’s markup. dstack offers a fourth option: declarative, ML-native, and free to self-host. The JetBrains DNA shows in how little time you spend reading docs before something actually runs. In our testing of the configuration flow, the distance from git clone to a running GPU dev environment was measured in minutes, not afternoons, and that first impression holds through the trickier multi-node scenarios.

The caveats are the caveats of youth. The community is smaller than Kubernetes by orders of magnitude, hosted pricing is not published, and cost optimization trails SkyPilot. None of those are disqualifying. If your team runs GPUs on more than one provider and nobody wants to own a scheduler full-time, deploy the open-source server this week; the worst case is you lose an afternoon, and the best case is you delete a pile of Terraform you never liked anyway.