We own the machines we rent you.
gpuserver.io is a small, independent operator. We buy the cards, we rack them in leased suites, and we rent them by the month. Nothing here is resold from someone else's cloud, which is the only reason we can promise that the specification on a row is the specification you get.
- 6data centres, all leased suites
- 12GPU models, every one bought new
- 41configurations, all available
- 0cards resold from another cloud
Renting a GPU had become the most complicated way to buy compute.
We started this because we were customers first, and because every option in front of us was wrong in the same three ways.
The hyperscalers wanted a company registration, a credit check and a quota request before they would let us near an A100 — and then billed by the second for a machine we intended to keep for a year. The marketplaces were cheaper, but the card was someone's spare desktop behind a residential line, and the node vanished mid-run. The specialist clouds were honest and good, and priced as though a GPU were a scarce metal.
None of them would simply rent us a machine for a month, take payment, and hand over root. So we bought the hardware and did that.
What we decided, in order
And in that order, because each one constrains the next
Owning the cards is what makes the other three possible. A reseller cannot promise a specification it does not control, and cannot refuse a KYC requirement imposed by whoever it buys from.
Four things we do not bend, including when it costs us.
Every one of these has already lost us a sale. They are on this page so that you can hold us to them.
No identity, ever
Not at signup, not at renewal, and not when the spend gets large. There is no verification tier: an eight-GPU node opens on exactly the same terms as the cheapest single card.
- No document, no selfie, no phone number
- Nothing you type is checked against anything
- No threshold that triggers a review
The price is the price
One number per configuration per term. No setup fee, no reinstall fee, no bandwidth overage, no support plan, and no discount that appears only if you ask.
- Nothing metered after the invoice
- The same price whoever you are
- Cancel a monthly term with no penalty
The listing is the machine
If we cannot deliver the exact card on the row, we tell you and refund the term. We do not ship a near-equivalent and hope the difference goes unnoticed.
- One tenant per physical card
- No MIG, no vGPU, no hypervisor
- Serial numbers recorded per chassis
We hold as little as we can
We keep what the service creates and nothing else, because data we do not have is data nobody can compel us to produce. Disks are erased between tenants, without being asked.
- No login inside your operating system
- No backup of your disks
- Nothing retained past its stated life
One suite at a time.
We have never raised money and we have never bought capacity we could not fill. Each site below was opened out of the previous one's revenue, which is slower than the alternative and the reason we are still here.
2021 · AMS
Amsterdam, Netherlands
Tier III, 100% wind. The first suite, and still the largest. Everything we learned about burn-in was learned here, mostly by getting it wrong.
2022 · STO
Stockholm, Sweden
Tier III, 100% hydro. Opened on the same pattern as the others: two carriers, a public exchange, on-site spares, and no machine sold before it has survived three days of load.
2023 · ZRH
Zürich, Switzerland
Tier IV, 96% carbon-free. Opened on the same pattern as the others: two carriers, a public exchange, on-site spares, and no machine sold before it has survived three days of load.
2023 · YUL
Montréal, Canada
Tier III, 99% hydro. Opened on the same pattern as the others: two carriers, a public exchange, on-site spares, and no machine sold before it has survived three days of load.
2024 · REY
Reykjavík, Iceland
Tier III, 100% geothermal + hydro. Opened on the same pattern as the others: two carriers, a public exchange, on-site spares, and no machine sold before it has survived three days of load.
2025 · SIN
Singapore, Singapore
Tier III, Grid + REC offset. Opened on the same pattern as the others: two carriers, a public exchange, on-site spares, and no machine sold before it has survived three days of load.
What we are not.
Half of choosing a provider is ruling one out quickly. Here is everything that should rule us out.
Not a cloud. There is no console, no autoscaling, no managed database and no object store. You get a machine and root on it.
Not a marketplace. Nobody lists their spare desktop here. Every card is ours, in a data centre, on a leased-line uplink.
Not hourly. If you need a card for two hours a week, we are the expensive option and you should use a per-second provider.
Not a training cluster. NVLink joins the cards inside a node. We do not sell a fabric between nodes, so a job needing 32 GPUs in one run is not for us.
Not an anonymity service. No identity check is not the same as no trace. Your machine has a public address and your traffic goes where you send it.
Not lawless. We do not ask who you are, which makes it more important that we act on what a machine does. The acceptable use policy is short and it is enforced.
The person who replies is the person with the spare card.
There is no first line and no ticket tier here, because there are not enough of us to have one. A message about a throttling GPU reaches someone who can walk to the rack.
That has a cost, and we would rather state it than dress it up: we are small. We do not staff a phone line, we do not offer a named account manager, and if you need a procurement process with a signed master agreement, we are not set up for it.
What we do offer is that the answer you get is from someone who knows the machine, and that it usually arrives within the hour.
Engineering, not support
Everyone who answers a message also racks hardware. That is why an answer about NVLink topology is specific rather than a link to a knowledge base.
Around the clock, honestly
Hardware failures are handled at any hour, with a four-hour target. Questions that are not urgent get answered during waking hours in Europe, and we would rather say so than pretend.
In writing, in English
One channel, one thread, no phone tag. Everything we tell you about your machine is written down where you can go back to it.
The machines are already racked.
Every configuration in the catalogue is powered, burned in and waiting in all 6 suites. Nothing is built on demand, which is why root arrives in under 5 minutes.
Related pages
- Hardware policyThe twelve NVIDIA GPUs we operate, the 72-hour burn-in before a node is sold, the four-hour replacement target, and how disks are erased between tenants.
- NetworkUnmetered ports up to 25 Gbit/s, two carriers and an IX per site, always-on DDoS filtering, routed IPv6 — and the four things we do not offer, stated up front.
- Acceptable useWhat you may and may not run on a gpuserver.io machine, what is explicitly allowed, what to ask about first, and the procedure when a machine crosses the line.
- GuidesEight practical guides: VRAM sizing, NVLink versus PCIe, image and video models, monthly versus hourly, self-hosting versus an API, serving Llama 70B, and QLoRA.