AMS
Operational
Amsterdam
Netherlands
Infrastructure
The network, described precisely.
Bandwidth is where hosting pages are vaguest, so this one is specific: what each port is, what it costs when you saturate it, who we buy transit from, and what we do not offer.
- 25 Gbit/son an 8-GPU node, unmetered
- $0overage, at any volume
- /64IPv6 routed to every machine
- 6sites, each on two carriers plus an IX
Port speed and traffic
The port is sized with the chassis, because a node that can saturate a card can saturate a link. Every port is unmetered in both directions: there is no included volume, no overage tier and nothing to watch on a graph.
| Node | Port | Monthly traffic | Upgrade | Cost of the upgrade |
|---|---|---|---|---|
| 1 × GPU | 1 Gbit/s | Unmetered | 10 Gbit/s | $49/mo |
| 2 × GPU | 2 Gbit/s | Unmetered | 10 Gbit/s | $49/mo |
| 4 × GPU | 10 Gbit/s | Unmetered | Already 10 Gbit/s or above | — |
| 8 × GPU | 25 Gbit/s | Unmetered | Already 10 Gbit/s or above | — |
| 10 × GPU | 25 Gbit/s | Unmetered | Already 10 Gbit/s or above | — |
Unmetered means what it says: you may hold the port at line rate for the whole month. We reserve the right to talk to you about it only if a single machine starts degrading the shared uplink for everybody else in the suite — and that conversation has happened twice in five years, both times about a misconfigured mirror.
Transit and peering
Every site buys from at least two independent carriers and sits on the local internet exchange. A carrier failure is a re-route, not an outage — which is the only reason we are willing to write 99.9% on a contract.
| Data centre | Internet exchange | Transit carriers | Uplink |
|---|---|---|---|
| Amsterdam | AMS-IX, NL-ix | Cogent, Arelion, Lumen | 2 × 100 Gbit/s |
| Stockholm | Netnod, STHIX | Arelion, GlobalConnect | 2 × 100 Gbit/s |
| Zürich | SwissIX, CIXP | Init7, Cogent | 2 × 100 Gbit/s |
| Reykjavík | RIX Reykjavík | Farice, Síminn | 2 × 100 Gbit/s |
| Montréal | QIX Montréal | Cogent, Zayo | 2 × 100 Gbit/s |
| Singapore | SGIX, Equinix Singapore | Arelion, Singtel | 2 × 100 Gbit/s |
Addressing
IPv4
One address, included
Routed to the machine, yours for the life of the term. Additional addresses are $3/month each, up to eight.
IPv6
A /64, included
Routed, not bridged. That is 18 quintillion addresses per machine, at no cost, and we will never ask you to justify one.
Reverse DNS
Yours to set
Forward and reverse records on every address you hold, changed by you, applied within a minute. No ticket.
Address history
Clean on delivery
An address that has been listed anywhere is held out of the pool until it is delisted. You do not inherit someone else's reputation.
Outbound port 25 is closed by default.
Not to inconvenience you — a GPU rental with open SMTP becomes a spam relay within days, and the whole range gets listed. If you have a genuine reason to send mail from the machine, ask us and we will open it on that address.
DDoS filtering
Volumetric filtering is on from the moment the machine is delivered. It is not an option, it is not a tier, and it does not appear on the invoice.
Always on, L3 and L4
Scrubbing at the edge of each site, in front of our own routers. Detection is automatic on rate and signature; nothing waits for a human.
Mitigation in seconds
Under ten seconds from the first anomalous packet to filtered traffic, up to the site's full transit capacity. Legitimate sessions are not reset.
No null-routing your address
Some hosts answer an attack by dropping your IP for 24 hours. We filter it. Your machine stays reachable, which is the entire point of paying for filtering.
GPU interconnect
Inside a node, what connects the cards to each other decides whether a large model is fast or merely possible. This is the part of the specification we are asked about most, and the part most listings leave out.
| Form factor | Cards | Between GPUs | Practical effect |
|---|---|---|---|
| SXM on an HGX board | A100 SXM4 80 GB, H100 SXM5 80 GB, H200 SXM5 141 GB, B200 SXM6 180 GB | NVLink via NVSwitch | Tensor parallelism across the whole board without the interconnect becoming the bottleneck. |
| PCIe cards | L4 24 GB, RTX 4090 24 GB, RTX 5090 32 GB, RTX A6000 48 GB, L40S 48 GB, A100 PCIe 40 GB… | PCIe Gen4 ×16 or Gen5 ×16 | Fine for pipeline parallelism, batch inference and one-model-per-card serving. Slower than NVLink for tensor parallelism across many cards. |
If you are unsure which of the two your workload needs: a model that fits on one card never needs NVLink, and a model split across four or more cards almost always benefits from it. The fit checker in the configurator tells you which case you are in before you pay.
Latency between sites
Measured from our own routers to the named cities, as a round trip, over thirty days. They are medians, not best cases.
STO
Operational
Stockholm
Sweden
ZRH
Operational
Zürich
Switzerland
REY
Operational
Reykjavík
Iceland
YUL
Operational
Montréal
Canada
SIN
Operational
Singapore
Singapore
What we do not offer
Stated here rather than discovered after payment. If one of these is a requirement, we are the wrong provider and it is cheaper for both of us to know now.
Multi-node InfiniBand. NVLink connects the cards inside one node. We do not sell a fabric between nodes, so a job that needs 32 GPUs in one training run is not something we can serve.
BGP sessions or bring-your-own-IP. We announce our own space only. There is no way to advertise your prefix from our routers.
Layer-2 links between data centres. A private VLAN joins your machines within one site. Between sites, you build the tunnel yourself.
Committed transit or a private uplink. The port is unmetered and shared upstream. We will not sign a dedicated-capacity contract on top of it.
Open SMTP on delivery. Port 25 outbound is closed until you ask, for the reason given above.
An anonymity network. Filtering protects your machine from attack. It is not a proxy, it does not hide where your traffic goes, and we do not pretend otherwise.
Every port above is on every machine.
There is no networking tier to choose. Pick the GPU, pick the site, and the port comes with the chassis.
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.
- Service statusLive status of every gpuserver.io component and data centre, probed every 60 seconds from five cities, with 90-day uptime and the incident log since 2022.
- Service level agreementThe 99.9% commitment: what counts as downtime, how it is measured from outside, five minutes of term back per minute lost, and the exclusions in full.
- DocumentationFirst SSH, verifying the hardware, CUDA containers, serving a model with vLLM, RAID, firewalling, IPMI and reinstalls — the commands, on a real machine.