100 Gbps guaranteed bandwidth

100Gbps Dedicated Servers: The Hardware Is the Easy Half

A guaranteed 100 Gbps port is the easy half of this purchase. Making a general-purpose Linux server saturate one is a memory placement and interrupt routing job, and that work is now yours.

X-ZoneServers 100 Gbps dedicated servers start at €4499 per month for a guaranteed 100 Gbps port — 3 bare-metal configurations with up to 32.4 PB of monthly transfer across 12 EU and US datacenter locations. The port is effectively unmetered: hold the full committed rate around the clock, with no overage charges and no fair-use throttling inside your tier.

4499
starting price per month
44.99
per Gbps per month
3
configurations
32.4 PB
transfer ceiling per month

Bandwidth is the cheap part of this tier. What it costs to move a packet has never been lower on our price list, and the chassis we install will accept a 100 Gbps port without complaint. The awkward truth is that no mainstream operating system ships ready to fill one. A default Linux install, with its interrupts landing wherever the scheduler feels like putting them and its socket buffers sized for a link from a decade ago, will typically stall somewhere in the tens of gigabits and stay there no matter how much CPU you throw at it. Reaching line rate is an exercise in where memory sits and which core handles the interrupt, long before it is an exercise in raw core count. Three configurations are offered here, and the right one depends less on the port, which is identical across all three, than on what your software does with each packet after it arrives.

What the port carries once the stack cooperates

At line rate the port moves 12500 MB/s of payload. That puts a 1 TB dataset across the wire in 1 minute 20 seconds and adds up to 32.4 PB over a 30-day month at full utilisation. In streaming terms it is 20000 simultaneous 1080p viewers at 5 Mbps, or 4000 viewers at 4K and 25 Mbps. Those figures describe the port, and the port is guaranteed, so nothing on our side throttles or charges extra within the tier. What arrives at your application is a separate question, decided by how the receive queues are spread across cores, which NUMA node holds the buffer a packet lands in, and how many times that packet is copied before your process sees it. The rest of this page is about closing that difference.

32.4 PB
Transfer per month
Sustained volume at full port utilisation across a 30-day month.
20,000
1080p viewers
Simultaneous streams at 5 Mbps each.
4,000
4K viewers
Simultaneous streams at 25 Mbps each.
1 min 20 s
Per terabyte
Wall-clock time to move 1 TB off the server at line rate.

Figures use decimal units (1 Gbps = 125 MB/s) across a 30-day month, at 5 Mbps per 1080p stream and 25 Mbps per 4K stream. They describe what the port can carry, not a guarantee of application throughput: disk, CPU and the far end all apply their own limits.

3 configurations at 100Gbps

The Dual Xeon Gold 6148 at €4499 brings 40 cores and 80 threads at 2.4GHz, 256GB of DDR4-2133 and six SSD bays. The Dual EPYC 9554 at €7379 steps up to 128 cores and 256 threads at 3.1GHz, 512GB of DDR5-4800 and 24 NVMe slots. The Dual EPYC 9754 at €7939 is the ULTIMATE build with 256 cores, 512 threads and 1024GB of DDR5-4800. The EPYC 9254 chassis are absent because their port options stop at 50Gbps. Memory generation is the specification to weigh here rather than clock speed, because every copy in the packet path costs both a read and a write, so the memory bandwidth you need is a multiple of the wire rate and not equal to it. All three builds are dual socket, which means two NUMA nodes and a network card physically attached to the lanes of only one of them. That single fact drives most of the tuning described below, whichever chassis you pick.

Dual Xeon Gold 6148

Ultimate Performance

4499
/month
40c / 80t @ 2.4GHz256GB DDR4 @ 2133MHz
6x SSD Slots100Gbps Guaranteed
AI Optimized Routing Included
Order Now

Dual EPYC 9554

Ultimate Performance

7379
/month
128c / 256t @ 3.1GHz512GB DDR5 @ 4800MHz
24x NVMe Slots100Gbps Guaranteed
AI Optimized Routing Included
Order Now
ULTIMATE

Dual EPYC 9754

Maximum Performance

7939
/month
256c / 512t @ 2.3GHz1024GB DDR5 @ 4800MHz
24x NVMe Slots100Gbps Guaranteed
AI Optimized Routing Included
Order Now

What a 100Gbps server is for

Cache and origin fleets serving large objects

The one shape of traffic that gets close to line rate with modest effort, because the kernel can hand whole files to the card with few copies. Spend your tuning budget on connection setup and TLS termination rather than on the payload path.

Packet inspection, scrubbing and traffic analytics

Anything that examines every packet rather than every file runs into per-packet kernel cost long before it runs into the port. Plan for a bypass framework and dedicated poll mode cores from the first day, not as a later optimisation.

NVMe storage targets and backup landing zones

NVMe-oF and object backends only hold their throughput when drive queues, network queues and memory all sit on the same NUMA node, which is why the 24-slot EPYC chassis are the sensible starting point for this role.

Dataset staging between sites and clusters

Moving training corpora or research archives between locations turns a working day into a coffee break, but only if the transfer tool is NUMA aware and does not funnel every byte through a single core near the wrong socket.

€44.99 per gigabit, and what the saving still costs you

At €4499 a month for a guaranteed 100 Gbps port, each gigabit costs €44.99. That is the lowest bandwidth price we publish, and it is a floor rather than a point on a straight line, which is why this tier is worth reaching for even if you expect to run it at half load through the first year. The saving is real, but it is not free money. Assume some engineer time for stack tuning, and possibly a licence or two if you take the bypass route, before the port earns the figure on the invoice. Budget it honestly and the arithmetic still works: the tuning is a one-off cost paid in days, while the price per gigabit is paid every month for as long as the server runs. What you should not do is buy this tier, install a stock image, measure 30 Gbps and conclude that the port is not there.

All guaranteed bandwidth tiers

The same hardware range is sold at nine committed port speeds. Compare the entry price of each.

100Gbps dedicated server questions

What does a server actually need in software to reach 100 Gbps?
Considerably more than a default install. In practice: a current kernel, receive side scaling with one queue per handling core, interrupt affinity pinned to cores on the network card's own socket, ring buffers and socket buffers raised well above their defaults, huge pages backing the buffer pool, and the segmentation and checksum offloads left switched on so the card does that work instead of a core. None of it is exotic and all of it is deliberate. A stock image has none of it configured for this rate, because the defaults are chosen for a machine with a 1 Gbps link.
Why does NUMA placement matter so much at this rate?
Every chassis here has two sockets, so it has two NUMA nodes, and the network card is attached to the PCIe lanes of exactly one of them. If the interrupt is handled on one node while the packet buffer and your worker thread live on the other, each packet makes extra trips across the socket interconnect and you pay for them in latency and in memory bandwidth you did not budget for. Pin the receive queues, the buffer allocations and the application threads to the node the card sits on, then verify the placement with numactl instead of assuming the scheduler got it right.
Can the Dual Xeon Gold 6148 drive this port as well as the EPYC builds?
For byte shaped traffic, yes. If you are sending large objects and the kernel is doing most of the moving, the Gold 6148 will fill the port, and at €4499 it is the cheapest way to own one. The gap opens when work happens per packet rather than per file. Encryption, deep inspection and userspace stacks all spend cycles and memory bandwidth on every packet, and 80 threads on DDR4-2133 leave far less margin for that than 256 threads on DDR5-4800. The short rule: buy the Xeon for delivery, buy an EPYC for processing.
When is a kernel bypass framework such as DPDK worth the trouble?
When your ceiling is packets per second rather than bytes per second. Small packets and per-packet decisions, so firewalling, scrubbing, telemetry and similar work, exhaust the kernel path long before the port is full. A poll mode driver removes that cost by taking the card and a few cores out of the operating system entirely. The price is that routing, retransmission and TLS become your code's problem. If your traffic is large objects, tune the kernel path and skip the framework. AF_XDP is a reasonable middle position when you are genuinely unsure.
How do I tell whether the port or my own software is the limit?
Test the link with a synthetic traffic generator first, so you hold a number that proves the port and the route, then run your real workload over the same path. If the generator reaches line rate and your application does not, the answer is on the server. Watch for a single core sitting at full soft interrupt load while the rest idle, which points at interrupt affinity, and for drop counters on the card or the socket, which usually point at buffer sizing. Our own system administrators will run that first test with you on request.

100 Gbps dedicated servers by location

Same hardware and guaranteed port, deployed from the datacenter closest to your users.

Tell us what your traffic looks like

Describe your workload to our system administrators, packets or files, and we will tell you which of the three builds fits and what the stack will need doing. Custom CPU, memory to 1TB DDR5, storage and bandwidth tier are quoted on request, and machines are online in about an hour.

Up to 1Tbps DDoS protectionDeployed in about 1 hour99.9% uptime SLA