Interview: Marvell says AI's copper wall is nearing, with custom silicon and optical I/O set to scale

Marvell’s aggressive stance at Computex 2026 signals a pivotal inflection in AI infrastructure: copper interconnects are hitting physical limits, forcing a systemic shift from electrical to optical I/O across data centers. This triggers cascading redesigns—from silicon photonics and laser components upstream to advanced packaging downstream—favoring foundries with co-packaged optics capabilities. Geopolitically, its executive visits to Taiwan, China underscore deep reliance on TSMC’s packaging capacity, heightening supply chain vulnerability; any U.S. export control escalation could delay custom optical chip deliveries by over 30%. Facing Broadcom’s vertically integrated stack and NVIDIA’s rumored in-house optical engines, Marvell’s bet on open interconnect standards is a strategic play for market access through technical neutrality. Within 18 months, optical I/O will migrate from premium AI clusters to mainstream servers, unlocking a multi-billion-dollar market—but only players mastering both SerDes IP and photonic integration will clear the cost barrier.