800G and 1.6T Optics: What Changes When You Go Above 400G

800G and 1.6T Transceivers, OSFP vs QSFP-DD

Fifteen times more optical engines by the middle of 2027

At GTC Berlin on 21 October, NVIDIA set out what its supply chain is being asked to build. Alongside a 2.5x expansion in substrate capacity and more than 300 percent additional interposer capacity, it named a fifteenfold increase in optical engine capacity with SPIL by the middle of 2027, and a further doubling targeted by the end of 2028. Read that as a statement about where the pressure sits. Substrates and interposers are about building accelerators. Optical engines are about connecting them. When the company designing the architecture expands optics faster than almost anything else in its supply chain, it is telling you which part of the build is going to be hard to get. For anyone specifying a cluster rather than manufacturing one, the practical question is narrower. Above 400 gigabit, the rules that governed optics for the last decade start to bend. This piece is about what actually changes.

The form factor stops being a detail

At 10 and 25 gigabit, the module format was a question of what fitted the switch. SFP+ and SFP28 were interchangeable in the sense that nobody agonized over the choice. At 400 gigabit and above there are two serious formats and the difference between them is thermal rather than electrical. QSFP-DD doubles the electrical lanes of the older QSFP while keeping backward compatibility with it, which means a QSFP-DD port can usually take a QSFP module. That compatibility is valuable in a mixed estate and it is the main reason QSFP-DD is widespread in general data center switching. OSFP is slightly larger and was designed from the start with higher power in mind. Many OSFP modules carry integrated heat-dissipating fins on the module itself rather than relying entirely on the cage and the switch's airflow. It is not backward compatible with QSFP without an adapter. That size difference is the whole argument. A module at 800 gigabit draws meaningfully more power than one at 400, and a module at 1.6 terabit more again. All of that power becomes heat inside a cage packed alongside thirty-one others across the face of a switch. The format that can shed it is the format that can run at the top speeds reliably, which is why 1.6T has gone to OSFP while 800G and 400G are available in both. The practical consequence for a buyer is that the format is set by the switch and is not a preference. Check the port, check whether the chassis expects finned or flat-top modules, and specify accordingly. A thermally mismatched module will link and then throttle or fail under sustained load, which is the worst kind of fault because it passes commissioning.

Breakout becomes the normal case, not the exception

At 40 and 100 gigabit, breakout was a useful trick for getting more ports out of a switch. In an AI cluster it is closer to the default topology. The reason is that the fabric has layers with different speeds, and a high speed port feeding several lower speed endpoints is simply how the tiers connect. A reference design for a modern GPU server hall illustrates the pattern: the highest speed modules sit between switches on the InfiniBand leaf to spine links, the next tier down connects the leaf switches to the network cards in each GPU server, and the tier below that handles north-south Ethernet and fabric management, often breaking out to half-speed links for the data processing units. Beneath all of that sits a much slower management layer running to management switches, firewalls, the internet edge and out-of-band controllers. That means a single cluster is not specified with one optic. It is specified with five or six speeds at once, each with its own count, each with its own cable, and the whole thing has to arrive together. Missing the 10 gigabit modules for the out-of-band network delays a hall just as effectively as missing the 1.6T links, and the out-of-band modules are the ones that get forgotten because they are not interesting. Breakout cabling follows the same discipline covered in our 40G and 100G guide: the parallel optic needs an MPO, the breakout harness fans it to the endpoint format, and the switch has to support the mode.

Compatibility, testing and the thing worth asking

Above 400 gigabit the gap between a module that meets the multi-source agreement on paper and a module that works in a specific switch gets wider, because there is less thermal and signal margin to absorb variation. Three things matter. The module should be built to MSA and original vendor specification, which covers the mechanical dimensions, the pull tab and the heat dissipating fins as well as the optics. It should have been tested against the device it will sit in, not just certified in the abstract. And the test should ideally be a link and interoperability test in a deployed cluster rather than a bench result. The mistakes repeat. Specifying by speed alone and discovering the port wanted the other form factor. Ordering the headline modules on a long lead time and the management layer optics casually, then waiting on the cheap ones. Assuming a module qualified in one vendor's switch will behave in another's. And treating optics as a commodity line at the exact moment the industry is expanding optical capacity fifteenfold because it is not one. The question worth putting to any supplier is simple. Has this exact module been tested in this exact switch, and can you tell me where. There is one more thing worth building into the schedule, which is what happens when a module fails. At these speeds a replacement is not something most sites hold on a shelf, and the lead time on a single spare is rarely better than the lead time on the original order. A small percentage of spares bought with the main order costs very little against the cost of a hall waiting. The figure people usually land on is somewhere between one and three percent per speed tier, weighted toward the tiers that are hardest to source rather than spread evenly. The other thing worth knowing is where the industry is going, because it changes how long a design stays current. Co-packaged optics move the optical engine onto the switch package itself, removing the pluggable module and most of the electrical path to it. That is a real power and density gain and it is clearly coming, which is what the capacity expansions are being built for. But pluggable optics do not disappear when it arrives. They remain how a hall is serviced, how a failed link is fixed in an hour rather than a quarter, and how an estate with mixed equipment ages gracefully. A design that assumes pluggables stay relevant for the life of the building is a safer design than one that does not.

Specifying a hall rather than buying a box

The reason optics go wrong is rarely the optic. It is that a cluster needs six speeds, several form factors, matching cable for each, and a single delivery date, and that list usually gets assembled by more than one person. Vocom supplies that list as one conversation. The optical transceiver range runs from 1.6T OSFP and 800G in OSFP and QSFP-DD down through 400G, 100G, 25G, 10G and 1G, built to MSA and original vendor specification, tested for compatibility and quoted within 24 hours, alongside the low-loss MPO, MTP and LC patch cords they terminate into. Vocom does not manufacture the modules; they are sourced through its manufacturing partner. Vocom is not affiliated with or endorsed by NVIDIA. If you have a cluster drawing where the optics line still says a speed and a quantity, that is the line worth finishing. Send the switch list.