Amazon EC2

Who owns it

EC2 is part of Amazon Web Services, a segment of Amazon.com, Inc., listed on NASDAQ as AMZN since May 1997 and run by Andy Jassy, with Matt Garman as chief executive of AWS since June 2024. EC2 launched in 2006. A UK or EU customer contracts with Amazon Web Services EMEA SARL in Luxembourg, a subsidiary of the US parent, which is the point that decides jurisdiction: the US CLOUD Act reaches the ultimate owner regardless of where the instance runs. AWS designs, buys and operates its own servers in its own and leased data centres and designs its own silicon, including Nitro, Graviton, Trainium and Inferentia. Its GPUs are bought from NVIDIA and run by AWS. There is no resold capacity and no marketplace of third-party hosts. AWS launched a separate European Sovereign Cloud in January 2026 with a first region in Brandenburg, Germany, a new European parent company and three German subsidiaries, independent root certificates and networking, an independent security operations centre and EU-resident staff only, with 90 services available on day one. Whether a GPU instance is among them is not something AWS has published.

## What it costs

Prices are in US dollars, exclusive of VAT, which Amazon Web Services EMEA SARL adds to the invoice, and were read on 22 September 2026 from the AWS price list, which is public and needs no account. In London, an H100 comes only inside a p5.48xlarge: eight H100 80 GB, 192 vCPU, 2,048 GiB of RAM, eight 3.84 TB NVMe drives and 3,200 Gbit of network for $71.552 an hour, which is $8.944 per H100-hour and $52,233 a month at 730 hours. A three-year all-upfront reserved instance costs $707,025, and that works out at $26.90 an hour for the node across 26,280 hours, or $3.36 per H100-hour, 62% below the on-demand rate. There is no one-year reserved option for p5 in London, only three-year. Savings Plans run one or three years and are advertised at up to 66% for Compute and up to 72% for EC2 Instance plans. Capacity Blocks let you book P4d, P5, P5e, P5en, P6-B200, P6-B300 and P6e-GB200 for up to six months in clusters of one to 64 instances, up to eight weeks ahead, but the prices "are dynamic and change based on available supply and demand" and appear in the price list only as zero placeholders, so committed GPU capacity cannot be compared before you are inside the console. Spot exists for G and P families, and its prices float and are not published.

The cheapest GPU in London is a g6.xlarge at $1.0216 an hour: one NVIDIA L4 24 GB with 4 vCPU, 16 GiB of RAM and a 250 GB NVMe drive. There is no fractional GPU on EC2. The instance hour is only the first of nine meters. In London you also pay $0.0928 per GB-month for gp3 storage, $0.0058 per provisioned IOPS-month and $0.0464 per provisioned MiBps-month above the free gp3 baseline, $0.053 per GB-month for snapshots, $0.09 a GB for the first 10 TB of internet egress each month after a 100 GB free allowance (falling to $0.085, $0.07 and $0.05 at 10 TB, 50 TB and 150 TB), $0.01 a GB for traffic between availability zones, $0.02 a GB out to any other AWS region, and $0.005 an hour for every public IPv4 address whether it is in use or idle. On top of that, a support plan is charged as a percentage of the bill: Business Support+ from $29 a month or 9% of spend falling to 3%, Enterprise from $5,000 a month or 10% falling to 3%. A worked month for the smallest sensible setup, a g6.xlarge with a 100 GB gp3 disk, one public IPv4 and 500 GB of egress: 730 hours at $1.0216 is $745.77, plus 100 GB of gp3 at $0.0928 is $9.28, plus 730 hours of IPv4 at $0.005 is $3.65, plus 400 chargeable GB of egress at $0.09 is $36.00, giving $794.70 before VAT and before any support plan. Billing is per second with a 60-second minimum on Linux, and stopping the instance stops the compute charge but not the storage, the snapshots or the IPv4. Billing is postpaid with no hard spending cap: AWS Budgets raises an alert after the money has been spent, and a p5.48xlarge left running costs $1,717 a day.

## What you get and what to watch

A new AWS account cannot launch a single GPU. The default On-Demand quota is 0 vCPUs for the G and VT families, 0 for P, 0 for HPC, and the Spot quotas for G, VT and P are 0 as well, against 5 vCPUs for the standard families. Even the $1.02 an hour g6.xlarge needs four vCPUs of G quota that the account does not have. The fix is a request in the Service Quotas console, which becomes a Support case with a case number, and AWS publishes no turnaround time for it anywhere in the documentation. On a free Basic support plan that case sits in the standard queue with no response commitment. Plan the request days before you need the hardware, not on the morning you want to train something.

London, eu-west-2, is the only UK region, and it carries H100 in the p5.48xlarge, A100 40 GB in the p4d.24xlarge, L4 in the g6 family and A10G in the g5 family. It does not carry H200, B200 or L40S: the p5en, p6-b200 and g6e families are in us-east-1 and elsewhere, and neither hpc7a nor hpc6id appeared in the London or N. Virginia price lists. Keeping the workload in the UK also costs more, between 27% and 30% more than us-east-1 for identical hardware: g6.xlarge is $1.0216 against $0.8048, p5.48xlarge is $71.552 against $55.04. The European regions are Ireland, Paris, Frankfurt, Stockholm, Milan and Spain, with Zurich outside the EU and the sovereign region in Brandenburg. Bare metal is available: a c5.metal in London, 96 vCPU and 192 GiB, is $4.848 an hour, which is $3,539 a month at 730 hours, on demand with no minimum term. For clusters, EFA gives OS-bypass networking optimised for NCCL at no extra charge, AWS ParallelCluster deploys Slurm or AWS Batch clusters and is free beyond the resources it starts, and EKS runs GPU node groups. Terraform, a mature CLI, a public API and AWS's own MCP servers are all there.

AWS holds ISO 27001, ISO 27017, ISO 27018, SOC 1, SOC 2, SOC 3 and PCI DSS globally, plus C5 in Germany, HDS in France and UK Cyber Essentials Plus, G-Cloud and PASF. The published EC2 uptime commitment is 99.99% at region level with service credits. Support is 24x7 by phone, chat and ticket on the paid plans; Basic is free but covers billing and account questions only, not technical help. Trustpilot rates aws.amazon.com 1.3 out of 5 from 427 reviews, 75% of them one star, on a profile AWS has never claimed and with no history of inviting reviews, so treat it as a complaint channel rather than a sample of a customer base measured in millions. The dated one-star themes in September 2026 are overcharging with refunds refused, accounts suspended without notice, more than 24 hours without a support response during a suspension, account closure requiring every service to be disabled by hand, and unauthorised access to a closed account's linked databases. One pricing change worth remembering: since February 2024 AWS charges for every public IPv4 address, in use or idle, which quietly added $3.65 a month per address to bills that had not changed.

First-party on the negative side: we have run 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. Saaskly has no commercial relationship with AWS.