Server Networking Buying Guide 2026: 10G, 25G & 100G NICs, SFP+ & QSFP+ Modules

Q: Can I use generic transceivers in Cisco / HPE / Juniper switches?

Yes, but you may need to enable a vendor unlock command (e.g. service unsupported-transceiver on Cisco IOS) and you forfeit OEM TAC support. For production fabrics, prefer OEM-coded; for test labs and edge, generic is fine.

Q: Do I need RDMA for storage?

If using NVMe-over-Fabrics, vSAN, or Storage Spaces Direct, RDMA (RoCEv2) is required for low-latency line-rate performance. Without RDMA the CPU becomes the storage bottleneck.

Q: What is the difference between SFP+ and SFP28?

SFP+ is 10 GbE. SFP28 is 25 GbE, same form factor, backwards compatible — a 25G SFP28 port accepts a 10G SFP+ module and runs at 10G.

Q: QSFP+ vs QSFP28 vs QSFP-DD?

QSFP+ is 40 GbE (4×10G). QSFP28 is 100 GbE (4×25G). QSFP-DD is 200/400 GbE (8 lanes, doubled stack). Each is backwards compatible with breakout cables to the prior generation.

Q: Do you supply Mellanox/NVIDIA ConnectX-7 and 400 GbE optics?

Yes. ICD sources ConnectX-7 (400G) and matching 400G QSFP-DD modules from verified US and European suppliers. Lead time is confirmed on the quote and depends on global allocation.

Request a Quote Contact Sales

Data-center networking in 2026 has effectively moved past 1 GbE for north-south server traffic. The volume bands now are 10 GbE for general compute, 25 GbE for converged and storage-aware fabrics, and 100 GbE for spine, AI fabric and storage front-end. Choosing the right NIC, transceiver and DAC/AOC cable depends on switch model, fabric topology (top-of-rack vs end-of-row), reach, and whether you need protocol offload (RoCE, iWARP, NVMe-oF). ICD’s catalog covers Mellanox/NVIDIA, Broadcom, Intel, Marvell QLogic, Cisco, Juniper and F5 modules across all major server platforms.

What to look for in 2026

Five drivers shape the right choice:

Port speed should match the switch ToR. 10G NIC into 25G switch downshifts to 10G. 25G NIC into 10G switch downshifts to 10G. To realize 25G or 100G, both ends must support the speed.
Transceiver type drives reach. SR (Short Reach, 850 nm multimode, 100-400 m), LR (Long Reach, 1310 nm single mode, 10 km), ER/ZR (40-80 km). For in-rack runs under 7 m, prefer passive DAC (Direct Attach Copper) at half the cost. For 7-30 m, AOC (Active Optical Cable). For >30 m or building-to-building, SR/LR optics.
OCP form factor for new builds. Dell, HPE and Lenovo Gen10/Gen11/V3 platforms offer OCP 3.0 slots that free up a PCIe slot and offer hot-swap NIC service. Prefer OCP NICs when ordering new servers.
RoCEv2 / NVMe-oF support. If the storage fabric is converged (Ethernet for both data and storage), the NIC must support RoCEv2 with proper congestion control (PFC / DCB). Mellanox/NVIDIA ConnectX-5/6/7 are the default. Broadcom NetXtreme E-Series and Intel E810 also work but verify firmware.
Coding strategy (FEC). 25G and 100G require Forward Error Correction. RS-FEC adds latency but is mandatory for some long-reach modules. Confirm FEC mode (CL74, CL91, RS-FEC) matches on both ends or the link will not come up.

Top picks by use case

Active SKUs across the speed bands:

W5HC8 — Broadcom 57454 Quad-Port 10GBASE-T RJ45 OCP NIC. For copper Cat6A short runs, OCP 3.0 form factor — common Dell R750 / R760 fitment.
QL41164HMCU-DE — Marvell QLogic FastLinQ QL41164 Quad-Port 10 GbE SFP+ adapter. RoCE/iWARP capable, PCIe Gen3 x8.
7ZT7A00536 — Lenovo ThinkSystem Intel I350-T4 ML2 1 GbE 4-Port RJ45. Management or lower-tier compute connectivity for ThinkSystem rack servers.
C-NIC-25G2D1-CM — Nutanix 25/10 GbE 2-Port NIC Expansion. For Nutanix-certified clusters requiring HCI-tested adapters.
MCX4431A-GCAN — Mellanox ConnectX-4 Lx EN 50 GbE single-port QSFP28 OCP 2.0. PCIe 3.0 x8, RoCEv2-capable for converged fabrics.
AB079026 — Dell NVIDIA ConnectX-6 Dx EN 100 GbE Dual-Port QSFP56 OCP 3.0. For Dell PowerEdge 15G/16G high-bandwidth nodes (AI hosts, all-flash storage).
105-000-492-00 — Dell EMC Isilon Mellanox CX314A 40 GbE Dual-Port QSFP NIC. For Isilon back-end fabric.
02313UBW — 2-Port 40 GE QSFP+ NIC Card. Generic 40 GbE for storage and HPC fabric.
CFP2-100G-LR4 — Cisco 100GBASE-LR4 CFP2 Transceiver Module, single-mode, 10 km LC. For Nexus 7000 / 9500 100G uplinks.
QFX-QSFP28-DAC-3M — Juniper 100G QSFP28 passive DAC 3 m. In-rack 100G short-reach, cost-effective for spine-leaf top-of-rack.
QSFP-H40G-CU3M — Cisco 40G QSFP+ DAC passive 3 m. Standard 40 GbE intra-rack connection.
SFP-GE-Z — Cisco 1000BASE-ZX SFP transceiver, 80 km. Long-haul gigabit, common in metro Ethernet.
DWDM-SFP10G-60.61 — Cisco 10G DWDM SFP+ 1560.61 nm, 80 km SMF. Long-reach 10G over DWDM-enabled infrastructure.
F5-UPG-QSFP+PIR4 — F5 40 GbE QSFP+ PIR4 transceiver, 2 km. For F5 BIG-IP appliance uplinks.
R0Q79A — HPE MSA 2062 16 Gb Fibre Channel LFF storage transceiver (note: FC, not Ethernet — listed for completeness for converged storage planning).

See Cisco Nexus 9000 Fixed and Cisco Catalyst 9300 for matching switches.

Compatibility considerations

Five compatibility rules:

OEM-coded transceivers. Cisco, HPE, Dell, Juniper and F5 lock transceivers via vendor codes. A generic 100G QSFP28 LR4 from one OEM will not light up in another’s switch without re-coding (which voids warranty). For OEM-locked deployments, ICD ships OEM-coded modules.
DAC/AOC length budget. Passive DAC is rated to 7 m at 25 G and 5 m at 100 G. AOC extends to 30 m. Beyond that, switch to SR/LR optics.
OCP slot version. OCP 2.0 (older) and OCP 3.0 (current) are different mechanical/electrical specs. Dell 14G uses OCP 2.0; 15G/16G uses OCP 3.0. Order the matching version.
PCIe lane budget. 100 GbE NIC requires PCIe 4.0 x16 or PCIe 3.0 x16 for line rate. PCIe 3.0 x8 caps 100G NIC at ~64 Gbps practical throughput.
Firmware/driver minimum. Mellanox ConnectX-6 / ConnectX-7 require recent OFED/MLNX_OFED drivers for full RoCEv2 and DPDK support. Verify your OS is supported.

Pricing factors (USD, quote-based)

ICD prices all networking in USD. Final pricing reflects: speed grade (10G < 25G < 100G < 400G), transceiver reach (DAC < SR < LR < ER/ZR), OEM-coded vs generic (OEM 30-100% premium where switch locks), condition (NEW factory-sealed, RENEWED, REFURB pulled-tested), warranty (3-year ICD warranty on networking and optics, extendable up to 10 years), volume tier and incoterms. Long-reach DWDM and ZR optics are typically quoted per project, with lead time confirmed on the quote.

ICD advisory + 24-hour quote

For a precise quote, share: switch model and version (Cisco IOS-XE / NX-OS, Juniper Junos, Arista EOS), server NIC slot (PCIe or OCP, generation), target speed (10/25/40/100 G), reach (in-rack, intra-row, building-to-building, metro), and any fabric requirement (RoCE, DCB, FCoE). ICD presales returns a verified-compatible BOM in USD with availability and lead time within 24 hours. Submit at icd3s.com/quote-request/ or email [email protected].

Frequently asked questions

Can I use generic transceivers in Cisco / HPE / Juniper switches? Yes, but you may need to enable a vendor unlock command (e.g. service unsupported-transceiver on Cisco IOS) and you forfeit OEM TAC support. For production fabrics, prefer OEM-coded; for test labs and edge, generic is fine.

Do I need RDMA for storage? If using NVMe-over-Fabrics, vSAN, or Storage Spaces Direct, RDMA (RoCEv2) is required for low-latency line-rate performance. Without RDMA the CPU becomes the storage bottleneck.

What is the difference between SFP+ and SFP28? SFP+ is 10 GbE. SFP28 is 25 GbE, same form factor, backwards compatible — a 25G SFP28 port accepts a 10G SFP+ module and runs at 10G.

QSFP+ vs QSFP28 vs QSFP-DD? QSFP+ is 40 GbE (4×10G). QSFP28 is 100 GbE (4×25G). QSFP-DD is 200/400 GbE (8 lanes, doubled stack). Each is backwards compatible with breakout cables to the prior generation.

Do you supply Mellanox/NVIDIA ConnectX-7 and 400 GbE optics? Yes. ICD sources ConnectX-7 (400G) and matching 400G QSFP-DD modules from verified US and European suppliers. Lead time is confirmed on the quote and depends on global allocation.