CPO & LPO: The Next Leap in Network Evolution presented by Micas Networks

This OCP Global Summit 2025 executive panel, moderated by a Micas Networks representative and featuring SemiAnalysis, Micas Networks, and Multilane, examines the state of Co-Packaged Optics (CPO) and Linear-Drive Pluggable Optics (LPO) as next-generation networking technologies for AI data centers. The panel reviews Meta's landmark CPO field data — including millions of hours at 400G, 5x performance vs. 3x optics, and an expected 65% operational power savings — and debates commoditization, serviceability, TCO, and market penetration trajectories. Panelists broadly agree that CPO adoption is inevitable given physical limits of LPO at 400G per lane, with the debate centered on timing, penetration rate, and the chicken-and-egg dependency between ecosystem scale and TCO justification.

Key Takeaways

Meta's CPO field data from ECO showed millions of operating hours at 400G, 5x better performance vs. 3x optics, improved mean time between failure, and an expected 65% operational power savings — shifting at least one major hyperscaler holdout to a pro-CPO position.
CPO reduces per-rack networking power from ~8 kW (pluggable) to 1.5-2 kW on a GB300 NVL72 rack (~150-160 kW total), yielding 10-20% networking power savings; LPO is an intermediate step at ~6 kW.
LPO is expected to work at 200G per lane but not at 400G, making CPO adoption a matter of 'when, not if' for the next switch generation; the 1.6T generation is forecast at 50-60% LPO and 10-30% CPO.
CPO penetration for the current 51.2T generation is optimistically forecast at 25-30% by Micas Networks; NVIDIA's active push for CPO is the single largest driver of industry adoption.
Supply chain commoditization concerns are real: CPO's tighter integration limits the field to fewer optical suppliers compared to the 10-15 players available today for pluggable optics.
CPO yield risk is concentrated at the Broadcom silicon assembly level; Micas Networks reports 100% revenue conversion on every Broadcom silicon module received.
Optimal CPO deployment strategy starts at the highest network aggregation layer and migrates down through tier 1 and tier 2 as the technology matures; clean-room manufacturing capability is required for switch vendors.
Jensen Huang has publicly suggested CPO could collapse AI training networks from three tiers to two — a potentially transformative architectural impact that is the key prize the industry is watching for.