---
title: "Best GPU Cloud and High Performance Compute for UK Business in 2026: OVHcloud Wins in Pounds, RunPod Is Fastest to a GPU, and the H100 Is Rarely in Britain"
description: "Twelve GPU clouds, bare-metal and dedicated-server providers compared on what one H100 really costs an hour, whether a new account can get one today, who owns the hardware and where it sits. OVHcloud wins in pounds, RunPod is fastest to a GPU, and only two providers had an H100 in Britain."
canonical: "https://saaskly.com/comparisons/best-gpu-cloud-high-performance-compute-2026"
last-updated: "2026-09-22"
---

# Best GPU Cloud and High Performance Compute for UK Business in 2026: OVHcloud Wins in Pounds, RunPod Is Fastest to a GPU, and the H100 Is Rarely in Britain

> Twelve GPU clouds, bare-metal and dedicated-server providers compared on what one H100 really costs an hour, whether a new account can get one today, who owns the hardware and where it sits. OVHcloud wins in pounds, RunPod is fastest to a GPU, and only two providers had an H100 in Britain.

- Type: Comparison
- Published: 2026-09-22
- Updated: 2026-09-22
- Categories: [GPU & High Performance Compute](https://saaskly.com/category/gpu-compute.md)

The price on a GPU cloud's pricing page is rarely the price you pay. Vultr advertises an H100 at $2.30 an hour, and its own API says $2.99 on demand, because $2.30 is the interruptible rate. CoreWeave's $49.24 an hour is for eight H100s, and you cannot buy fewer. Lambda charges more for one H100 than it does per GPU for eight. Nebius has told its customers the H100 goes from $3.85 to $4.50 an hour on 1 October 2026. So for every provider here we priced the same thing: one NVIDIA H100 80 GB for one hour, on demand, in the currency the vendor shows a UK visitor, then the cheapest committed rate actually published, then what the bill does when you stop the machine. Four providers keep charging for a stopped GPU in full; Nebius and Amazon stop the compute charge and keep billing the disk.

Then we asked whether a small team can get the GPU at all, because in this category the paperwork is the price. CoreWeave has no self-serve signup: its sales team approves your organisation and sends the invitation. A new Amazon account has a GPU quota of zero vCPUs and no published waiting time for the request that lifts it. OVHcloud raises GPU quotas manually and says eligibility depends on account seniority and paid invoices. DigitalOcean and Civo lift limits by request. RunPod, Nebius, Hyperstack and Lambda will let a card-paying account launch the reference GPU the same day. And then there is stock: on 22 September 2026 every one of Vultr's multi-GPU node plans returned an empty list of locations, and Scaleway's own availability API flagged every H100 type as scarce or in shortage.

And we asked where the hardware sits and who owns it. Only two of the twelve could sell you an H100 in a British data centre when we looked: Civo in London and Amazon EC2 in London. OVHcloud's H100 is in Gravelines, DigitalOcean's in Amsterdam, Nebius's in Finland, and Vultr, which runs London and Manchester, had no GPU of any kind in either. Five providers own their data centres (OVHcloud, Scaleway, Hetzner, Nebius and Amazon); six own the servers and rent the halls; RunPod's cheaper tier is a marketplace of other people's machines, and its terms say so.

Twelve providers. Vultr is partly first-party: Saaskly and Fractional Teams run their ordinary servers on it and have not run its GPUs, so its GPU verdict rests on the research alone. Amazon EC2 is first-party on the negative side: we have run client workloads on it and moved them because the bill was hard to work out in advance and carried charges driven by usage and bandwidth we had not expected. Nebius is listed and marked down hard under our origin rule, and its profile says exactly why. Genesis Cloud was on the list and is not on the page: the German company entered liquidation in August 2025 and its site was offline when we checked, so Civo took its place. Prices were read on 22 September 2026 in the currency each provider shows a UK visitor: pounds at OVHcloud, euros at Scaleway and Hetzner, dollars everywhere else. Everyone but Vultr and Amazon is labelled "Not yet tested by Saaskly".

> **Note:** Quick take: OVHcloud wins because one H100 costs £2.41 an hour in pounds, the same per GPU whether you take one or four, with free data transfer, no setup fee anywhere in the range, a bare-metal line from £8.89 a month, an official MCP server and the broadest certifications on the page, from an EU company that owns its data centres; go in knowing the H100 is in France, GPU quotas are raised by hand, and there is no spot or committed discount. RunPod is the fastest route to a GPU today, per second, from $3.49. Civo is the only UK-owned provider with an H100 in London.

## The winner

### OVHcloud: the H100 in pounds, with nothing bolted on

OVHcloud wins on the arithmetic a UK buyer actually does. One H100 80 GB with 30 vCores and 380 GB of RAM is £2.41 an hour, and it stays £2.41 per GPU at two and four GPUs, where most rivals charge a premium for the smaller slice. Take it monthly and a single H100 is £1,670. Data transfer in and out is free in the European regions, the public IPv4 is free, block storage runs from £0.03 to £0.06 per GB a month, and there is no setup fee on anything in the GB catalogue. Below the GPU line sits the cheapest dedicated server on this page, the KS-B at £8.89 a month, or £101.35 for a year paid upfront, and the cheapest hourly bare metal at £0.435 an hour. The company builds its own servers and data centres, publishes the entire UK price list through an API that needs no login, and holds ISO 27001, SOC 1, 2 and 3, HDS, SecNumCloud, C5 and G-Cloud UK, the longest compliance list here. There is an official Terraform provider, a CLI and an official MCP server.

Three things to know before you sign. The H100 is not in Britain: it runs at Gravelines in northern France, the eight-way H200 node (£36.12 an hour) is in Paris, and London Erith carries only 2019-era Tesla V100S cards in the price data while OVHcloud's own availability page says London has no GPU at all. Second, the discounts do not reach the GPUs: the 15% and 30% Savings Plans exclude GPU instances, so the only saving on an H100 is the 5% for monthly billing, and there is no spot tier. Third, quotas are manual. GPU increases depend on account seniority and paid invoices, go through support and carry no published turnaround, so a brand-new account should not plan to launch an H100 this afternoon. On the record: the 2021 fire destroyed the SBG2 data centre in Strasbourg, some paid-for backups were held in the same building, and a French court ordered €250,000 to two customers who lost data. Trustpilot is 3.8 on 8,629 invited reviews, half five-star and 36% one-star, with billing the one-star theme. Not yet tested by Saaskly.

## The runner-up

### RunPod: a GPU in the time it takes to add credit

RunPod is the answer to the question most small teams are actually asking, which is how to have an H100 running before lunch. You add prepaid credit through Stripe, pick a region and deploy, with no sales call, no business verification and no quota form; the only ceiling is an $80 an hour spending limit that lifts itself as the account builds history. One H100 SXM is $3.49 an hour in Secure Cloud and $2.69 in Community Cloud, billed per second, and data transfer costs nothing in either direction. The cheapest GPU on the page that you can actually launch is here too: an RTX A5000 at $0.16 an hour. RunPod holds ISO/IEC 27001:2022 under a dated certificate, plus SOC 2 Type II and SOC 3, and it ships the runpodctl CLI and a first-party hosted MCP server, which almost nobody else in this category does.

The two tiers are the thing to understand. Secure Cloud runs in what RunPod calls T3 and T4 partner data centres. Community Cloud is a marketplace of third-party hosts, and the terms of service say RunPod "is only the venue to connect Hosts with consumers", with no availability warranty; RunPod has also stopped accepting new Community hosts, so that pool is shrinking. There is no UK region: the European sites are the Netherlands, France, Sweden, Czechia, Romania and Iceland. Storage on a stopped pod costs $0.20 per GB a month, double the running rate, and a pod without a network volume is destroyed rather than paused when the credit runs out, which is what the one-star review of 24 August 2026 describes. Savings Plans exist on 3 and 6 month terms but the discount is not published, and no spot rate is either. Trustpilot is 3.8 on 315 invited reviews with 19% one-star, and the dated complaints are GPU supply and support replies of 36 to 72 hours. Not yet tested by Saaskly.

### Why OVHcloud edges it

The bill and the contract. OVHcloud prices in pounds, contracts from an EU company, runs its own data centres and publishes the certifications to prove it; RunPod prices in dollars, contracts from a US company under Delaware law, rents its Secure halls from partners and runs a marketplace beneath them. On the H100 itself the two are close, and RunPod is cheaper to leave idle because it bills per second. Where RunPod wins outright is speed: it will have a GPU running for a new account today, and OVHcloud may not. If you need a GPU this afternoon and can live with a US contract and no UK region, buy RunPod. If you want the bill in pounds and the hardware in Europe under an EU contract, and you can wait for the quota, buy OVHcloud.

## The UK pair

**Civo (3.9)** is the only UK-owned provider on the page with an H100 in a British data centre. Civo Limited, London, company 09568551, sells an H100 SXM at $2.99 an hour in its London region, $2.49 on a 36-month term, with committed terms of 6, 12, 24 and 36 months published rather than quoted, an eight-GPU node at $23.92 an hour, free and unlimited data transfer, a 99.95% SLA with a 100% credit below 95% uptime, ISO 27001 and SOC 2, an official Terraform provider and CLI, and $250 of credit for a new card. The catches are real: prices are in US dollars excluding VAT from a UK company, a stopped instance is billed "exactly the same as an active instance", an unpaid invoice means suspension after eight days and deletion after two weeks, quota numbers are published nowhere and rise only after a period of proving yourself, SMS verification is mandatory, virtual and prepaid cards are refused, and the GPUs sit in London only with no spot tier and no cluster fabric mentioned. Trustpilot is 4.0 from three reviews, which is no signal either way. Not yet tested by Saaskly.

**DigitalOcean GPU Droplets (3.8)** is the trusted name with the wrong map. The H100 is $4.41 an hour on demand and $3.26 on a published 12-month reservation, a 35% gap you can take without a sales call; an eight-GPU node is $35.28 an hour; the RTX 4000 Ada starts at $0.76. Single-GPU plans include 10,000 GiB of transfer with a cent-a-gigabyte overage, managed Kubernetes takes GPU node pools, and there is an official Terraform provider, the doctl CLI and a published MCP server, though that server is built around App Platform. Trustpilot is 4.5 on 2,572 uninvited reviews, the best record of any US provider here. But London has no GPU Droplets of any kind; Amsterdam is the only EU GPU site and carries H100 only. A stopped GPU Droplet bills in full with no monthly cap, the GPU SLA is 99% against 99.99% for CPU Droplets and a claim needs written notice within 24 hours, the new-account GPU limit is unpublished and lifted by request, and the bare-metal GPU nodes are sales-only. Not yet tested by Saaskly.

## Full comparison

| Provider | Rating | H100 per GPU-hour* | Committed H100 | Cheapest GPU | 8xH100 node | Spot / interruptible | Cheapest dedicated server | Setup fee | Billing | Egress | Storage (per GB/mo) | Self-serve signup | GPU quota | Hardware** | UK location | H100 in the UK | EU locations | Locations | Owned in | Owner | Origin | Cluster networking | Kubernetes | Slurm | Bare metal | API | Terraform | MCP server | Uptime SLA (%) | Trustpilot (reviews) |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| [OVHcloud GPU & Bare Metal](https://saaskly.com/category/gpu-compute/ovhcloud-gpu.md) | 4.2/5 | £2.41/h (NVIDIA H100 80 GB (h100-380)) | None | £0.34/h (NVIDIA RTX 5000 (rtx5000-28)) | Not on demand | No | £8.89/mo (KS-B (Eco): Intel Xeon E5-1620v2 4C/8T, 1U OVH frame, 99.9% SLA (GBP 101.35 for 12 months paid upfront)) | None | Per hour | Included | £0.060 | Yes | Quota ticket | Own data centres | Yes | No | 7 | 15 | EU-owned | OVH Groupe SA (Klaba family control) | — | No | Yes | No | Yes | Yes | Yes | Yes | — | 3.8 (8,629) |
| [RunPod](https://saaskly.com/category/gpu-compute/runpod.md) | 4.0/5 | $3.49/h (NVIDIA H100 SXM 80 GB (Secure Cloud)) | Sales only | $0.16/h (NVIDIA RTX A5000 24 GB (Community Cloud)) | Not on demand | No | None | — | Per second | Included | $0.100 | Yes | Automatic limit raise | Own hardware, colocated | No | No | 5 | 18 | US-owned | Independent | — | Yes | No | Yes | No | Yes | No | Yes | — | 3.8 (315) |
| [Civo](https://saaskly.com/category/gpu-compute/civo.md) | 3.9/5 | $2.99/h (NVIDIA H100 SXM 80 GB) | $2.49/h (6 mo), 17% off | $1.09/h (NVIDIA A100 40 GB) | $23.92/h | No | None | — | Per hour | Included | $0.110 | Yes | Quota ticket | Own hardware, colocated | Yes | Yes | 1 | 5 | UK-owned | Independent | — | No | Yes | No | No | Yes | Yes | No | 99.95 | 4.0 (3) |
| [DigitalOcean GPU Droplets](https://saaskly.com/category/gpu-compute/digitalocean-gpu.md) | 3.8/5 | $4.41/h (NVIDIA H100 80 GB) | $3.26/h (12 mo), 26% off | $0.76/h (NVIDIA RTX 4000 Ada) | $35.28/h | Yes | None | — | Per second | $0.010/GB | $0.100 | Yes | Quota ticket | Own hardware, colocated | Yes | No | 2 | 16 | US-owned | DigitalOcean Holdings, Inc. (NYSE: DOCN) | — | Yes | Yes | No | Yes | Yes | Yes | Yes | 99.00 | 4.5 (2,572) |
| [Vultr Cloud GPU & Bare Metal](https://saaskly.com/category/gpu-compute/vultr-cloud-gpu.md) | 3.5/5 | $2.99/h (NVIDIA H100 80 GB (eight-GPU node; preemptible-only self-serve)) | Sales only | $0.06/h (NVIDIA A16 fraction (2 GB vGPU, 2 vCPU / 8 GB)) | $23.92/h | Yes | $120/mo (Intel E3-1270, 4 cores / 8 threads at 3.8 GHz, 32 GB RAM, 2 x 240 GB SSD, 5 TB bandwidth (not sold in the UK; cheapest UK bare metal is the Intel E-2286G in London at $185 a month)) | None | Per hour | $0.010/GB | $0.100 | Yes | Automatic limit raise | Own hardware, colocated | Yes | No | 7 | 33 | US-owned | The Constant Company, LLC | — | Yes | Yes | No | Yes | Yes | Yes | Yes | — | 1.6 (582) |
| [Hyperstack](https://saaskly.com/category/gpu-compute/hyperstack.md) | 3.4/5 | $3.20/h (NVIDIA H100 SXM 80 GB) | $2.72/h, 15% off | $0.15/h (NVIDIA RTX A4000) | Not on demand | Yes | None | — | Per minute | Included | $0.073 | Yes | None | Own hardware, colocated | No | No | 0 | 3 | UK-owned | Artemis 2 Limited | — | Yes | Yes | No | No | Yes | Yes | No | — | 2.5 (8) |
| [Hetzner Dedicated & GPU Servers](https://saaskly.com/category/gpu-compute/hetzner-dedicated.md) | 3.3/5 | €1.92/h (NVIDIA RTX PRO 6000 Blackwell 96 GB (GEX131; no H100 sold)) | None | €0.34/h (NVIDIA RTX PRO 4000 Blackwell SFF 24 GB (GEX45)) | Not on demand | No | €59/mo (AX41: AMD Ryzen 5 3600 (6 core), 64 GB DDR4, 2x 512 GB NVMe, unlimited traffic, 1 IPv4) | None | Per hour | Included | €0.003 | Yes | None | Own data centres | No | No | 3 | 3 | EU-owned | Hetzner Online GmbH (privately held, no listing) | — | No | No | No | Yes | Yes | No | No | 99.90 | 3.0 (2,971) |
| [Scaleway GPU & Bare Metal](https://saaskly.com/category/gpu-compute/scaleway-gpu.md) | 3.1/5 | €2.87/h (NVIDIA H100 PCIe 80 GB (H100-1-80G)) | None | €0.79/h (NVIDIA L4 24 GB) | €25.33/h | No | €27.99/mo (EM-A116X-SSD: Intel Xeon E3 1220 4C/4T, 32 GB RAM, 2 x 1.02 TB SSD, 500 Mbps (a RISC-V EM-RV1 server is cheaper at EUR 15.99 a month)) | None | Per minute | Included | €0.095 | Yes | Quota ticket | Own data centres | No | No | 3 | 3 | EU-owned | Iliad SA | — | Yes | Yes | No | Yes | Yes | Yes | No | — | 1.5 (351) |
| [Lambda](https://saaskly.com/category/gpu-compute/lambda.md) | 3.0/5 | $4.29/h (NVIDIA H100 SXM 80 GB) | Sales only | $0.69/h (NVIDIA Quadro RTX 6000 24 GB) | $31.92/h | No | None | — | Per minute | Included | $0.200 | Yes | None | Own hardware, colocated | No | No | 1 | 14 | US-owned | Independent | — | Yes | Yes | Yes | No | Yes | No | No | — | 2.6 (4) |
| [CoreWeave](https://saaskly.com/category/gpu-compute/coreweave.md) | 2.9/5 | $6.16/h (NVIDIA H100 80 GB SXM (sold only as an eight-GPU HGX node)) | Sales only | $1.25/h (NVIDIA L40 (sold only as an eight-GPU node)) | $49.24/h | Yes | $4,687/mo (AMD Genoa 9274F CPU-only node) | None | Per hour | Included | $0.070 | No | Quota ticket | Own hardware, colocated | Yes | No | 0 | 7 | US-owned | Independent | — | Yes | Yes | Yes | Yes | Yes | Yes | No | 99.90 | — |
| [Amazon EC2](https://saaskly.com/category/gpu-compute/amazon-ec2.md) | 2.7/5 | $8.94/h (NVIDIA H100 80 GB (p5.48xlarge, eight GPUs, London)) | $3.36/h (36 mo), 62% off | $1.02/h (NVIDIA L4 (g6.xlarge, London)) | $71.55/h | Yes | $3,539/mo (c5.metal (London), 96 vCPU / 192 GiB RAM, EBS only, 25 Gbit network, $4.848 an hour on demand) | None | Per second | $0.090/GB | $0.093 | Yes | Quota ticket | Own data centres | Yes | Yes | 6 | 39 | US-owned | Amazon | — | Yes | Yes | Yes | Yes | Yes | Yes | Yes | 99.99 | 1.3 (427) |
| [Nebius](https://saaskly.com/category/gpu-compute/nebius.md) | 2.6/5 | $3.85/h (NVIDIA H100 NVLink (SXM) 80 GB) | Sales only | $1.55/h (NVIDIA L40S PCIe) | $30.80/h | Yes | None | — | Per second | Included | $0.071 | Yes | None | Own data centres | Yes | No | 5 | 10 | EU-owned | Nebius Group N.V. | Spun out of Yandex N.V. in July 2024; founded and led by Yandex co-founder Arkady Volozh | Yes | Yes | Yes | No | Yes | Yes | Yes | 99.50 | 2.9 (2) |

* 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.

## The one we expected to win: Vultr

**Vultr Cloud GPU & Bare Metal (3.5)** is the cloud we run Saaskly on, and we expected it to win this category the way it won VPS & Cloud Servers. The research would not let it. On 22 September 2026 there was no GPU plan of any kind in London or Manchester; Frankfurt was the only European GPU region and carried two A16 fractions; and every multi-GPU node plan, H100, B200, MI325X and the rest, returned an empty list of available locations. The advertised $2.30 per H100 hour is the preemptible rate; the API says $2.99 on demand and marks the H100, B200 and MI325X nodes as not deployable on demand at all, so the self-serve route was interruptible only. Committed pricing is sales-only. What Vultr does have is the cheapest way onto a GPU on this page, a 2 GB slice of an A16 with 2 vCPU and 8 GB of RAM for $0.059 an hour, $43 a month, across fifteen fractional plans; CPU bare metal in both London and Manchester from $185 a month; $0.01 a gigabyte egress; and an official MCP server with 191 tools that is read-only unless you enable writes, alongside Terraform and a CLI. Trustpilot is 1.6 on 582 uninvited reviews, 69% one-star, with account suspensions and data lost on termination the themes. We would still run a web server there. We could not, on the day, run a GPU there in Britain. Partly first-party: we run ordinary Vultr servers and have not run its GPUs; the verdict on those rests on the research.

## The UK company with the foreign GPUs: Hyperstack

**Hyperstack (3.4)** is the only provider here besides Civo whose contract is with a UK company: NexGen Cloud Limited, London, company 12556681, whose controlling shareholder changed on 14 May 2026. The H100 SXM is $3.20 an hour on demand and "from $2.72" reserved, though the term behind that discount is not published anywhere; spot prices are published, which is rare, with the H100 PCIe at $2.00; the entry GPU is an RTX A4000 at $0.15 an hour; egress and ingress are free. Then the traps. The data centres are in Norway, Canada and Texas, so a UK contract buys no UK hardware. A stopped machine bills in full, and running out of credit does not stop the meter, it takes the balance negative, both by Hyperstack's own documentation and both the subject of the most recent one-star review. The SLA link in the footer returned a 404, ISO 27001 was promised for early 2026 and is still unevidenced, and the security page contradicts a support article on whether SOC 2 Type 2 exists. Trustpilot is 2.5 from eight reviews. Not yet tested by Saaskly.

## The dedicated-server benchmark: Hetzner

**Hetzner Dedicated & GPU Servers (3.3)** sells no H100, no H200 and no multi-GPU node, and it is on this page because it is the number every dedicated server in Europe is priced against. The AX41, a six-core Ryzen with 64 GB and two NVMe drives, is €59 a month with no setup fee, no minimum term, unlimited free traffic on a 1 Gbit port and 24/7 phone and email support at no charge. Dedicated is now billed hourly with a monthly cap that works out at 624 hours, so a weekend on the GEX45 GPU server is about €16 of compute. The GPU line is workstation cards: the GEX45 with an RTX PRO 4000 Blackwell at €214 a month plus a €209 setup fee, and the GEX131 with a 96 GB RTX PRO 6000 Blackwell at €1,199 plus €599 setup, quoted at 10 to 12 workdays to provision. Germany and Finland only, ISO 27001 and C5, no Kubernetes, no Slurm, no Terraform or CLI for the dedicated line, no MCP server. Trustpilot is 3.0 on 2,971 invited reviews. Not yet tested by Saaskly.

## The rest of the field

**Scaleway GPU & Bare Metal (3.1)** is the European sovereign option for GPUs: one H100 PCIe at €2.8665 an hour billed per minute, the L4 at €0.7875, bare metal from €27.99 a month or €0.077 an hour with no setup fee, a dedicated two-H100 server at €3,499.99 a month, hardware Scaleway owns in Paris and Warsaw, ISO 27001 and HDS, and a stated freedom from the US CLOUD Act. Against it: no committed rate, no reserved rate and no spot tier, so an H100 is €2,092.55 a month however long you stay; the availability API flagged every H100 type as scarce or in shortage on the day; new accounts need identity checks and a support ticket to hold more than one GPU; the eight-GPU node is priced at €25.3308 an hour but missing from the list of orderable machines; no UK region and no pounds; and Trustpilot is 1.5 on 351 reviews, 79% one-star. Not yet tested by Saaskly.

**Lambda (3.0)** is the ML team's default GPU cloud and the one that treats a small buyer worst. One H100 SXM is $4.29 an hour, and eight are $3.99 per GPU-hour, so the single card costs the most. There is no spot tier, reserved pricing is sales-only, and the cheapest 1-Click Cluster, 16 H100s for two weeks at $6.16 per GPU-hour, is about $33,100 prepaid. Egress is free and billing is per minute in arrears, and compliance is the strongest of the specialists: SOC 2 Type II plus ISO 27001, 27017, 27701 and 22301 with an evidence portal. But there is one EU region, in Germany, no UK region, no UK or EU contracting entity, no published SLA for the self-serve product, no stop state (you terminate or keep paying), filesystems that bill on after the instance is gone, and a signup that refuses debit and prepaid cards and flags VPN users as suspicious. Trustpilot is 2.6 from four reviews, all one-star, two of them describing data deleted after a payment failure. Not yet tested by Saaskly.

**CoreWeave (2.9)** is the platform the big labs rent, and it is not for sale to the reader of this page. There is no self-serve signup: CoreWeave's documentation says the sales team approves the organisation and sends the invitation before anyone can log in. The smallest H100 purchase is a whole eight-GPU node at $49.24 an hour, about $35,945 a month; the cheapest node of any kind, eight L40s at $10.00 an hour, is $7,300 a month. What it has is genuine: two live UK data centres, Crawley since October 2024 and London Docklands since December 2024, a UK contracting entity, free egress, published spot at about 60% off, a 99.9% multi-region SLA, Kubernetes, Slurm and InfiniBand. Committed pricing is "up to 60%" with no terms published, and the certification reports sit behind an NDA. Trustpilot has no reviews. Not yet tested by Saaskly.

## The two we marked down

**Amazon EC2 (2.7)** has the H100 in London, which only Civo can also say, and every GPU, HPC and bare-metal instance type there is. It is marked down for the reason we moved our own clients off it: the bill. A London H100 hour is $8.944, the p5.48xlarge divided by eight, and the only way to bring it down is a three-year all-upfront reservation of $707,025, which works out at $3.36 per GPU-hour. One GPU instance is billed on nine meters: instance hours, EBS capacity, IOPS and throughput, snapshots, internet egress at $0.09 a gigabyte after the first hundred, cross-zone and inter-region transfer, and the public IPv4, with a support plan charged as a percentage on top; a g6.xlarge with a 100 GB disk, an IPv4 and 500 GB of egress comes to $794.70 in a month, not the $745.77 on the instance line. A new account has a GPU quota of zero vCPUs and AWS publishes no waiting time for the request that lifts it. London costs 27% to 30% more than Virginia for the same hardware, and H200, B200 and L40S are not sold there. There is no hard spending cap. Trustpilot for AWS is 1.3 on 427 reviews. First-party on the negative side: we ran client workloads on EC2 and moved them to Vultr because the bill was hard to work out in advance and carried unexpected charges driven by usage and bandwidth.

**Nebius (2.6)** would rank near the top on the platform alone, and its profile says so: per-second billing with no compute charge on a stopped machine, free egress, a default quota of 12 GPUs for a new card-paying account, free support with a 30-minute first-response target on critical tickets, a published 99.50% SLA, SOC 2 Type II and ISO 27001, Kubernetes, Slurm, InfiniBand, Terraform, a CLI and an official MCP server, and UK regions live since November 2025, though the H100 itself is in Finland. It is last because of where it comes from. Nebius Group N.V. is the former Yandex N.V., the Dutch parent of the Russian search company; it sold its Russian businesses in July 2024, its chairman said "All connections with Russia have now been severed", and it relisted on Nasdaq that October. Its founder and chief executive is Yandex co-founder Arkady Volozh, on the EU sanctions list from June 2022 to March 2024, and about a thousand of its engineers came across from Yandex. Our rule is that a vendor with Russian links is listed, the facts are stated, and the score is marked down hard. The facts are stated in the profile, without adjectives. The H100 is $3.85 an hour today and $4.50 from 1 October 2026. Not yet tested by Saaskly.

## Bottom line

- **Price the hour, then price the stop.** One H100 on demand is £2.41 at OVHcloud, €2.87 at Scaleway, $2.99 at Civo and at Vultr (interruptible only), $3.20 at Hyperstack, $3.49 at RunPod, $3.85 at Nebius, $4.29 at Lambda, $4.41 at DigitalOcean, $6.16 at CoreWeave (eight at a time) and $8.94 on Amazon EC2 in London. Then ask what a stopped machine costs: everything at Hyperstack, Civo, DigitalOcean, Vultr and Lambda; the disk only at Nebius and Amazon.
- **Ask if you can have it today.** RunPod, Nebius, Hyperstack and Lambda will launch the reference GPU for a new card-paying account. OVHcloud, DigitalOcean, Civo and Scaleway raise quotas by request. Amazon starts you at zero. CoreWeave will not open the door without its sales team.
- **Ask where it is.** Only Civo and Amazon EC2 had an H100 in a British data centre on the day. Vultr had no GPU in the UK at all; DigitalOcean's London region has none; OVHcloud's is in Gravelines.
- **OVHcloud** for the H100 in pounds with free egress, no setup fees and an EU company on the contract, once your quota is raised. **RunPod** for a GPU this afternoon, per second, if a US contract and no UK region are acceptable. **Civo** for a UK company and a London H100 with published committed terms, eyes open on the stopped-instance bill.
- **Hetzner** for the cheapest dedicated server in Europe at €59 with free traffic and phone support, not for GPUs above one card. **Vultr** for fractional GPUs from $43 a month and UK bare metal, not, today, for an H100.
- **Read the origin and the first-party lines.** Nebius is last under our origin rule with the facts stated; Amazon EC2 is marked down because we have paid its bills.

*Prices checked 22 September 2026 on each provider's UK-facing pages or public price APIs, in the currency shown to a UK visitor: GBP ex VAT for OVHcloud; EUR ex VAT for Scaleway and Hetzner; USD for RunPod, Civo, DigitalOcean, Vultr, Hyperstack, Lambda, CoreWeave, Amazon EC2 and Nebius, with the VAT basis stated per profile where the vendor states one. Monthly equivalents use 730 hours. Trustpilot scores are from uk.trustpilot.com on the same date; where a profile has a handful of reviews (Civo, Hyperstack, Lambda, Nebius) or none (CoreWeave) we say so. Vultr is partly first-party (ordinary servers, not GPUs); Amazon EC2 is first-party on the negative side; the rest are labelled "Not yet tested by Saaskly". Saaskly may earn a commission from some links on this page, each disclosed on the profile; no one on this page is sponsored, and commercial relationships never affect scores.*

---

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/comparisons/best-gpu-cloud-high-performance-compute-2026 · Markdown index: https://saaskly.com/llms.txt