For more than a decade, 10 Gigabit Ethernet was the default answer for connecting a server. That has changed. NVMe storage, dense virtualisation and GPU workloads now move data faster than a 10GbE link can carry it, and 25GbE has quietly become the most common high-speed port on current Dell PowerEdge and HPE ProLiant servers. Meanwhile 40GbE has become a dead end, and 100GbE has dropped from the data centre core down to individual storage and hypervisor hosts.
So which network speed should a new or refurbished server actually use in 2026? The answer depends less on headline speed than on lanes, cabling, PCIe bandwidth, switch cost, and what the workload really needs.
Standardise on 25GbE for servers. Use 100GbE where links aggregate.
Match the network to the storage behind it: hard drives and SATA SSDs rarely push past 10GbE; NVMe servers replicating to each other will.
- The server runs file, print, AD or light applications
- Storage is hard drives or SATA SSDs
- You want RJ45 over existing Cat6a cabling
- You run a virtualisation or HA cluster
- Storage is NVMe or shared across hosts
- The server already has 25GbE ports onboard
- Budget is tight and the hardware is refurbished
- It's a lab or a direct storage link
- You accept a replacement, not an upgrade, later
- The link is a switch uplink or spine
- You run all-NVMe Ceph, vSAN or storage nodes
- GPU servers or RDMA workloads are involved
Understanding server network speeds
All four are Ethernet and carry the same frames, VLANs and protocols. What separates them is the physical layer: how many electrical lanes the port uses and how fast each lane runs. That single detail explains almost everything else in this guide, from cable choice to why 40GbE fell behind.
10GbE
One 10 Gb/s lane — SFP+ or RJ45 10GBASE-T. The workhorse from 2012 to around 2020.
- Lowest cost per port, new or refurbished
- 10GBASE-T runs on existing Cat6a cabling
- Universal support across servers and switches
- A single NVMe drive can saturate it
- No longer the default on current servers
25GbE
One 25 Gb/s lane in the SFP+ cage size. The most common high-speed server port in 2026.
- 2.5× the bandwidth of 10GbE on the same port size
- Breaks out cleanly from 100GbE switch ports
- Dual-port cards fit a PCIe 3.0 x8 slot
- Fibre or DAC only — no practical RJ45
- FEC must match on both ends of the link
40GbE
Four 10 Gb/s lanes in a QSFP+ port. Superseded, but very cheap refurbished.
- Lowest cost per gigabit on the refurbished market
- Breaks out to 4 × 10GbE for existing servers
- Good fit for labs and budget storage links
- Dead end: no new platforms are built around it
- SR4 optics need MPO trunk cabling
100GbE
Four 25 Gb/s lanes in a QSFP28 port. Uplinks, all-NVMe storage and GPU servers.
- 10× the bandwidth of 10GbE, enough for all-NVMe
- Lowest cost per gigabit on new switches
- RDMA (RoCEv2) for storage and AI fabrics
- Needs PCIe 3.0 x16, or 4.0 for both ports
- Overkill unless storage can fill it
Performance comparison
Raw link speed is the obvious difference. Converted into the units storage and applications use, the gap is clear:
Put those figures next to storage and the problem with 10GbE becomes obvious. A SATA SSD reads at about 0.55 GB/s, which 10GbE handles comfortably. A PCIe Gen3 NVMe drive reads at around 3.5 GB/s and a Gen4 drive at around 7 GB/s. One Gen3 NVMe drive fills a 25GbE port; one Gen4 drive needs more than half of a 100GbE port. For the full storage picture, see our NVMe vs SAS vs SATA comparison.
The lane diagram explains why 25GbE replaced 40GbE. A 40GbE port is four 10G lanes. A 25GbE port is a single, faster lane, so it delivers more than half of 40GbE's bandwidth on one cable pair, on a smaller and cheaper port — and four of those lanes together make 100GbE. Once switch chips moved to 25G lanes, 40GbE had nowhere to go.
| Spec | 10GbE | 25GbE | 40GbE | 100GbE |
|---|---|---|---|---|
| Transceiver | SFP+ | SFP28 | QSFP+ | QSFP28 |
| Electrical lanes | 1 × 10G | 1 × 25G | 4 × 10G | 4 × 25G |
| Throughput | 1.25 GB/s | 3.1 GB/s | 5 GB/s | 12.5 GB/s |
| Time to send a 1500-byte frame | 1.2 µs | 0.48 µs | 0.3 µs | 0.12 µs |
| RJ45 copper option | Yes, 10GBASE-T | Not practical | Not practical | No |
| Breakout | None | None | 4 × 10GbE | 4 × 25GbE or 2 × 50GbE |
| NIC PCIe slot (dual port) | PCIe 2.0 x8 | PCIe 3.0 x8 | PCIe 3.0 x8 (shared) | PCIe 3.0 x16 / 4.0 x16 |
| Typical optic power | About 1 W (SR) | About 1 W (SR) | About 1.5 W (SR4) | About 2.5–3.5 W |
| Status in 2026 | Mainstream, ageing | Current standard | Legacy | Uplinks and storage |
How long does it take to move the data?
Throughput becomes tangible when it's turned into time. The figures below are at about 90 percent of line rate, realistic for a well-configured link with fast storage at both ends.
| Task | 10GbE | 25GbE | 40GbE | 100GbE |
|---|---|---|---|---|
| Copy 1 TB | About 15 min | About 6 min | About 4 min | About 1.5 min |
| Live-migrate a 256 GB VM | About 4 min | About 1.5 min | About 1 min | Under 25 sec |
| Full backup of 20 TB | About 5 hours | About 2 hours | About 1.2 hours | About 30 min* |
| Rebuild a failed 8 TB Ceph OSD | Network-bound | Disk-bound | Disk-bound | Disk-bound |
*At 100GbE the network is rarely the limit. Backup and copy jobs at that speed are usually bound by the source disks, the target disks or the backup software, so faster links only pay off when the storage behind them can keep up.
Cabling, optics and reach
Cabling is where the real cost and complexity of each speed shows up. Inside a rack, direct attach copper (DAC) cables are the cheapest and lowest-power option at every speed. Between racks and rooms, fibre optics take over, and the choice of optic decides which fibre you need.
| Connection | 10GbE | 25GbE | 40GbE | 100GbE |
|---|---|---|---|---|
| Passive DAC (in rack) | Up to 7 m | Up to 5 m | Up to 7 m | Up to 5 m |
| Active optical cable | Up to 100 m | Up to 100 m | Up to 100 m | Up to 100 m |
| RJ45 copper | 100 m Cat6a, 55 m Cat6 | Not practical | Not practical | Not available |
| Short-reach multimode | SR: 300 m OM3, 400 m OM4 | SR: 70 m OM3, 100 m OM4 | SR4: 100 m OM3, 150 m OM4 | SR4: 70 m OM3, 100 m OM4 |
| Multimode connector | Duplex LC | Duplex LC | MPO-12 | MPO-12 |
| Long-reach single-mode | LR: 10 km | LR: 10 km | LR4: 10 km | LR4: 10 km, CWDM4: 2 km |
25GbE link won't come up? Almost always the cause is a forward error correction (FEC) or auto-negotiation mismatch between the NIC and the switch. Set FEC to the same mode on both ends — RS-FEC is the usual choice for 25G DAC and SR optics — before suspecting the cable.
Planning a new build? Our guides to SFP, SFP+ and QSFP transceivers and Cat5e, Cat6 and Cat6a structured cabling cover optics and copper in more depth.
Real-world workload guide
Most business servers don't need more than 25GbE, and many still don't need more than 10GbE. A server with SATA SSDs or hard drives will rarely push past 10GbE; a cluster of NVMe servers replicating to each other will make 10GbE the slowest component in the building.
-
File servers, domain controllers, small officesUsers connect at 1GbE, so 10GbE is ample — and 10GBASE-T reuses existing Cat6a.10GbE
-
Virtualisation clusters (Proxmox, Hyper-V, VMware)Fast live migration and HA failover; split VM, storage and migration traffic on 2 × 25G.25GbE
-
Backup targets and NASShrinks backup windows 2.5× over 10GbE; inexpensive with refurbished ConnectX-4 Lx cards.25GbE
-
Ceph, vSAN and hyper-converged storageReplication multiplies every write; 10GbE bottlenecks rebuilds. All-NVMe clusters want 100GbE back-ends.25GbE min · 100GbE NVMe
-
AI, GPU servers and HPCDatasets are terabytes; RDMA (RoCEv2) cuts CPU load and latency. Multi-GPU clusters move to 200G and 400G.100GbE+
-
Labs, test environments, budget storage linksVery low cost per gigabit on used NICs and switches — plan to replace rather than expand.40GbE refurb
Server and NIC compatibility
Every network card has to fit both a physical slot and the PCIe bandwidth behind it. A dual-port 100GbE card needs 200 Gb/s, which a PCIe 3.0 x16 slot (about 126 Gb/s usable) cannot supply, so on older servers it runs as one active port plus a failover. Our article on PCIe Gen3 vs Gen4 in enterprise servers explains the difference.
| Server generation | PCIe | Onboard / mezzanine | Practical maximum |
|---|---|---|---|
| Dell 13G (R630, R730) / HPE Gen9 | 3.0 | rNDC or FlexibleLOM: 1GbE or 10GbE | 25GbE, or single-port 100GbE via PCIe card |
| Dell 14G (R640, R740) / HPE Gen10 | 3.0 | rNDC or FlexibleLOM: 10GbE or 25GbE | 100GbE via PCIe 3.0 x16 card |
| Dell 15G (R650, R750) / HPE Gen10 Plus | 4.0 | OCP 3.0: 10GbE or 25GbE | Dual-port 100GbE at line rate |
| Dell 16G (R660, R760) / HPE Gen11 | 5.0 | OCP 3.0: 10GbE, 25GbE or 100GbE | 200GbE and 400GbE adapters |
| Speed | Common server NICs | PCIe requirement |
|---|---|---|
| 10GbE | Intel X520-DA2, X540-T2, X550-T2, X710-DA2 / DA4; Broadcom 57810S | PCIe 2.0 x8 or 3.0 x8 |
| 25GbE | Mellanox ConnectX-4 Lx, ConnectX-5, ConnectX-6 Lx; Intel XXV710, E810-XXVDA2; Broadcom 57414 | PCIe 3.0 x8 |
| 40GbE | Mellanox ConnectX-3 Pro; Intel XL710-QDA2; Chelsio T580 | PCIe 3.0 x8 |
| 100GbE | Mellanox ConnectX-4, ConnectX-5, ConnectX-6 Dx; Intel E810-CQDA2; Broadcom P2100G | PCIe 3.0 x16 (one port) or 4.0 x16 (two ports) |
Before buying a NIC, check three things: a free slot with the right lane width, the riser it sits in, and whether the optics are coded for your switch. Many switches reject third-party or differently coded transceivers, while most NICs accept a wider range.
Cost vs performance
The cost of a network speed is never just the NIC. It's the switch port, the optic or DAC at each end, the fibre, and the power. In 2026 the refurbished market has made 10GbE and 40GbE extremely inexpensive, 25GbE NICs very affordable, and 25GbE and 100GbE switches the main cost of moving up.
| Speed | Typical switch | Cost per port | Cost per Gbps |
|---|---|---|---|
| 10GbE | 48 × SFP+ with 40G or 100G uplinks (e.g. Dell S4048-ON, Cisco Nexus 3172) | Lowest | Moderate |
| 25GbE | 48 × SFP28 with 6 to 8 × 100G (e.g. Dell S5248F-ON, Cisco Nexus 93180YC-FX) | Moderate | Low |
| 40GbE | 32 × QSFP+ (e.g. Dell S6000, Cisco Nexus 3132Q) | Low (refurbished) | Lowest (refurbished) |
| 100GbE | 32 × QSFP28 (e.g. Dell S5232F-ON, Cisco Nexus 9336C-FX2) | Highest | Lowest (new) |
A practical design for most growing businesses: 25GbE to every server, 100GbE between switches, and 10GbE or 1GbE to the access layer.
100GbE switch ports with 4 × 25G breakout cables let the same switch serve both roles. Our network switch buying guide covers layer 2 and layer 3 choices.
Where server networking is heading
Ethernet is moving faster now than at any point in its history, driven almost entirely by AI clusters. The speed of each lane keeps doubling, and the server port follows a few years behind.
Three trends matter for business buyers. First, 25GbE is becoming what 1GbE once was: the assumed minimum on every new server. Second, 100GbE is falling into the price range of mid-sized businesses, especially refurbished, as hyperscalers move to 400GbE and 800GbE. Third, Ethernet is absorbing workloads that used to need InfiniBand, with RDMA over Converged Ethernet (RoCEv2) and the Ultra Ethernet specification aimed at AI and storage traffic.
For a server bought in 2026, that points to 25GbE as the safe default, with 100GbE reserved for storage, uplinks and GPU servers. A 10GbE network installed today will still work for years, but it will be the first part of the infrastructure to feel slow.
Conclusion
25GbE is the best choice for most servers in 2026. It delivers two and a half times the bandwidth of 10GbE on the same size of port, fits a PCIe 3.0 x8 slot, comes onboard with current Dell and HPE platforms, and breaks out neatly from 100GbE switches. 10GbE remains perfectly adequate for file servers, domain controllers and anything backed by hard drives, and it's the only one of the four that runs on standard RJ45 cabling.
100GbE belongs on switch uplinks, all-NVMe storage, hyper-converged clusters and GPU servers, where the storage can actually fill it. 40GbE is worth buying only as inexpensive refurbished hardware for labs and point-to-point links, knowing it will be replaced rather than expanded.
Common questions
Is 25GbE worth it over 10GbE?
For virtualisation, NVMe storage and backup targets, yes: 2.5× the bandwidth on the same size of port, and dual-port cards fit a PCIe 3.0 x8 slot. For file servers and anything on hard drives, 10GbE is still ample.
Can an SFP28 port use 10G SFP+ optics?
Usually, yes. SFP28 ports generally accept 10G SFP+ optics and run at 10GbE, which makes stepping up gradually straightforward.
Why won't my 25GbE link come up?
Almost always a forward error correction (FEC) or auto-negotiation mismatch. Set FEC to the same mode on both the NIC and the switch — RS-FEC is the usual choice for 25G DAC and SR optics.
Does a dual-port 100GbE card need PCIe 4.0?
To run both ports at line rate, yes. A PCIe 3.0 x16 slot supplies about 126 Gb/s, so on older servers the card runs as one active port plus a failover.
