Overview
Paperspace was founded in 2014 and is headquartered in New York, USA. It is a GPU-focused cloud server platform acquired by DigitalOcean in 2023 and folded into its AI/ML product line. The platform offers NVIDIA H100, A100, A5000, and RTX 4000 GPU instances with per-second on-demand billing, helping ML teams avoid the high upfront cost of building their own GPU infrastructure.
Paperspace's core products include Gradient (ML development platform), Machines (GPU cloud instances), and Deployments (model inference), covering the full path from model fine-tuning to production inference. Data centers are located in the US (New York, San Francisco) and Europe (Amsterdam), and the company claims up to ~70% savings on compute costs compared with major public clouds. For budget-conscious teams that need high-performance GPUs for AI model hosting, Paperspace offers strong value.
Key Strengths
- Per-second billed GPU compute: H100 80GB instances can be spun up on demand and stopped when done; the company claims up to ~70% savings vs. major public clouds, ideal for sporadic and bursty training jobs.
- Full range of NVIDIA GPUs: H100, A100 40/80GB, A5000, A4000, and RTX 4000 cover everything from inference to large-scale training. See GPU cloud server comparison.
- One-stop Gradient environment: Built-in Jupyter Notebooks, experiment versioning, dataset mounting, and one-click inference deployment let developers skip server ops and go from training to production.
- No long-term commitments: On-demand pricing with instant instance resizing gives teams flexible control over cloud server budgets.
- DigitalOcean ecosystem synergy: Post-acquisition, GPU resources can be managed alongside Droplets, managed databases, and more, reducing multi-cloud complexity.
Product Ecosystem
Gradient (ML Platform)
Gradient is Paperspace's AI development platform, providing Jupyter-based Notebooks, experiment tracking, dataset and model repository management, and Deployments for inference. Developers can move from a Notebook prototype to GPU training and then to a one-click API endpoint in minutes, making it ideal for quickly validating model fine-tuning approaches.
Machines (GPU Cloud Instances)
Machines provides per-second billed bare GPU instances with NVIDIA H100, A100, A5000, and RTX 4000 options, supporting on-the-fly resizing and snapshots. Teams needing flexible compute for AI infrastructure can scale per task without paying for idle capacity.
Core (Cloud Workstations)
Core delivers cloud-based Linux/Windows desktop workstations with high-performance remote environments, suitable for data labeling, remote work, and scenarios needing unified compute.
Limitations
- Limited data center coverage: Only US (New York/San Francisco) and EU (Amsterdam) regions; latency is higher for users in China and Asia-Pacific, requiring CDN acceleration or other regional providers.
- Small free quota: The free plan offers limited GPU hours and quotas, so sustained training effectively requires paid instances, which can overshoot individual learners' budgets.
- Leaner managed ecosystem: Compared with AWS/Azure, adjacent services such as object storage and fully managed Kubernetes are fewer, so complex production architectures need DIY assembly. See server selection guide.
Use Cases
- AI/ML training and fine-tuning (★★★★★): Per-second billed H100/A100 suited to LLM and diffusion model training plus fine-tuning.
- Data science prototyping (★★★★★): Gradient ships with Jupyter Notebooks out of the box, removing environment setup.
- Model inference deployment (★★★★): Deployments publish models as API endpoints in one click, following AI model hosting best practices.
- GPU rendering and video processing (★★★★): Rent high-end GPUs on demand and avoid one-time hardware purchases.
- Low-latency China-facing workloads (★★): No China nodes; real-time China-facing businesses should evaluate domestic cloud providers.
Pricing
| Instance tier | Approx. hourly rate | Typical specs and use |
|---|---|---|
| Entry GPU | $0.40–$0.60 | RTX 4000 / M4000, light inference and prototyping |
| Mid-range GPU | $1.2–$1.5 | A5000, medium training and rendering |
| High-end GPU | $2.7–$2.9 | A100 80GB, large model training |
| Flagship GPU | ~$3.9 | H100 80GB, large-scale distributed training |
Note: All instances are billed per second on demand; the company claims up to ~70% savings vs. major public clouds. Reserved instances and team subscriptions are also available. Actual prices may vary; check the official website.
FAQ
- How is Paperspace billed? Per-second on-demand billing: you pay only while instances run and can stop and release resources anytime, with no long-term contracts. See cloud server pricing guide.
- What is the relationship between Paperspace and DigitalOcean? Paperspace was acquired by DigitalOcean in 2023 and now serves as its AI/ML GPU platform, manageable alongside Droplets and other products.
- Is there free GPU available? The free plan provides limited entry-level GPU and Gradient resources for trials, but sustained training is best done on paid instances. See GPU cloud server comparison.
- How large a model can it train? Single machines support H100/A100-class cards for small-to-medium fine-tuning and training; very large cluster training requires reviewing network and quotas. See GPU cloud server comparison.
- Can users in China access it reliably? Access from China may have high latency; consider CDN acceleration or a domestic GPU cloud provider.