Nvidia's OOI architecture paves way for optical interconnects in GPUs

Nvidia's OOI architecture—embedding optical engines on the silicon interposer—is a paradigm shift from copper-dominant to opto-electronic hybrid compute. Technical cascade: Sharing the interposer with GPU dies structurally breaks copper's bandwidth-power ceiling. Silicon photonics foundries (GlobalFoundries, UMC in Taiwan, China) face an order-mix reset. The HBM memory wall eases as optical I/O bandwidth jumps, but package-level thermal complexity escalates exponentially. Supply chain risk: Interposer capacity concentrates in advanced packaging lines in Taiwan, China; InP/GaAs epitaxial wafers scatter across Japan and Europe. U.S. export controls extending to silicon photonics would choke the final packaging mile of China's AI self-sufficiency push. Competitive dynamics: AMD's MI350 will accelerate its 3D-stack + optical I/O roadmap. Broadcom and Marvell fight for custom ASIC optical interconnect outside Nvidia's orbit. Intel's Foveros, if mass-produced by 2026, compresses Nvidia's first-mover window. Outlook: Within 18 months, optical interconnect shifts from optional to mandatory. Copper retains a cost moat below 50 cm, but at 10K+ GPU scale, opto-electronic hybrid packaging is the only viable density path. For the first time, advanced packaging houses—not fabless designers—hold pricing power in the AI chain.