In AI rack cabling, five things matter more than the rest: keeping network power out of a power budget that is already tight, avoiding cable bulk that restricts airflow, matching each link to the shortest technology that covers it, keeping links serviceable while the rack is running, and locking the specification so later racks match the first.

An AI rack is a different cabling problem to a general compute rack. The power budget is tighter, the thermal margin is thinner, the link count is higher, and the cost of taking the rack down for service is measured in stalled training jobs.

Power headroom is already spent before the network is considered

GPU racks tend to arrive at the facility power limit on compute alone. Whatever the network draws comes out of that same allocation, and then again out of the cooling budget.

The practical consequence is that copper should be the default at short reach. Passive DAC adds essentially nothing. ACC and AEC add small amounts in exchange for reach. Optics draw the most, which is correct when distance requires it and wasteful when it does not.

Cable bulk is a thermal decision

Thick bundles behind a switch restrict exhaust airflow and raise temperatures across the chassis. Higher module temperature means higher module power, which produces more heat. In a rack already running warm, that loop matters.

Thinner assemblies at the same reach are worth the small premium. So is spending time on bundle geometry rather than treating cable management as a finishing task.

Map the links before choosing products

AI racks have a small number of repeating link types. Server to top of rack in the same cabinet. Server to switch in a tall or shared cabinet. Adjacent rack. Row level. Leaf to spine.

Each maps cleanly to a technology. Under passive DAC reach, use passive DAC. To around 5 m, ACC. To around 7 to 10 m at 400G and 800G, AEC. Beyond that, AOC or pluggable optics with structured fibre. ATOP builds all of these in 400G and 800G, which allows the map to be filled in properly rather than compromised.

Serviceability while the rack runs

A rack running paying or production workloads cannot be powered down for a cable swap. That affects layout more than most designs allow for.

Fixed assemblies such as DAC, ACC, AEC and AOC are cheaper and lower power but mean a failure replaces the whole cable. Pluggable optics on patched fibre cost more and draw more but allow a module to be swapped without disturbing the cable plant. Mixing the two deliberately, by criticality rather than by convenience, is usually the right answer.

Specification lock across phases

AI capacity is built in waves. Rack fifty should behave like rack one, which means the cable qualified in phase one has to be available in the same build in phase three.

That depends on the supplier being a manufacturer rather than an assembler. ATOP designs, builds and tests in its own facilities with component-level traceability, and sells direct, so a specification can be locked and reordered without drifting between batches.