A 100Gbps dedicated server is a specialist tool. Most websites never come close to filling a 1Gbps port, and a 10Gbps port already moves 3.24 PB in a month. At 100Gbps the arithmetic changes scale again, and so do the ways a server can fail to use the port it is paying for.
This guide covers who genuinely needs 100Gbps, which NexonHost builds can carry it and at what price, and the four things that have to be true inside the server before the port becomes the limit.
What 100Gbps means in practice
| Measure | 10Gbps | 100Gbps |
|---|---|---|
| Per second | 1.25 GB | 12.5 GB |
| Per hour | 4.5 TB | 45 TB |
| Per 30-day month | 3.24 PB | 32.4 PB |
| Concurrent 4K streams at 15 Mbps | 666 | 6,666 |
| A 50 GB game patch to 10,000 players | about 4.6 days | about 11 hours |
| Copying 1 PB | about 9.3 days | about 22 hours |
All of these assume the port runs at line rate, which real traffic rarely does for long. They are ceilings, not forecasts. The stream figure uses Netflix's own recommendation of 15 Mbps for 4K.
Who needs 100Gbps, and who does not
The workloads that justify a 100Gbps port share one trait: a small number of servers pushing a very large amount of data to a very large number of clients, or moving a very large dataset in a hurry.
- Video delivery. Origin and edge servers for live and on-demand video, where thousands of concurrent streams leave one machine.
- Software and game distribution. Launch-day patches, where the whole player base downloads the same files in the same few hours.
- Backup, replication and dataset moves. Petabyte-scale copies between sites, where the transfer window is the constraint.
- Proxy and scrubbing nodes. Machines that carry traffic for many other servers at once.
Most other workloads do not need it. A busy web shop, a SaaS application or a game server usually peaks well below 10Gbps, and paying for 100Gbps buys headroom it will never use. How much bandwidth does a server need? works through the sizing from page views and peak hours. If your graph peaks at a few gigabits, a 10Gbps port is the right answer.
Which builds carry 100Gbps
The port has to be matched by the network card and by the platform behind it. On NexonHost dedicated servers, ports run from 1Gbps up to 100Gbps depending on the build:
| Build | Network card | Port options | 100Gbps option | Total with 100Gbps |
|---|---|---|---|---|
| 2× Xeon E5 v3 / v4 | — | 1–20 Gbps | Not available | — |
| 1× or 2× EPYC 9254 | 2× 25 Gbps SFP28 | 1–50 Gbps | Not available | — |
| 2× EPYC 9554 | 2× 100 Gbps QSFP28 | 1–100 Gbps | €6,500 a month, guaranteed | €7,379 a month |
| 2× EPYC 9754 | 2× 100 Gbps QSFP28 | 1–100 Gbps | €6,500 a month, guaranteed | €7,939 a month |
| 2× EPYC 7702, Bucharest | — | 1–100 Gbps | €4,000 a month, unmetered | €4,989 a month |
The totals replace the default 1Gbps port with the 100Gbps option on each build's base price.
Price per gigabit
The faster ports cost less per gigabit, not more:
| Port on the EPYC 9004 builds | Monthly | Per Gbps |
|---|---|---|
| 1 Gbps | €80 | €80 |
| 10 Gbps | €800 | €80 |
| 25 Gbps | €1,875 | €75 |
| 50 Gbps | €3,500 | €70 |
| 100 Gbps | €6,500 | €65 |
At full utilisation, a 2× EPYC 9554 with a 100Gbps port costs €7,379 a month for up to 32.4 PB, which is about €0.23 per terabyte. The Bucharest EPYC 7702 route works out at €40 per gigabit for the port, if Bucharest suits your users.
Four bottlenecks to fix before the port is the limit
1. The platform
100Gbps needs a 100Gbps network card and enough PCIe bandwidth behind it. AMD's EPYC 9004 processors provide 128 lanes of PCIe 5.0 each, against 40 lanes of PCIe 3.0 on the Xeon E5 v3 and v4. The builds that offer 100Gbps are the ones with 100Gbps cards and the newest I/O; Xeon vs EPYC compares the platforms in detail.
2. Storage
Serving 100Gbps from disk means reading 12.5 GB every second, without pause. The 100Gbps builds carry 512 GB or 1 TB of DDR5, and AMD rates each EPYC 9004 socket at 460.8 GB/s of memory bandwidth, so the practical design is to keep the hot set in RAM, where the operating system's page cache does most of the work, and to spread the rest across many NVMe drives. The 9554 and 9754 builds take a 24-drive NVMe option: 24× 1.92 TB is €820 a month.
3. TCP
One TCP connection can only have as much data in flight as its window allows. To fill 100Gbps over a 30 ms round trip, the window has to hold 375 MB (100 Gbps × 0.03 s ÷ 8). Linux auto-tunes its buffers only up to a default ceiling that depends on RAM: 4 MB for sending, and 6 MB for receiving (32 MB from kernel 6.16). At 30 ms, 4 MB in flight is about 1.1 Gbps, and only part of a buffer becomes usable window, so an untuned single connection tops out around a gigabit on that path. Many parallel connections, which is the normal case for a server with thousands of clients, avoid the problem. For the few long transfers that need a single fast flow, raise the maximum values of net.ipv4.tcp_rmem and net.ipv4.tcp_wmem on both ends towards the bandwidth-delay product.
4. Packets per second
Bits are not the only limit. At 100Gbps with full-size 1,500-byte packets, a server handles about 8.1 million packets per second. With minimum-size 64-byte frames it would be 148.8 million, once Ethernet's preamble and inter-frame gap are counted. Normal traffic sits near the first number. Attacks aim for the second, which is why DDoS mitigation is quoted in packets per second as well as bits: the Xeon and EPYC 9004 builds include 40 Gbps / 4 Mpps, and the mitigation tiers go up to 1 Tbps / 100 Mpps.
One 100Gbps server or several 10Gbps servers?
Before ordering, ask whether the traffic needs to leave from one machine. Ten 10Gbps servers in different cities put capacity closer to users, survive the loss of any one of them, and let you grow in steps; choosing a European server location covers the trade-offs. One 100Gbps server wins when the data lives in one place, such as a single origin library, a replication target or a scrubbing node, or when operating ten machines costs more than the difference in port price.
Measure before you decide. A month of five-minute samples tells you both your peak and your sustained rate, and 95th percentile bandwidth billing explains why the shape of that graph matters as much as its height.
Where to start
The 100Gbps builds, along with every other port option, are listed with NexonHost's 100Gbps dedicated server range, and the full configuration list is on the dedicated servers page. The 100Gbps options are a flat monthly port price, not a charge per terabyte, so a busy month costs the same as a quiet one.
Specifications and pricing are as of 28 September 2026 — check the product pages for current figures.
Sources
- Internet connection speed recommendations, Netflix Help Center. 15 Mbps or higher for 4K. Retrieved 28 September 2026.
- IP Sysctl, Linux kernel documentation. The
tcp_wmemmaximum defaults to at most 4 MB, and thetcp_rmemmaximum to at most 6 MB (32 MB from kernel 6.16), depending on RAM. Retrieved 2 October 2026. - AMD specification pages for the EPYC 9554 and EPYC 9754: PCIe 5.0 x128, 460.8 GB/s per socket. Retrieved 28 September 2026.
- Intel Xeon Processor E5-2699 v4 specifications, Intel: PCIe 3.0, 40 lanes. Retrieved 28 September 2026.
- NexonHost store configurator: network cards, port options and prices, NVMe options, DDoS mitigation tiers. Retrieved 28 September 2026.
- Computed: 100 Gbps ÷ 8 = 12.5 GB/s; × 3,600 = 45 TB/hour; × 2,592,000 s = 32.4 PB per 30-day month; 100,000 Mbps ÷ 15 Mbps = 6,666 streams; 500 TB × 8 ÷ 100 Gbps = 40,000 s; bandwidth-delay product 100 Gbps × 30 ms ÷ 8 = 375 MB; 4,194,304 bytes × 8 ÷ 0.03 s ≈ 1.1 Gbps; packets per second = line rate ÷ ((frame + 20 bytes of preamble and gap) × 8).
