Best 100G QSFP28 Transceivers and Adapters

100GBASE-SR4 is the short-reach standard this category is built on. Four lanes at 25G each, multimode fiber at 850 nm, an MPO-12 connector, and a rated reach measured in tens of metres rather than kilometres — which is exactly right for a link that stays inside a row. PSM4 is the parallel single-mode answer for the same MPO connector when the run leaves the row, and it uses four single-mode fibers instead of four multimode pairs. Both are parallel optics, and both are why a 100G project buys MPO trunks and cassettes rather than duplex patch cords.

The trap is that a 1G or 10G module uses duplex LC and a 100G SR4 module does not, so a team that has cabled a rack for years without thinking about connectors will find the fiber is the blocker. Retrofitting MPO into an existing duplex plant is a cabling project, not a parts order. Establish which connector the fiber terminates in before the transceiver line is priced, and if the answer is duplex LC, the conversation moves to breakout cassettes or a different standard rather than a cheaper module.

Spine and leaf switches are bought for fabric bandwidth, and a build almost always ends up with more 100G cages than 100G endpoints. A QSFP28-to-SFP28 adapter closes that gap for the price of a cable: seat the adapter, seat a 25G module, and the port serves a server NIC instead of sitting empty. It is the single most useful cheap part in this category and it is routinely left off a bill of materials because nobody thinks to ask for it.

The alternative — a 4x25G breakout from one QSFP28 into four SFP28 endpoints — is a different part and a different topology, and it is the right answer when four servers sit close enough to share one cage. Both approaches beat buying a second switch to terminate a handful of 25G links. Decide which one the rack layout supports before the port count is finalised, because the harness lengths are fixed and short.

Optics are rarely the biggest line on a fabric order and they are almost never the longest-lead one. A 100G build carries the switches themselves, the operating-system and feature licensing those switches need, MPO trunks and cassettes sized to the row, the patch panels they land in, structured cabling labour, and the staging and configuration work that turns a pallet into a running fabric. On an AI or GPU cluster, add the InfiniBand side and the management network, both of which have their own optics.

Uniqcli quotes those together — hardware, licensing, and the deployment and lifecycle services — on one document rather than leaving half of it to a second purchase order. For a build of any size the practical route is the bill-of-materials upload rather than a cart: it prices the whole fabric at once, keeps the country-of-origin and §889 screening attached to every line, and produces a single quote a programme can actually approve.