---
title: "DigitalOcean GPU Droplets review: pricing, pros and cons"
description: "Hourly NVIDIA and AMD GPU servers you can start with a card, for UK and EU teams training, fine-tuning or serving models without a sales call."
canonical: "https://saaskly.com/category/gpu-compute/digitalocean-gpu"
last-updated: "2026-09-22"
---

# DigitalOcean GPU Droplets review: pricing, pros and cons

> Saaskly verdict: **3.8/5**. Hourly NVIDIA and AMD GPU servers you can start with a card, for UK and EU teams training, fine-tuning or serving models without a sales call.

- Category: [GPU & High Performance Compute](https://saaskly.com/category/gpu-compute.md)
- Vendor website: https://www.digitalocean.com/products/gpu-droplets
- Last updated: 2026-09-22

## Pros and cons

**Pros**

- H100 at $4.41 per GPU-hour on demand drops to $3.26 on a published 12-month reservation, a 35 per cent gap, with no sales call needed for either
- GPU Droplets include 10,000 GiB of outbound transfer on single-GPU plans and 15,000 GiB on eight-GPU plans, pooled across the team, with overage at $0.01 per GiB and free ingress
- Managed Kubernetes takes GPU node pools on H100, H200, MI300 and MI325, alongside an official Terraform provider, the doctl CLI and a published MCP server

**Cons**

- A stopped GPU Droplet is still billed in full because the hardware stays reserved, and GPU Droplets get no 672-hour monthly cap, so an idle month costs 730 hours
- London has no GPU Droplets of any kind. Amsterdam is the only EU GPU site and carries H100 only, so UK data residency and GPU compute cannot be combined here
- The GPU SLA is 99 per cent a month against 99.99 per cent for CPU Droplets, credits stop at 50 per cent, and a claim needs written notice within 24 hours of the downtime
- New accounts face an unpublished GPU Droplet limit that is raised by request rather than automatically, and September 2026 reviews report prepayment of up to $1,000 being asked for

## Pricing

Month-to-month price first (what you pay without committing), then the annual-commit rate.

- **H100 80 GB, one GPU**: $3,219/mo month-to-month; $2,380/mo on an annual commitment (month-to-month costs 35% more)
  - Includes: One NVIDIA HGX H100 80 GB (20 vCPU / 240 GiB RAM / 720 GiB NVMe boot / 5 TiB NVMe scratch / 15,000 GiB transfer) at $4.41 per GPU-hour on demand, $3,219 a month at 730 hours; reserved from $3.26 per GPU-hour ($2,380 a month) on a 12-month commitment, a 35 per cent gap. Eight H100s on demand are $35.28 an hour, about $25,754 a month. USD, VAT not stated, read 22 Sep 2026.
- **RTX 4000 Ada 20 GB, one GPU**: $555/mo, same price on any term
  - Includes: One NVIDIA RTX 4000 Ada 20 GB (8 vCPU / 32 GiB RAM / 500 GiB NVMe / 10,000 GiB transfer) at $0.76 an hour, $555 a month at 730 hours; no reserved rate published. Volumes $0.10 per GiB a month, billed while detached; transfer overage $0.01 per GiB; ingress free. Toronto only.

Minimum commitment: **$3,219** ($3,219/mo × 1 seat × 1 month (no commitment)).

## Specification

| Attribute | Value |
| --- | --- |
| H100 per GPU-hour* | $4.41/h (NVIDIA H100 80 GB) |
| Committed H100 | $3.26/h (12 mo), 26% off |
| Cheapest GPU | $0.76/h (NVIDIA RTX 4000 Ada) |
| 8xH100 node | $35.28/h |
| Spot / interruptible | Yes |
| Cheapest dedicated server | None |
| Billing | Per second |
| Egress | $0.010/GB |
| Storage (per GB/mo) | $0.100 |
| Self-serve signup | Yes |
| GPU quota | Quota ticket |
| Hardware** | Own hardware, colocated |
| UK location | Yes |
| H100 in the UK | No |
| EU locations | 2 |
| Locations | 16 |
| Owned in | US-owned |
| Owner | DigitalOcean Holdings, Inc. (NYSE: DOCN) |
| Cluster networking | Yes |
| Kubernetes | Yes |
| Slurm | No |
| Bare metal | Yes |
| API | Yes |
| Terraform | Yes |
| MCP server | Yes |
| Uptime SLA (%) | 99.00 |
| Trustpilot (reviews) | 4.5 (2,572) |
| Committed pricing | Published |
| Committed term (months) | 12 |
| Dedicated server term | Contract with DigitalOcean sales, term not published |
| Stopped instance billed | Yes |
| Free credit | $200 for 60 days on signup (third-party sourced, not confirmed on a DigitalOcean page). Hatch startup programme reported at up to $100,000 over 12 months, partner introduction required |
| Sales-only above | Bare Metal GPUs (8-GPU H100, H200 and MI300X nodes) |
| UK cities | London (LON1, no GPUs) |
| HQ | USA |
| Publicly listed | Yes |
| Sovereign offering | None |
| ISO 27001 | Yes |
| SOC 2 | Yes |
| NVIDIA GPUs | Yes |
| AMD GPUs | Yes |
| Multi-GPU nodes | Yes |
| Dedicated servers | No |
| Object storage | Yes |
| Block storage | Yes |
| Private networking | Yes |
| CLI | Yes |
| 2FA | Yes |
| Phone support | No |
| Support channels | Email, Live chat, Slack, Video call |

* One NVIDIA H100 80 GB per GPU-hour, on demand, in the currency the vendor shows a UK visitor; where the vendor sells no H100 the nearest GPU is named. Committed prices are the cheapest published rate and its minimum term.

** Who owns and runs the GPUs. "Third-party hosts" means the vendor is a marketplace for other people's machines and the SLA is theirs, not the vendor's.

## Our review

## Who owns it

DigitalOcean Holdings, Inc., headquartered in New York and listed on the New York Stock Exchange as DOCN since its March 2021 IPO. It is a US company with US ownership, and the trust pages do not name a separate UK or EU contracting entity. The GPU platform came from Paperspace, acquired in July 2023 and still only partly folded in: the Bare Metal GPU page states that a Paperspace account is required to manage payments, and the Paperspace storefront still publishes its own, higher, prices. What was retired is the software layer around it, with the legacy Gradient and Core API endpoints deprecated for Paperspace users from 15 July 2024 and the Gradient Python package carrying a retirement date of 15 August 2026.

The hardware is colocated rather than owned. DigitalOcean runs its own servers, but the buildings belong to third parties: the trust page lists ISO 27001 against its "colocated data center facilities", naming AMS2, AMS3, ATL1, BLR1, FRA1, LON1, NYC1 to NYC3, SFO1 to SFO3, SGP1, SYD1, TOR1, RIC1, MEM1 and MKC1.

## What it costs

Prices are shown in US dollars to a UK visitor, with VAT not stated. All figures read on 22 September 2026.

An NVIDIA HGX H100 with 80 GB is $4.41 per GPU-hour on demand, which is about $3,219 a month at 730 hours. Commit for 12 months and the same card is $3.26 an hour, roughly $2,380 a month, so the on-demand rate carries a 35 per cent premium. The GPU arrives in a Droplet with 20 vCPU, 240 GiB of RAM, a 720 GiB NVMe boot disk and a 5 TiB NVMe scratch disk, and you can buy one GPU or eight. Eight H100s on demand is $35.28 an hour, about $25,754 a month. An H200 is $4.47 an hour on demand and $3.40 reserved, an AMD MI300X is $2.59 and $1.91, an MI325X is $3.80 and $2.88. The cheapest card of any kind is the RTX 4000 Ada with 20 GB at $0.76 an hour, about $555 a month, in a smaller Droplet with 8 vCPU and 32 GiB of RAM. An L40S or an RTX 6000 Ada is $1.57. Spot capacity exists on the newest parts, but the published spot rates for B300 at $11.19 and MI355X at $4.50 sit above the equivalent on-demand and reserved rates, which contradicts DigitalOcean's own documentation that spot is always at or below on demand.

Around the compute, block storage volumes are $0.10 per GiB a month and are charged whether or not they are attached to a Droplet. Spaces object storage is $5 a month for 250 GiB of storage and 1 TiB of outbound transfer, with overage at $0.02 per GiB stored and $0.01 per GiB moved, and a cold tier at $0.007 per GiB. Outbound transfer included with a GPU Droplet is 10,000 GiB on the single-GPU plans and 15,000 GiB on the eight-GPU class, pooled across the team, with overage at $0.01 per GiB and free ingress. Billing is per second, although DigitalOcean's pricing page and its documentation disagree on the minimum, the first saying a five-minute minimum billing period and the second 60 seconds or $0.01, whichever is higher. Two things to plan around: a powered-off Droplet is still billed in full, because "the compute resources stay reserved", so destroying it is the only way to stop the meter, and GPU Droplets have no 672-hour monthly cap of the kind CPU Droplets get, so a month of continuous running is 730 billed hours. Billing is postpaid by card.

For the persona, the smallest useful bill is an RTX 4000 Ada left running, at $0.76 an hour or about $555 a month before storage. A short job is cheaper still: a 20-hour fine-tune on one H100 is 20 x $4.41, about $88, plus $0.10 per GiB a month for any volume you keep afterwards. Bare metal has no published price at all.

## What you get and what to watch

Signup is self-serve with a card, and a GPU Droplet can in principle be launched the same day. What stands in the way is the account limit. DigitalOcean does not publish the default GPU Droplet quota for a new account, and the route to raise it is a request in the control panel under Settings and Profile, or a support ticket, rather than an automatic increase. September 2026 Trustpilot reviews describe a new account restricted from the larger advertised Droplets despite valid payment, with access said to require months of billing history or up to $1,000 prepaid, and a DigitalOcean community question from November 2025 shows a GPU limit increase request going unanswered on the page. Bare Metal GPUs are a different purchase entirely: 8x H100, 8x H200 or 8x MI300X nodes with dual Xeon Platinum CPUs, 2,048 GiB of RAM, 61.44 TiB of NVMe and RDMA interconnect up to 3.2 Tbps, but they are "provisioned via contract with DigitalOcean sales (not on-demand)", with no public price.

Where the GPUs sit matters more here than at most providers. DigitalOcean has 16 data centres across 13 regions including London, but GPU Droplets are in only a handful of them. H100 is in New York (NYC2), Amsterdam (AMS3) and Toronto (TOR1). H200 is in New York and Atlanta. AMD MI300X is in Atlanta. The L40S, RTX 4000 Ada and RTX 6000 Ada are in Toronto only. Spot parts sit in Atlanta, Richmond, Memphis and Kansas City. London has no GPU Droplets of any kind, and Amsterdam is the only GPU site in the EU. Bare metal is listed in New York, Amsterdam and Atlanta, although the product page names only the first two.

The platform is the familiar DigitalOcean one. There are eight-GPU nodes, managed Kubernetes with GPU node pools covering H100, H200, MI300 and MI325 (AMD by request, on Kubernetes 1.33.1-do.5 or higher with the AI/ML Ready images), block and object storage, VPC networking, a public API, an official Terraform provider, the doctl CLI and two-factor authentication. There is an official MCP server at github.com/digitalocean/digitalocean-mcp, although it is built around App Platform rather than GPU Droplets. There is no Slurm. Compliance covers SOC 2 Type II and SOC 3 Type II audited by Schellman, ISO 27001 on the colocated facilities, PCI-DSS SAQ-A validation under a stated zero-footprint card policy and CSA STAR Level 1. HIPAA is not claimed, and the CLOUD Act is not addressed anywhere on the trust pages.

Support is sold in four tiers: free with a response under 24 hours by email, Developer at $24 a month under eight hours with live chat, Standard at $99 a month under two hours with a dedicated Slack channel, and Premium at $999 a month under 30 minutes with video calls. There is no phone support at any tier, and DigitalOcean states that response times do not vary by severity. The SLA is per product, and the GPU one is the weaker: 99 per cent monthly uptime measured at the server node, with credits of 10 per cent between 95 and 98.99 per cent, 25 per cent between 90 and 94.99 per cent and 50 per cent below 90 per cent, against 99.99 per cent and a full 100 per cent credit for CPU Droplets. Claims need written notice within 24 hours of the downtime and a request inside three months, and credits are vouchers valid 90 days, issued only above $1 and never refundable. On Trustpilot, DigitalOcean scores 4.5 from 2,572 reviews, 76 per cent five-star and 19 per cent one-star, and the reviews are not invited. The one-star themes in September 2026 are about account handling rather than uptime: government ID and a biometric selfie demanded in order to cancel (2 September), an account locked without explanation after a basic server setup with support declining to explain, refund or confirm billing had stopped (15 September), the new-account Droplet restriction described above (7 September), and a lockout on an $8 a month service with no clear route to a human (15 September).

*Not yet tested by Saaskly. Assessed on published pricing, vendor documentation, company filings, the vendor's own status and trust pages, and third-party reviews.*

---

Researched and drafted with AI assistance; reviewed and approved by Tim Meredith. See https://saaskly.com/ai-transparency.md and https://saaskly.com/editorial-policy.md.

Canonical HTML: https://saaskly.com/category/gpu-compute/digitalocean-gpu · Markdown index: https://saaskly.com/llms.txt