Can you use third-party optics in NVIDIA switches?

An optical transceiver is not a proprietary part. Pluggable form factors — SFP+, SFP28, QSFP28, QSFP-DD, OSFP and the rest — are defined by Multi-Source Agreements, industry standards that fix the mechanical shape, the electrical interface, and the optical behavior so that modules from different makers are interchangeable in the same cage. That is the entire point of an MSA: a 100G QSFP28 LR4 is a 100G QSFP28 LR4 whoever built it.

So "compatible" is not a fuzzy marketing word here. It means the module matches the port on four concrete axes:

• Form factor — the physical cage (a QSFP-DD module does not go in an SFP port).
• Speed — the lane rate the switch ASIC drives (100G, 400G, 800G).
• Reach and media — multimode vs single-mode, and the distance grade (SR/DR/FR/LR/ER), matched to the fiber that is actually run.
• Coding — the identity the module reports to the switch, covered next.

Match those four and the optic behaves exactly like the branded one, because underneath the label it very often is the same silicon: OEMs rarely fab their own optics — they qualify and rebrand modules from the same handful of manufacturers everyone else buys from.

Every transceiver carries a small EEPROM that the host reads on insertion: vendor name, part number, serial, and the optical parameters (wavelength, reach, power). This is what a switch shows when you inspect a port. Some network operating systems keep a list of "qualified" vendor strings and will flag anything else — a log line, a warning, sometimes an unsupported transceiver state.

The key distinction: on most modern platforms that flag is a warning, not a block — the port still passes traffic — and where a platform is stricter, third-party optics are enabled with a single configuration setting, or the module is simply coded to report a vendor string the switch already trusts. Coding to a platform is a routine, above-board practice; it is how the entire compatible-optics market operates, and it is why a reputable third-party module ships specified for a given switch family rather than generically.

The practical takeaway is not "will any optic work" but "is this optic coded for my switch." That is a per-module, per-platform fact to confirm before you buy — not a reason to pay the OEM premium by default.

The most repeated reason to avoid third-party optics — "it voids the switch warranty" — does not survive contact with the law in the United States. The Magnuson-Moss Warranty Act prohibits a manufacturer from conditioning a product's warranty on the use of its own branded parts or supplies unless those parts are provided free of charge. A switch vendor cannot lawfully void your hardware warranty simply because a third-party optic is plugged into it.

The honest caveat is about support, not warranty. If you open a case and the vendor suspects the optic, you may be asked to reproduce the issue with a qualified module — normal first-line triage, not a warranty forfeiture. And a cheap, poorly-made optic that misreports its power budget is a real operational risk. The answer to that is buying quality modules from a source that stands behind them and confirms the coding — not paying a 5–10× brand premium across thousands of ports.

The failure mode that strands a cluster is almost never the vendor on the module. It is a reach or media mismatch — the one spec people skim.

• Multimode vs single-mode. SR/VR modules run over multimode fiber (OM3/OM4/OM5) for tens of metres inside a rack or row; DR/FR/LR modules run over single-mode for hundreds of metres to kilometres. Put an SR optic on a single-mode run, or the reverse, and the link does not come up.
• Distance grade. Even within single-mode, DR (~500 m), FR (~2 km) and LR (~10 km) are different parts. Under-spec the reach and the link is marginal or dark; over-spec it and you overpay per port — which, multiplied across a fabric with thousands of optics, is real money.
• Breakout. An 800G port often breaks out to 2×400G or 8×100G, and the optic and cabling have to match the intended split.

Get the reach and media right and a correctly-coded third-party module in the correct form factor is, functionally, the branded part at a fraction of the cost. Get it wrong and the most expensive OEM optic in the catalog still will not light.

Optical transceivers are one of the few genuinely commodity layers of an AI fabric: MSA-standardized, spec-searched, and available across the full range — 1.25G through 800G, in OSFP, QSFP-DD, QSFP28, SFP28 and SFP+, across the SR/DR/FR/LR/ER reach ladder, plus coherent DCO for metro spans.

We source that full line, matched to the switch and NIC platforms in your fabric. Because coding and reach are per-link facts, the exact module — including whether it is coded for your specific switch family — is confirmed at quote rather than asserted from a spec sheet. Browse the networking catalog, and if you are still weighing the transport underneath, InfiniBand vs Ethernet for AI covers the fabric choice the optics ride on.