Nvidia recently committed $4 billion in equity investments to two optical component suppliers, Lumentum and Coherent. The $4 billion equity investments signal a critical shift in how the company plans to power its next generation of AI supercomputers. AI's insatiable demand for data bandwidth is accelerating, but the physical limitations of electrical signaling are creating a bottleneck. Companies that fail to integrate advanced optical interconnect technologies into their AI infrastructure risk falling behind in the race for computational supremacy.
The Looming Bandwidth Wall for AI
Electrical loss in traditional switches can reach 22 dB for 200 gigabit-per-second channels, severely limiting data transfer. At 200G per lane, active electrical cables (AECs) are effective for only 3 meters, making large-scale AI systems impossible to build across multiple racks, according to developer and io-fund. The 3-meter effectiveness limit of active electrical cables creates an insurmountable bandwidth wall for future AI scale with electrical-only solutions.
The global Silicon Photonics Market, projected to grow from US$ 2.62 billion in 2025 to US$ 34.34 billion by 2035 (Datamintelligence), confirms the urgent need for alternatives. Companies relying on traditional electrical interconnects for their AI infrastructure are already operating with obsolete technology, facing an insurmountable bandwidth wall for future scale.
Nvidia's Optical Future: Rubin and CPO Systems
NVIDIA's upcoming CPO-based systems will deliver industry-leading bandwidth of up to 409.6 Tb, slated for commercial availability in 2026 (developer). The 409.6 Tb bandwidth directly addresses the 22 dB electrical loss, confirming optical interconnects as the only viable path for future AI data demands.
The Rubin platform, for instance, will offer NVL72, NVL144, and NVL576 configurations, connecting two and eight racks into a single NVLink scale-up domain (io-fund). Rubin also doubles NVLink scale-up bandwidth versus Blackwell, achieving 3.6 TB/s of bidirectional GPU-to-GPU bandwidth on the sixth-generation NVLink 6 interconnect. Nvidia's aggressive CPO roadmap, with systems delivering 409.6 Tb by 2026, means the race for AI dominance will be won or lost on photonics, leaving competitors without similar optical strategies years behind.
