Paperspace

Paperspace's pitch is the unglamorous part of AI work: getting a GPU without an infrastructure project. The stack splits into Core, which gives you fully-managed cloud GPU servers with an elegant management interface and a full API for programmatic access, and Gradient, the managed ML platform. Gradient is where most individuals will live — 1-click hosted Jupyter notebooks including free GPU tiers, pre-configured PyTorch and TensorFlow templates so you go from signup to training in seconds, distributed training via Workflows (beta), deployment of trained models as scalable API endpoints, a model repository, and automatic versioning, tagging and life-cycle management. Compute is billed per second on paid instances, and the marketing claim is savings up to 70% versus major public clouds — worth testing against your own workload rather than taking at face value, because the monthly platform fee ($0 Free, $8/mo Pro, $39/mo Growth) is separate from the hourly instance rate, and anything past your tier's storage allowance is billed at $0.29/GB. The team tiers change shape entirely: T0 is $0 plus utilization with 10GB storage and a single running notebook, T1 is $12/user/mo plus utilization with 500GB storage and a 10-notebook running limit, and T2 with unlimited notebooks and scalable storage is contact-sales. That notebook-count framing matters more than it first appears — it's a collaboration limit, not just a compute limit. Where Paperspace is genuinely differentiated: cost control, with configurable auto-shutdown (12-hour limit on Free and T0) and per-second billing so you stop paying when the instance stops; breadth of GPU choice, from basic instances up to high-end H100/A100/V100; and Graphcore IPU access on demand via the documented partnership, which is rare to find next to NVIDIA capacity in the same console. There's also Portal, low-latency desktop streaming software in limited preview, and a Gradient installer for running Gradient on your own cluster. Where it loses: high-end GPU availability varies with demand, so H100 capacity is not something you can assume; the free and entry tiers cap storage hard; and this is not a no-code tool — you choose instance types, so teams wanting fully serverless inference at massive scale without that decision will find it fussy. If you need integrated feature stores or deep model monitoring, this isn't that platform. And the strategic caveat hanging over all of it: Paperspace is part of DigitalOcean, the site now says the full GPU suite lives at DigitalOcean.com, and new buyers should treat that consolidation as the direction of travel. Use Paperspace if you are an individual ML engineer or a small team that wants notebooks, training and deployment cheaply in one account; go elsewhere if your priority is deep MLOps maturity or a single vendor for everything.