The physical layer is an energy system. Every watt of optical power that enters a switch leaves as heat that must be removed. At AI cluster scale, this is not a minor concern -- it is a primary design constraint that cascades into data centre infrastructure, cooling system capacity, and operational costs.
The switch power arithmetic
A 64-port 400G switch with QSFP-DD modules: 64 ports x 3.5W per module = 224W in optics Switch ASIC (25.6 Tbps tier) ~= 250W Total switch power: ~500W
A 64-port 800G switch with OSFP modules: 64 ports x 14W per module = 896W in optics Switch ASIC (51.2 Tbps tier) ~= 500W Total switch power: ~1,400W
The step from 400G to 800G nearly triples switch power -- from 500W to 1,400W -- because optics scale faster than ASIC power. The ASIC roughly doubles power when you double capacity. The optics nearly quadruple (4x more ports at 800G vs 400G if you consider that 800G replaces two 400G switches in some designs, but at roughly 4x per-port power).
The BasePOD power profile
A DGX BasePOD consists of:
- 32 DGX H100 nodes: each draws ~10.2 kW under full GPU load Total DGX power: 32 x 10.2 kW = 326 kW
- 8 leaf switches (400G): 8 x 500W = 4 kW
- 32 spine switches (400G): 32 x 300W = 9.6 kW (spine switches are smaller -- 8-port per switch)
- Storage switches, management switches, out-of-band: ~5 kW
Total BasePOD power: approximately 345 kW in a standard DGX H100 deployment.
Distributed across the racks:
- 4 DGX compute racks + 2 switch racks
- Average power density per rack: 57 kW/rack
This is why AI clusters require direct liquid cooling, rear-door heat exchangers, or high-density air-cooling with elevated airflow rates. Standard enterprise rack cooling (which handles 5-15 kW/rack) is completely inadequate for AI cluster infrastructure.
When you upgrade from 400G to 800G switching in a future refresh:
- Leaf switch power: 8 x 1,400W = 11.2 kW (up from 4 kW)
- Additional cooling load from switching alone: 7.2 kW
This 7 kW delta must be accommodated in the data centre's cooling design. It is not an afterthought -- it must be specified in the building infrastructure before the hardware is installed.
What CPO and LPO actually solve
The motivation for CPO and LPO is partly cost and partly thermal.
In a CPO switch, the DSP's heat is managed by the switch ASIC heatsink -- an already-engineered, already-paid-for thermal solution. There are no dozens of individual hot pluggable module heads radiating heat into the switch airspace. CPO switches can potentially deliver the same throughput as pluggable switches at significantly lower total power.
Early prototypes suggest CPO can reduce per-port power by 30-50% compared to pluggable equivalents. At the scale of a hyperscale AI cluster with thousands of switch ports, a 40% reduction in per-port switching power represents tens of megawatts of data centre power capacity that does not need to be built.
This is the economic case that is driving CPO from academic research to production deployment. It is not primarily about cost per module. It is about infrastructure capacity -- the ability to deploy more compute per megawatt of facility power.