25G vs 10G vs 100G — which server network speed do you need
Three Ethernet speeds cover almost every decision in a modern server room: 10GbE, 25GbE and 100GbE. Older 40GbE still turns up on the refurbished market but is a technological dead-end, and 400GbE remains the preserve of AI clusters and hyperscale. This guide compares the three that actually matter, so you can put the right speed at the server and the right speed between switches.
The lane story — why 25G exists at all
The single most useful thing to understand is how these speeds are built.
- 10GbE runs on one lane — one electrical lane in a DAC, one fibre pair in an SR/LR optic — through an SFP+ cage.
- 25GbE also runs on one lane, through an SFP28 cage. It was standardised (IEEE 802.3by, 2016) by taking a single lane of a 100G link and defining it on its own.
- 100GbE is four 25G lanes bundled into one QSFP28 cage.
That is why 25G, not 40G, became the modern server default. On the same connector type, the same cabling class and the same single pair as 10G, you get 2.5× the throughput. And because 100G is simply four of those 25G lanes, a 100G switch port can break out into four 25G server links — 10G and 25G and 100G form one coherent family, while 40G (four 10G lanes) sits awkwardly outside it.
Head-to-head
| 10GbE | 25GbE | 100GbE | |
|---|---|---|---|
| Per-lane rate | 10G × 1 | 25G × 1 | 25G × 4 |
| Cage / module | SFP+ | SFP28 | QSFP28 |
| IEEE standard | 802.3ae / 802.3an | 802.3by (2016) | 802.3ba + later |
| Signalling | NRZ | NRZ | NRZ (4 lanes) |
| In-rack (DAC) | 1–5 m | 1–5 m (FEC rules apply) | 1–5 m |
| Multimode optic (OM4) | SR: 400 m | SR: 100 m | SR4: 100 m |
| Single-mode optic | LR: 10 km; ER: 40 km | LR: 10 km; ER: 40 km | DR: 500 m; FR: 2 km; LR: 10 km |
| Copper RJ45 option | 10GBASE-T: 100 m Cat6a | — | — |
| Breaks out to | — | — | 4 × 25GbE |
| Typical role | Budget/general server access | Modern server access (ToR) | Uplinks, spine, storage, AI |
Who each speed is for
10GbE — the budget floor. Still the most cost-effective way to connect a server, and refurbished 10G switches and SFP+ optics are plentiful. It is the right choice for general-purpose servers, small-business storefronts, and iSCSI/NAS storage that does not saturate the link. It also has a unique advantage: 10GBASE-T over RJ45 copper reaches 100 m on Cat6a, so where structured copper is already in the walls, 10G can reuse it without new fibre.
25GbE — the modern default. For any new server deployment, 25G is the sweet spot. It uses the same SFP28 cage size and the same single-pair cabling as 10G, runs on the same in-rack DACs, and delivers 2.5× the bandwidth — for a cost and power budget now close to 10G. Crucially it lines up with 100G uplinks: a top-of-rack switch such as the Dell PowerSwitch S5248F-ON offers 48 × 25GbE server ports and 100GbE uplinks on the same box.
100GbE — uplinks, spine and heavy workloads. 100G is where you connect switches to each other (leaf-to-spine), attach shared storage back-ends, and feed virtualisation hosts and GPU/AI servers that a 25G link would bottleneck. A single QSFP28 port can also break out to four 25G server links, so even a compact 100G switch — the MikroTik CRS504-4XQ-IN carries four QSFP28 ports and fans each out to four 25G devices — does double duty as an aggregation point and a server access switch.
Cabling and transceiver implications
- In the rack, use DACs. For 10G, 25G and 100G alike, a passive direct-attach copper cable (roughly 1–5 m) is the cheapest, coolest-running server-to-switch link. Reserve optics for runs a DAC cannot make.
- Between rows or across a building, use optics. Multimode SR-type optics are cheaper and reach up to a few hundred metres; single-mode LR-type optics reach 10 km. Note that at 25G, SR reach on OM4 is 100 m — shorter than 10G's 400 m — so a long multimode run that carried 10G may not carry 25G without re-checking the fibre.
- Mind FEC at 25G and up. Forward error correction must match at both ends. On 25G DACs, short cables need no FEC while longer ones require BASE-R or RS-FEC; a mismatch is a classic cause of a link that will not come up.
- Breakout is the money-saver. One 100G QSFP28 uplink port plus a breakout cable equals four 25G server ports — often cheaper and denser than buying separate 25G ports.
Cost, power and the refurbished market
Without quoting prices, the trade-offs are consistent:
- Per-port cost and power rise with speed and reach. DAC < AOC < multimode optic < single-mode optic, at every speed. 100G optics draw meaningfully more power than 25G, which matters for switch thermal and power budgets.
- 25G has closed the gap on 10G. Because it reuses 10G's cabling and cage and runs on one lane, 25G's per-port economics are now close enough that new builds rarely justify starting at 10G unless reusing existing 10G kit.
- The refurbished market favours 10G and 40G. Both are plentiful and inexpensive second-hand — 10G is a genuinely smart budget choice, but 40G is not, because it cannot evolve toward 25/50/100G lanes (see below).
Why to skip 40GbE
40GbE looks like a bargain on the refurbished market, but it is built from four 10G lanes in a QSFP+ cage. That means:
- It offers no single-25G-lane path, so a 40G port cannot be broken out into useful 25G server links (unlike 100G → 4×25G).
- QSFP+ optics and switches do not carry forward into the 25/50/100G roadmap.
- 100GbE (QSFP28) is both faster and more flexible for the same uplink role.
Unless you are extending an existing 40G island, put new spend into 100GbE for uplinks and 25GbE for servers instead.
Recommendation matrix
| Scenario | Recommended speed |
|---|---|
| Budget server access; reuse existing Cat6a copper | 10GbE (10GBASE-T) |
| General-purpose server access, budget-constrained | 10GbE |
| New server / top-of-rack access | 25GbE |
| Virtualisation, VDI or database hosts | 25GbE (100GbE if very dense) |
| Switch-to-switch uplinks (leaf–spine) | 100GbE |
| Shared storage back-end (SAN/NAS at scale) | 100GbE |
| GPU / AI training servers | 100GbE (or higher) |
| Extending an existing 40G island only | 40GbE (otherwise skip) |
The short answer for most South African server rooms: 25GbE to the server, 100GbE between the switches, 10GbE as the budget floor — and 40GbE only to extend what you already own.
Switches, NICs and transceivers, priced in rand:
The Server Hub briefing
South African IT hardware news, once a week.
What’s new, what it costs in rand, and what it means for the kit you run — servers and storage, networking, backup power, surveillance and print. Every claim checked against a named source.