25GbE (and 40/100GbE) - The Tier Above 10GbE

Computed live from in-stock parts: one Intel E810 (QSFP28) as the origin, one QSFP28-to-4x25G breakout cable, and 4 × Intel E810 (SFP28) - one per host. That connects 4 separate 25G hosts to a single 100G-capable origin for $1163.84 total, $290.96 per host, not counting the origin card itself since it is shared across all four legs.

ipolex ipolex QSFP28 card (Intel E810 claimed) $405.99 · 10Gtek breakout DAC $45.89 · H!Fiber H!Fiber SFP28 card (Intel E810 claimed) $177.99 each

25GbE (SFP28) network cards in stock now

SFP28 direct-attach cables

Passive direct-attach copper, by length. A DAC is a fixed-length cable with the transceivers moulded on - you buy the whole run at once rather than a cable plus two modules.

SFP28 transceivers and active optical cables

Every "compatible with" phrase below is the seller's own wording, recorded as a claim and badged. We have not tested any module against any switch.

Why the used market dominates this tier too

The same economics that made 10GbE affordable repeat one generation up, and our own catalogue shows it rather than just asserting it. Every ConnectX-5 listing we currently track is Mellanox-branded rather than a third-party board, and both ConnectX-5 and the ConnectX-4 Lx card in the 10GbE cluster below carry NVIDIA's own end-of-life notice - cards a datacentre bought by the pallet and has since replaced are exactly what fills a secondary market like this one.

Intel E810 shows the opposite half of the same pattern: four of the five listings we track are third-party boards from 10Gtek, H!Fiber and Ipolex asserting an Intel E810 controller, against one first-party Supermicro card. That is the third-party-undercut finding the 10GbE cluster already documents for X520 and X710, repeating a generation up rather than being a coincidence of this specific card.

The switches follow the same pattern at a larger scale. Arista's own End-of-Sale advisory archive carries a 2016 notice retiring select 7050SX/7050TX/7050QX models, and separately, 2026-dated notices retiring the newer 7050SX3-48YC12 and 7050SX3-48YC8/48C8 - switches formally end-of-saled by their own maker are exactly the hardware that lands secondhand at a fraction of new pricing. A used 7050-family switch bought today may already be running silicon Arista itself no longer sells new.

The form-factor map: SFP28, QSFP+ and QSFP28

Three connector families cover this whole tier, and which one a card or switch port has decides every cable and module decision after it.

SFP28 is the 25G cage - physically the same size and shape as the SFP+ cage the 10GbE cluster uses, and backward compatible with it, the same relationship 10GBASE-T has with the older 1000BASE-T. A ConnectX-5 or Intel E810 25G card, an SFP28 DAC, and an SFP28 switch port are the building blocks of a 25GbE build.

QSFP+ is the 40G cage - four lanes in one connector, standardised for 40GbE. A ConnectX-4 (non-Lx) card, the Arista 7050-family and Dell S4048-ON switches in our table, and QSFP+ DAC cables all share this family.

QSFP28 is the 100G cage - the same four-lane idea at the higher per-lane rate, used by ConnectX-5's and Intel E810's 100G SKUs, the Dell S5248F-ON and several QNAP and Mellanox switches in our table, and QSFP28 DAC and breakout cabling.

None of the three plugs into either of the others without a breakout or hybrid cable built for exactly that pairing - the same non-interchangeability rule the SFP+/SFP28/QSFP+/QSFP28 family follows throughout this cluster and the cabling guide covers in full.

Where breakout cables change the math

A QSFP28 100G port is electrically four 25G lanes bundled into one connector, and a breakout cable splits them back out - one QSFP28 end, four SFP28 legs, each running an independent 25G link. The same relationship holds one tier down: a QSFP+ 40G port breaks out to four SFP+ 10G legs.

That is not a trick or a workaround; it is how the standard defines the port, and it is frequently the cheapest way to wire several hosts to one switch or one server. Computed from our own live catalogue rather than assumed: one 100G-capable card, one QSFP28-to-4x25G breakout cable, and four 25G cards - one per host - is a real, priced arithmetic problem, and the answer is on this page.

The trade is flexibility rather than cost in the usual case: a breakout cable's four legs go to four separate devices, not one device with four times the bandwidth, and the switch or card doing the breaking-out has to support it (documented per-model in the switch table below, not assumed). Where a build genuinely needs one fast link rather than four independent ones, a straight same-speed DAC is the simpler answer.

Where to go next

The 10GbE hub is the tier below this one, and the step-up arithmetic between them lives at 10GbE vs 25GbE · 40GbE and 100GbE are this cluster's own siblings, each with their own live builds · the cabling guide covers every DAC, AOC, transceiver and breakout cable in this cluster in one place · rack disk shelves and NAS devices are usually what sits at the other end of a build like this · used servers are the other common host - check the registry for which models expose a compatible PCIe slot.

Frequently asked questions

Sources

Specification claims on this page were checked against the pages below on 4 August 2026. Where a widely-repeated figure did not survive that check, we report what the source says now and say so in the text rather than repeating the familiar number.

1. NVIDIA (Mellanox) - ConnectX-3 EN product line and lifecycle status (nvidia.com, 2026)
   Used for: that ConnectX-3 has reached end-of-life status with NVIDIA, and the MCX311A / MCX312A / MCX314A part-number family. EOL means no new firmware or driver work from the maker; it does not mean the card stops working, and the page says so in those terms.
2. The Linux kernel project - mlx4, mlx5, ixgbe and i40e in-tree Ethernet drivers (kernel.org, 2026)
   Used for: that mlx4 (ConnectX-3), mlx5 (ConnectX-4 and later), ixgbe (Intel 82599/X520, X540, X550) and i40e (X710) ship in the mainline Linux kernel. This is why an end-of-life card can still be a sane purchase, and it is a documented fact rather than a forum claim.
3. DatacenterDisk - Live catalogue - NIC, DAC and transceiver listings (datacenterdisk.com, 2026-08-11)
   Used for: every price on this page, read from our catalogue at request time. Seeded 11 August 2026: 24 network cards across 6 models, 24 DAC cables, 8 transceivers. Every card is split tier 1 (the listing is that part) or tier 2 (a third-party board asserting that silicon).
4. DatacenterDisk - Sourcing note - power figures and compatibility claims (datacenterdisk.com, 2026-08-11)
   Used for: We publish no per-port wattage for either cabling path. The relationship - 10GBASE-T draws more power and runs hotter than SFP+ direct attach - is documented at product level and is stated here. The magnitude is not, because we have not fetched and logged a source we can cite for it. Separately: every 'compatible with' phrase shown on this site is stored as a seller claim at tier 2 and badged wherever it renders. We have not tested any module against any switch.
5. NVIDIA - ConnectX-5, ConnectX-4 Lx and ConnectX-4 EN user manuals - end-of-life status (networking-docs.nvidia.com, 2026-09-13)
   Used for: that NVIDIA's own manuals state "The ConnectX-5 product line has moved to end-of-life" and "The ConnectX-4 Lx product line has moved to end-of-life" in those exact words (connectx5enhw and connectx4lxhw respectively, both dated 2026). The standard ConnectX-4 EN manual (connectx4enhw) carries an "EOL'd Ordering Part Numbers" table for every MCX413A/414A/415A/416A SKU but not the same plain-English banner sentence - stated as that asymmetry rather than smoothed into one claim. Also the per-SKU PCIe generation and lane count (PCIe 3.0 x8 or x16 depending on SKU; PCIe 3.0/4.0 on ConnectX-5 Ex variants) from the same manuals.
6. NVIDIA - Linux Kernel Upstream Release Notes - driver and firmware support table (networking-docs.nvidia.com, 2026-09-13)
   Used for: that ConnectX-4, ConnectX-4 Lx and ConnectX-5 all use the mlx5 driver family per NVIDIA's own "Supported Uplinks to Servers" table, dated June 2026 - current documentation, confirming the in-kernel driver is still maintained for hardware NVIDIA itself calls end-of-life for ordering. Also that NVIDIA's current DOCA-Host device list (docs dated September 2026) includes ConnectX-5 and ConnectX-4 Lx but not the standard ConnectX-4 - a real difference in ongoing software-package support between the two, stated as found rather than smoothed over.
7. The Linux kernel project - i40e and ice in-tree Ethernet driver documentation (kernel.org, 2026-09-13)
   Used for: that the i40e driver explicitly lists Intel Ethernet Controller XXV710 as supported hardware alongside X710/XL710/X722 (same page also notes XXV710's speed cannot be forced via ethtool - auto-negotiation only). The companion ice driver page (device_drivers/ethernet/intel/ice.html) is titled for the "800 Series" - the family E810 belongs to - without naming E810 verbatim in the rendered portion, which is stated as the limit of what this specific page confirms rather than stretched into a direct quote. Intel's own ARK product pages did not render in this sourcing pass (see the 100GbE and NIC pages' own note); PCIe generation and EOL status for XXV710 and E810 parts are marked not sourced for that reason.
8. Arista Networks - 7050X and 7050X3 Series switch specifications, and End-of-Sale advisories (arista.com, 2026-09-13)
   Used for: the 7050X family's own published figures: "10GbE Typical Power: 1.2W per port", and for the 7050QX-32S specifically "Typical Power Draw: 150W (4.5W per port)" - the only per-port power figures this cluster publishes, because they are the only ones a vendor datasheet actually rendered and stated. Also Arista's own End-of-Sale advisory archive (arista.com/en/support/advisories-notices/endofsale), which carries a dated 2016 end-of-sale notice for "Select Models of the Arista 7050SX, 7050TX and 7050QX Series" and, separately, 2026-dated end-of-sale notices for the newer 7050SX3-48YC12 and 7050SX3-48YC8/48C8 models - a used 7050-family switch may be running silicon its own maker has since end-of-saled, stated with the notice's own date rather than ours.
9. DatacenterDisk - Live catalogue and sourcing note - 25/40/100GbE cluster (datacenterdisk.com, 2026-09-13)
   Used for: every price in this cluster's tables, read from our catalogue at request time. Seeded 13 September 2026: 12 NICs across 3 models (ConnectX-5, ConnectX-4, Intel E810), 40 DAC cables (25G/40G/100G including breakout), 18 transceivers, 11 used-enterprise and current-SMB switches. Two vendor documentation gaps stated plainly rather than filled: NVIDIA/Mellanox's per-model switch datasheets (SN2010/SN2100/SN2410/SN2700) did not render in this sourcing pass beyond a series-level marketing page, so those four switches carry no sourced power, acoustic or EOL figures; Intel's ark.intel.com did not render at all in this pass, so XXV710 and E810 PCIe generation and formal EOL status are marked not sourced rather than assumed from the 700/800-series family pattern.