Technical note: power and cooling for AI data centers
What co-packaged optics measurably does to an AI data center's power bill, with the denominator attached to every number, and the three places a CPO figure usually means something other than you think.
Measured on the Micas M2-W6940-128X1-FR4 (Broadcom Tomahawk 5 Bailly CPO) in a validated design published by Supermicro with AMD and Micas, April 2026.
Measurements, test setup, sources, and a formula for your own fabric.
Get the technical noteFree 14-page PDF. Opens a short form in a new tab.
Vendor power figures usually arrive as one percentage with no denominator. Ours starts with the measurement: two switches on the same Broadcom Tomahawk 5 ASIC at the same 51.2 Tb/s, one with pluggable optics and one with the optics co-packaged in the ASIC package.
Source: Supermicro, AMD and Micas validated design, April 2026. Results are specific to the tested configuration.
Fourteen pages written for the people who have to defend the number: network architects, facilities teams and whoever owns the power budget.
754.85 W vs 465.87 W on the same Tomahawk 5 ASIC, and exactly what changes inside the box to produce it.
Network power becomes heat. At PUE 1.3 the measured delta removes about 376 W of facility load per switch.
Zero flaps across more than 4 million port-hours, and why a stalled collective burns GPU power redoing paid-for work.
Per-switch delta × switches × cooling multiplier, with a worked example and no vendor modeling in between.
Most arguments about CPO savings are really arguments about which layer is being measured.
The biggest number. Measured at the optical interconnect alone, where CPO removes the pluggable module and its DSP entirely. Dramatic, but your power distribution unit doesn't meter optics.
The measured number. The whole chassis, at the plug, against an equivalent pluggable switch on the same ASIC. This is the one to put into a power budget.
The conservative number. What we'll stand behind across port counts, chassis and generations we haven't measured ourselves. Our March 2025 benchmark on fully populated switches measured more than 40% against standard pluggables and more than 25% against LPO.
“Early test results indicate that CPO can deliver not only significant power savings but also reductions in link flapping, which could provide an even greater advantage for AI training workloads.”
Sameh Boujelbene, Vice President, Data Center Switch and AI Networks Market Research, Dell'Oro Group, October 2025
Measured platform: M2-W6940-128X1-FR4. 51.2 Tb/s, 128 × 400G FR4, eight 6.4 Tb/s optical engines co-packaged with a Broadcom Tomahawk 5 Bailly ASIC. Air-cooled, so it fits existing network rows without new plumbing. In volume production since March 2025 and built in Micas's company-operated factory in Malaysia.
In a validated design with Supermicro and AMD, the Micas CPO switch drew 465.87 W against 754.85 W for a pluggable-optics switch on the same Broadcom Tomahawk 5 ASIC: 288.98 W, or 38.28%, less at the plug. At the optics layer alone, savings are larger, around 60–65%. Results vary by configuration.
The Micas CPO switch has logged more than 4 million cumulative 400G-equivalent port-hours without a single link flap. Fewer flaps mean fewer stalled collectives and checkpoint rollbacks during distributed training.
It sums link-hours across many ports, normalized to 400G. It is not wall-clock runtime. It's the right unit for comparing optical reliability, so ask any vendor for the port count behind their "hours of testing."
Our March 2025 benchmark on fully populated switches measured more than 25% power savings against LPO and more than 40% against standard DSP-based pluggables.
No. The Tomahawk 5 generation CPO switch is air-cooled and drops into existing air-cooled network rows. Some Tomahawk 6 generation platforms are designed for liquid cooling.
Multiply the measured 288.98 W per-switch delta by the number of switches in your fabric and by your facility's cooling multiplier, such as PUE. At 1,000 switches and PUE 1.3, that's about 376 kW of facility load removed.
Download the full note, or send us your topology, port counts and PUE and we'll run the arithmetic with you, including the parts that don't favor CPO.
Get the technical note Talk to a Micas engineerPerformance figures reflect the specific validated configuration described in the technical note and will vary by deployment. Cluster-level savings are estimates derived from measured per-switch deltas. Broadcom, Tomahawk, Supermicro, AMD and other marks are the property of their respective owners.