Tuesday, July 14, 2026
DarkSubscribe
AI Infrastructure · News & Analysis
HomeChips & HardwareReport
Chips & Hardware · Report

Marvell Technology's co-packaged optics (CPO) mass production is delayed to 2029, while NVIDIA's Rubin GPU will continue using all-copper interconnect solutions.

CPO delays push 3D interconnect roadmaps back, favoring conservative thermal approaches and reducing competitive pressure on NVIDIA's proprietary link technologies.
Trade pressSlicast · July 11, 2026 · US · Source: Google News
importance 65

On July 10, Marvell Technology (MRVL) shares declined 3.38% to $235 following reports from SemiAnalysis that large-scale co-packaged optics (CPO) deployment will be delayed until late 2028 or 2029. SemiAnalysis founder Dylan Patel attributed the delay to inadequate manufacturing yields, immature chip designs, and supply chain challenges that have not yet met the standards required for mass production.

Nvidia's upcoming Rubin architecture and its successor Feynman will continue to rely on all-copper solutions, meaning CPO adoption on the GPU side will need to wait several additional generations of chip iterations. Recent design changes further complicate the timeline—notably the removal of the 800V design in Rubin Ultra's Kyber variant has delayed CPO rollout even further.

This extended timeline creates a significant near-term opportunity for copper cable providers. Amphenol and similar connectors manufacturers will benefit more than previously anticipated as copper-based connectivity remains a critical, high-growth bottleneck for AI infrastructure. While CPO represents the long-term industry trajectory and will eventually replace copper cables, the shift in timeline provides a substantial window for copper solutions to generate growth in the short-to-medium term.

The market reversal marks a sharp contrast to sentiment from just one month earlier, when Marvell stock surged over 32% in a single day following remarks by Nvidia CEO Jensen Huang at Computex Taipei, where Huang stated that Marvell was positioned to become the next trillion-dollar market cap company. Huang had emphasized that modern large-model training architecture requires distributed execution across clusters of thousands of chips, with high-speed data sharing between chips directly determining overall cluster computing efficiency and posing a core bottleneck constraining compute power expansion.

Read the original
Marvell Technology's co-packaged optics (CPO)… · Slicast