Choose 400G and 800G optics by working through five things in order: link reach, host platform compatibility, power and thermal headroom in the switch, cable plant already installed, and supply continuity over the life of the cluster. Reach decides the technology. Copper covers the shortest links, multimode covers in-row, and single-mode covers everything beyond that. Everything else follows from there.

Reach is the first filter, not speed

Speed is usually fixed before the conversation starts. The switch and the NIC decide whether a link is 400G or 800G. What is still open is how that link is carried, and reach is what decides it.

Passive direct attach copper handles the shortest runs inside a rack with almost no added power and no added latency. Active copper cables extend that to roughly 5 m at 800G by adding equalisation in the connector heads. Active electrical cables push further again, to around 7 to 10 m at 400G and 800G, using a DSP or retimer in each end to rebuild the signal.

Past that, the link becomes optical. Multimode SR8 modules cover up to around 100 m, which is enough for most in-row switch to server connections. DR-class single-mode modules reach 500 m. FR-class modules, including 800G 2xFR4, reach 2 km and cover building and campus spans.

ATOP builds across all of these, so the decision can be made on the link rather than on what a supplier happens to stock.

Host compatibility decides whether the link comes up at all

A module that meets the MSA electrically can still be rejected by a host that expects specific EEPROM content. This is the most common cause of a link that will not come up, and it has nothing to do with optical performance.

ATOP owns its firmware and does vendor-specific coding in house across major switching platforms. That matters for two reasons. Coding requests do not queue behind a third party, and when a platform changes its firmware expectations, the fix comes from the same team that wrote the original code.

Before ordering, confirm the host platform, the software version in use, and whether that platform enforces a coding check on the port type in question.

Power and thermal headroom are cluster-level constraints

A single module's power draw looks trivial. Multiplied across a 32-port or 64-port switch, and then across every switch in a fabric, it becomes a real line in the power budget and a real constraint on how densely racks can be filled.

Two things follow from that. First, front-panel power affects the cooling design, not just the electricity bill. Second, in dense fabrics the copper links are effectively free power, which is why the copper-versus-optical decision at short reach carries more weight in AI clusters than it did in general-purpose data centers.

Silicon photonics and linear pluggable optics both reduce module power by simplifying the electrical path. ATOP's 800G DR8 LPO silicon photonics variant sits in this category. Ask for measured figures under the thermal conditions you actually run, not best-case bench numbers.

The installed cable plant sets hard limits

The fibre already in the ground shapes what is realistic. OM4 multimode supports SR8 at up to around 100 m. It does not support DR or FR modules, which need single-mode. MPO trunk polarity and connector type also have to match, and getting that wrong is a slow and expensive thing to discover during commissioning.

Where an existing plant is single-mode and correctly terminated, a 400G to 800G upgrade can often be done by changing modules and switches without pulling new fibre. That is worth checking early, because it changes the shape of the whole project.

Supply continuity matters more than unit price in a build phase

AI clusters are built in waves. The modules qualified in wave one need to be available in the same specification for waves two and three, often twelve to eighteen months later.

This is where sourcing model becomes a technical question rather than a commercial one. Products that are built to a fixed design in owned facilities behave the same way batch after batch. Products assembled from whatever is available at the time can drift, and that drift shows up as intermittent link faults in a cluster that was stable at first commissioning.

ATOP designs, manufactures and tests in its own facilities, with component-level traceability and full bills of materials behind every module. Fulfilment runs from hubs in Denmark, the United States and Singapore, which shortens delivery into most build regions.

A working selection sequence

• Measure the actual link length, including slack and cable management routing, not the rack-to-rack distance on the floor plan
• Confirm the host platform, port type and firmware version, and whether coding is enforced
• Check what fibre is installed and how it is terminated
• Calculate front-panel power across the full switch, not per module
• Confirm the operating temperature range required in the deployment environment
• Agree supply continuity and specification lock for the full build programme, not just the first order

Working through those six points removes most of the risk from a 400G or 800G selection. ATOP sells direct to end users, so these questions can be put to the people who designed and built the product rather than passed along a distribution chain.