Nvidia Bets $4B on Photonics for AI

The investment in photonics addresses a critical, and expensive, problem in AI development: GPUs are often idle. In some multimodal AI environments, as much as 84% of GPU computing power is wasted due to inefficiencies in legacy infrastructure, with GPUs sitting idle up to 40% of the time while waiting for data. This bottleneck stems from the physical limitations of copper wiring, which struggles to keep up with the massive parallel processing demands of large-scale AI. As AI models grow, the sheer volume of data that needs to move between thousands of GPUs creates a "networking wall." Traditional copper interconnects consume too much power and introduce latency, which is the delay in data transfer. This latency can significantly slow down AI model training, effectively doubling the time it takes to complete and underutilizing the expensive GPU resources. Silicon photonics offers a solution by using light to transmit data, which is significantly faster and more energy-efficient than electrical signals over copper. This technology allows for the creation of high-bandwidth, low-latency connections between chips, which is essential for feeding the massive appetite of modern GPUs. The market for silicon photonics is expected to grow dramatically as a result of the demands of AI. Nvidia's $4 billion investment is split evenly, with $2 billion each going to Lumentum and Coherent, two key players in optical technologies. The deals are not just for research and development; they also include multi-billion dollar purchase commitments and securing future production capacity, ensuring Nvidia has a stable supply chain for these critical components. Coherent has a 20-year relationship with Nvidia, highlighting the long-term strategic importance of this partnership. This investment will directly support the development of "co-packaged optics" (CPO), which involves integrating optical components directly with silicon chips in the same package. This tight integration is a key enabler for the next generation of AI hardware, promising significant reductions in power consumption and signal loss. The fruits of this investment will be seen in Nvidia's upcoming networking platforms. The Spectrum-X Photonics Ethernet switches, expected in 2026, and the Quantum-X Photonics InfiniBand switches will leverage this technology. These platforms are designed to create a high-speed, efficient fabric connecting tens of thousands of GPUs, a critical step towards building the "AI factories" of the future.