Vultr Cloud GPU & Bare Metal

Who owns it

Vultr is the cloud product of The Constant Company, LLC, a privately held US company based in West Palm Beach, Florida, founded in 2014 by David Aninowsky, who is still chief executive. The parent brand is presented as Constant. There is no UK or EU contracting entity, so a London or Manchester customer signs with the US LLC even when the server sits in Britain, and Vultr publishes no position on the US CLOUD Act. Vultr owns its own servers and runs them in leased space in third-party data centres across 33 regions. It does not resell another cloud's capacity and does not run a marketplace of third-party hosts, which puts it in a different category from the community-GPU platforms. Its own words on the dedicated line: bare metal is "single-tenant, which means that only you have access to the resources of your server". A Sovereign Cloud and Private Cloud offering was announced in April 2024 for nations and enterprises, but it is sold through sales rather than the console.

## What it costs

Prices are shown to a UK visitor in US dollars, with no VAT basis stated on the pricing page, and were read on 22 September 2026. The reference GPU, an NVIDIA H100 80 GB SXM, is sold only inside an eight-GPU node: 216 vCPU, 1,914 GB of RAM, 13 TB of NVMe and 15 TB of bandwidth. Vultr's API gives that node an on-demand rate of $23.92 an hour, which is $2.99 per GPU-hour, or $17,462 a month at 730 hours. The pricing page advertises $2.300 per GPU-hour, and that figure is the preemptible rate, $18.40 an hour for the node. Vultr's own API marks the H100, B200 and MI325X node plans as preemptible-only for self-serve deployment, so the advertised price is the interruptible price and the words spot, preemptible and interruptible appear nowhere in the public copy. Reserved and contract rates are not published at all: "Contact us for additional contract duration and payment options."

The cheapest way in is far cheaper and genuinely useful. A 2 GB slice of an NVIDIA A16 with 2 vCPU, 8 GB of RAM, 50 GB of NVMe and 1 TB of bandwidth is $0.059 an hour, $43 a month. Fifteen fractional plans split an A16, an A40 or an L40S by video memory, from 2 GB up to 192 GB. Whole-GPU rates run from $1.671 per GPU-hour for an L40S to $2.800 for an A100 80 GB SXM, with AMD at $1.850 for an MI300X and $2.000 for an MI325X. Block storage is $1 a month per 10 GB, object storage is $18 a month for a terabyte of capacity and a terabyte of transfer, snapshots are $0.05 per GB a month and bandwidth beyond the generous instance allowance is $0.01 a GB. Billing is hourly, capped at the monthly price, with no per-second option, and a stopped instance continues to reserve and bill for its resources. Accounts are prepaid credit or postpaid card, self-serve. For the small buyer, the realistic monthly floor is the A16 slice at $43 plus 100 GB of block storage at $10, so around $53 a month before any bandwidth overage, and the realistic H100 floor is the whole node at $17,462.

## What you get and what to watch

The thing to settle before anything else: on 22 September 2026 there was no GPU plan of any kind available in London or Manchester. Frankfurt was the only European region with GPU stock, and it carried just two A16 fractions, the 2 GB and the 4 GB. Every multi-GPU node plan, H100, B200, A100, L40S, MI300X, MI325X, MI355X and GH200, returned an empty list of available locations, as did 11 of the 15 fractional plans. The UK regions are real and well stocked for CPU work: London and Manchester both run cloud compute, block storage, load balancers and managed Kubernetes, and London carries most of the bare metal range from $185 a month for a six-core Xeon E-2286G up to $1,450 a month for a 32-core EPYC 9354P. The $120 a month entry server, an Intel E3-1270 with 32 GB of RAM and 5 TB of bandwidth, is sold in the United States and Amsterdam but not in Britain, so the cheapest UK bare metal is the $185 machine in London.

Signing up is self-serve with a card and no sales call, but new accounts carry a spend and instance limit that will block a higher-cost GPU plan until it is raised. The increase is requested inside the console: confirm your profile, state your organisation size, describe your use case and sign to confirm it. Vultr publishes no turnaround time for that request. The platform itself is broad. Eight-GPU nodes come with 8 x 400 Gbps of GPU network and a 25 to 200 Gbps front end, although Vultr does not name the fabric type for its NVIDIA nodes. Vultr Clusters creates on-demand clusters of bare metal servers "without manual requests or reservations". There is managed Kubernetes in both UK regions, though whether it supports GPU node pools is not documented anywhere we could read. Developers get a public API, an official Terraform provider, a CLI and an official MCP server at github.com/vultr/vultr-mcp with 191 tools across 33 product areas, read-only unless writes are explicitly enabled, which is further than most infrastructure vendors have gone.

Vultr's compliance material lists SOC 1 Type II, SOC 2 Type II, ISO 27001 and PCI DSS, and says its data centre processes have been reviewed for GDPR. The compliance page blocks automated reading, so that list is the company's own account and has not been read from the page itself. There is no C5 and no SecNumCloud. Support is ticket-based through the customer portal with a community forum, a docs site and a public status page; the legal and support pages also block automated reading, so the published uptime commitment, the service credits and the paid support tiers could not be captured. Trustpilot is the loudest signal on the profile: 1.6 out of 5 from 582 reviews, 69% of them one star, with no history of Vultr inviting reviews, on a profile claimed in October 2024. The dated one-star themes through July and August 2026 are accounts suspended without notice shortly after a deposit, an endless Cloudflare CAPTCHA loop locking users out, charges continuing after cancellation, urgent tickets going unanswered, and production servers and data lost after an unexpected termination with no recovery window.

Partly first-party: Saaskly and Fractional Teams run their ordinary servers on Vultr and have not run its GPUs or bare metal; the GPU verdict rests on published pricing, vendor documentation and third-party reviews. Saaskly may earn a commission if you buy through the link on this page; the verdict came first and the link does not affect the score.