Paperspace Gradient review

Paperspace Gradient is a hosted machine-learning platform: a SaaS workspace for developing, training and deploying models, rather than a library or a local-only CLI. Its product flow starts with hosted notebooks for proofs of concept, moves to GPU machines for training or fine-tuning, then turns models into scalable API endpoints through Deployments.

That places Gradient across the development, compute and serving layers of an ML stack. The homepage describes it as a multi-cloud machine-learning platform and lists end-to-end MLOps workflows. It also advertises Gradient Installer as a way to run Gradient on any compute cluster, alongside its hosted services. The homepage says Paperspace is now part of DigitalOcean and directs users to DigitalOcean for its GPUs and other AI products. Treat the hosted notebook and machine workflow as the main entry point; confirm the installer’s requirements and supported cluster configurations in the docs before planning a self-managed deployment.

The homepage advertises a full API for programmatic access, but does not name an API protocol, SDK, language or CLI. Treat the API as the integration surface and confirm protocol, SDK support and current endpoint documentation in the vendor’s docs before choosing a client library. The product flow also exposes notebooks, machines and deployments in the hosted platform.

Paperspace’s homepage links to a docs hub, tutorials, a quickstart guide, machine types and team management documentation. It also points to sample projects that can be cloned into an account. A Reddit commenter points users to a Run on Gradient link-builder for bringing a GitHub repository or Docker container into a cloneable notebook; verify that workflow in the current docs before using it as a deployment path. The homepage does not name language-specific integrations, webhooks or plugins, so teams should confirm those details with DigitalOcean rather than assume compatibility.

The Free plan is $0 per month and lists public projects, 5GB storage, basic instances and 12-hour auto-shutdown. Pro (first listing) is $8 monthly, with private projects, 15GB storage and mid-range instances; another listing calls Pro (second listing) $12 monthly with similar inclusions. The pricing page says billing runs monthly and lists no annual pricing. Confirm which Pro offer applies before budgeting.

Growth is $39 monthly and lists 50GB storage, high-end instances and expert support. Separate one thread listings show a free tier with 10GB persistent storage, plus a $12-per-user monthly plan with 500GB; a further one thread tier has custom pricing. Enterprise is also custom-priced and includes hosting Gradient on your own infrastructure.

Paid-instance compute is billed separately at hourly rates: listed examples are CPU C1 at $0.0045/hour and GPU M4000 at $0.45/hour. Storage overage is $0.29/GB. Estimate notebook-plan charges and compute usage separately.

The vendor does not publish an open-source license for Gradient on the vendor's site, so do not assume you can inspect or fork its source. The homepage advertises Gradient Installer as a way to run the platform on any compute cluster, while Enterprise lists hosting Gradient on your own infrastructure. That is a self-hosting option, but the published description does not establish whether it is available outside a custom Enterprise agreement; ask DigitalOcean about licensing, deployment requirements and support.

The homepage and pricing information do not describe a data-export process or a supported migration path. A Reddit user reports that files stored on Gradient disappeared, which is an individual report rather than a platform-wide reliability finding. Before committing datasets, model artifacts or notebook state, confirm export formats, retention rules and backups. Keep code and important data in systems you control where possible, and test that you can retrieve them before relying on Gradient for production.

Gradient’s homepage says users can scale training with a range of GPU options and no runtime limits, but the Free plan lists 12-hour auto-shutdown. Paid plans list configurable auto-shutdown. Check how those settings affect long-running jobs, notebook persistence and deployment availability, and get current capacity and rate-limit details from DigitalOcean before setting production expectations. The vendor does not publish rate limits on the vendor's site.

User reports are mixed. One Reddit commenter said the free machine was slower than Colab but more predictable about shutdowns; another reported GPU detection problems with Disco Diffusion. A different user said they stopped paying after files vanished. These are individual experiences, not guarantees about current service behavior. Confirm backup and retention policies, security controls, compliance coverage, support terms and incident procedures with the vendor. For a production rollout, test the full path from notebook or training job to deployed endpoint and include compute and storage overages in cost monitoring.