Paperspace (Gradient)

What Paperspace (Gradient) actually is

Paperspace is a GPU cloud and managed ML platform that started in New York in autumn 2014 as a browser-accessible cloud PC, went through Y Combinator's Winter 2015 class, and pivoted toward GPU compute and machine-learning tooling as the market moved. The product today has two halves: Paperspace Machines (Linux and Windows virtual machines with persistent storage, GPU options, and free unlimited bandwidth, aimed at HPC-style workloads) and Paperspace Gradient (managed Jupyter notebooks, workflows, and deployments for ML teams, sold on subscription plans).

Then ownership rewrote the roadmap. DigitalOcean acquired Paperspace for $111 million in cash in July 2023, and the integration has been moving in one direction since: DigitalOcean's own pages now say all AI and GPU services are unified under the DigitalOcean brand, that existing Paperspace users can still access legacy products, and that new users should look at the DigitalOcean AI/ML suite — GPU Droplets, bare metal GPUs, serverless inferencing — instead. The legacy Gradient and Core APIs were shut off on 15 July 2024 in favor of one unified Paperspace API, and 2026 reviews describe the standalone product as being folded into GPU Droplets, with rate cards changing along the way.

We say this up front because it is the single most important fact in this review: you would be buying into a product line that its own owner is migrating away from as a standalone. For the /category/gpu-cloud shortlist, that belongs in the first paragraph.

Pricing: what you really pay

Gradient's subscription plans are cheap and clearly published: Free at $0 (one member, public projects, 5 GB storage, basic instances), Pro at $8 per month (private projects, 15 GB storage, mid-range instances, faster free machines when available), Pro team at $12 per user per month (two members max), and Growth at $39 per user per month (five members max, high-end machines, expert support). Enterprise is listed as "coming soon". The pricing formula is Total = subscription + utilization: the plan sets which machine types and collaboration features you can use, and compute time is billed per hour on anything outside the free tier. Storage overages cost $0.29 per GB.

Machine rates on the public price list: the M4000 ("GPU+", 8 vCPU, 30 GB RAM, 16 GB dedicated GPU) at $0.45/hour, P4000 at $0.51/hour, A4000 at $0.76/hour, A6000 at $1.89/hour, with multi-GPU variants like P5000 x2 at $1.56/hour and P6000 x2 at $2.20/hour. Modern silicon is priced differently: H100 at $5.95/hour on-demand under a stated promo price, or $2.24/hour on a three-year commitment. Monthly dedicated GPU machines run $268 (M4000), $303 (P4000), and $461 (P5000). The worked example on the site: a Pro subscription ($8) plus 12 hours on a P6000 ($1.10/hour in their example) totals $21.20 for the month.

Note that our database claimed "from $0.07/hour on M4000 GPUs" — the current published rate is $0.45/hour, and we have corrected the row. The "cheap GPU" framing still holds relative to hyperscaler pricing for older Pascal-era cards, but not at seven cents.

Performance & infrastructure

The footprint is small and deliberate: three Paperspace datacenters across three regions — NY2 (New York), CA1 (California), and AMS1 (Amsterdam). Machines run in all three; the ML platform surface (Notebooks, Deployments, Workflows) runs in NY2 only. GPU availability is tiered by region: the M4000 sits in NY2 and AMS1, the workhorse cards (P4000, P5000, P6000, RTX4000, A4000) in all three, V100 in NY2 and AMS1, and the newest silicon — A100, A100-80G, H100, RTX5000 — is NY2-only. If you want an H100 near Amsterdam, this product does not have one.

Machines come with persistent storage and what DigitalOcean's documentation calls free unlimited bandwidth, which is a genuine differentiator for data-heavy training jobs; GPU machines are dedicated rather than time-sliced on the paid tiers. On the ML side, Gradient covers managed notebooks, experiment tracking, workflows, and model deployments — the "notebook to endpoint" path that data science teams want without building MLOps plumbing.

We have not benchmarked Paperspace hardware, and we would not attach numbers to GPU throughput without running them; performance here depends heavily on the card generation and whether your slot is dedicated. What we can say editorially: the older Pascal and Turing cards are cheap for a reason, and for inference or small training jobs they are fine value, while anyone pricing modern training runs should compare DigitalOcean GPU Droplets and the hyperscalers card for card before assuming this is the low-cost option.

The migration mechanics deserve one practical warning. When a platform changes hands, the first things to move are APIs and billing, and both have already moved here: the unified Paperspace API replaced the legacy Gradient and Core endpoints in July 2024, and payment and account management increasingly route through DigitalOcean (their bare metal GPU pages already require a Paperspace account for billing). If you automate against this platform, pin your integration to the unified API and keep an eye on DigitalOcean's release notes rather than Paperspace's blog.

Where it falls short

The transition is the problem. Product pages still describe Gradient in standalone terms while the owner's messaging sends new customers elsewhere; APIs have already been replaced once; and third-party reviews report rate-card changes inside the last year, including a monthly gate on access to high-end silicon. None of that makes the product unusable today, and legacy access is promised for existing users. It does mean that a three-year plan built on Paperspace-specific tooling carries migration risk that a comparable plan on a stable platform does not.

Second, the ML platform is NY2-only and the region count is three. For European teams, Amsterdam covers Machines but not Notebooks or Deployments; for Asia-Pacific teams there is nothing. That is thin coverage in a category where competitors offer a dozen regions.

Third, the free tier is real but narrow — one member, public projects, 5 GB storage, basic instances — and the plan ladder caps members at two on Pro team and five on Growth, which makes the "team" tiers feel like trial steps rather than team products. Storage overages at $0.29/GB are the kind of line item that grows quietly.

There is also a comparison worth doing against DigitalOcean's own GPU Droplets, because the two product lines overlap and the pricing models differ: Paperspace bills subscription plus per-hour utilization on named GPU machines, while GPU Droplets follow DigitalOcean's usual per-hour Droplet model. Depending on card generation and utilization pattern, either can come out cheaper — model your actual training or inference hours rather than comparing headline rates.

Who should pick this

We pick Paperspace today for two cases. First, existing users: if your notebooks, workflows, and machines already live here and the pricing works, staying through the migration is likely cheaper than moving. Second, individuals and small research efforts wanting managed GPU notebooks without infrastructure work — the Free and Pro plans are genuinely cheap ways to get a GPU-backed Jupyter environment.

For everyone else, our pick is to start the comparison at DigitalOcean's GPU Droplets and Paperspace Machines side by side, and to look at Lambda, RunPod, or the hyperscalers for modern GPU generations. Teams needing EU-resident ML tooling should look harder at providers with European ML regions; there is no APAC presence here at all.

One more angle for ML students and hobbyists: the free tier's real value is not the compute credits but the environment — managed notebooks with GPU access, saved state, and a path to deployments when a project outgrows a laptop. For coursework and small experiments it remains one of the lowest-friction ways to get a GPU-backed Jupyter session running, and that is a use case the transition does not change overnight. It is also, notably, the one thing here that costs nothing.

Key takeaways

• GPU machines plus a managed ML platform (notebooks, workflows, deployments), founded 2014 in New York, YC W2015.
• Acquired by DigitalOcean for $111M in July 2023; being folded into DigitalOcean's AI/ML suite, legacy APIs retired July 2024.
• Gradient plans: Free $0, Pro $8/month, Growth $39/user/month — plus per-hour compute; storage overages $0.29/GB.
• M4000 GPU machines at $0.45/hour (not $0.07 as our database claimed); H100 at $5.95/hour on-demand.
• Three datacenters (NY2, CA1, AMS1); the ML platform runs in NY2 only, newest GPUs are NY2-only.
• Our pick for cheap legacy GPU notebooks and existing users; new strategic workloads should compare DigitalOcean GPU Droplets first.

Sources

• https://paperspace.com/pricing
• https://paperspace.com/gradient/pricing
• https://docs.digitalocean.com/products/paperspace/machines/details/availability/
• https://docs.digitalocean.com/platform/regional-availability/
• https://digitalocean.com/solutions/paperspace-joins-digitalocean
• https://docs.digitalocean.com/release-notes/paperspace
• https://techcrunch.com/2023/07/06/digitalocean-acquires-cloud-computing-startup-paperspace-for-111m-in-cash
• https://blog.paperspace.com/paperspace-closes-13m-to-fuel-growth
• https://aisotools.com/blog/paperspace-review-2026