Kishore Atreya, Senior Director of Cloud Platform at Marvell, presents the deployment challenges and industry standardization needs for co-packaged optics (CPO) at OCP Global Summit 2025. The talk argues that CPO's path to mass deployment hinges not just on optical technology maturity but on solving three system-level problems: field serviceability (especially laser FRUs), system density (shrinking from 4-5U to 1U), and rack-scale fiber management (30,000+ fibers per rack). Marvell demonstrated a 1U liquid-cooled CPO reference system with 1,152 fibers at the summit and called on OCP to drive standardized deployment playbooks across the ecosystem.

Key Takeaways

• CPO is the necessary interconnect technology for AI scale-up domains exceeding ~72 GPUs per rack, driven by the latency and power demands of connecting up to 1,000 accelerators — but deployment, not technology, is now the bottleneck.
• The laser is the primary CPO failure point (confirmed at ECOC 2025); it must be kept field-replaceable (FRU), and the industry needs standardized, simplified access to lasers independent of compute trays.
• Existing market CPO systems are 4-5U air-cooled designs; the industry must achieve a 1U liquid-cooled form factor with smaller connectors (MXC-equivalent connectors are too large) to maintain compute rack density at 72-128+ GPUs.
• A single CPO rack with 128 GPUs running 100T generates ~36,000 fibers; a 1,000-GPU pod requires hundreds of thousands of fibers, necessitating new rack form factors, fiber shuffles, optical backplanes, and PM/SMF fiber management standards.
• Marvell demonstrated a 1U liquid-cooled CPO reference system with 1,152 fibers at OCP 2025 and is pushing for OCP-led standardization of CPO deployment playbooks across connector vendors, fiber vendors, system manufacturers, and silicon/photonics vendors.
• Moving lasers outside of compute trays (beyond ELSFP) is identified as the path to achieving 200T+ system densities by freeing up faceplate real estate.