The AI industry's primary bottleneck is transitioning from semiconductor and GPU supply to the interconnects between these powerful chips. While the initial years of the AI race focused on hardware, the current challenge lies in enabling hundreds of thousands of AI-optimized semiconductors to function as a single, enormous computer. This requires robust communication between chips, as the entire system operates at the speed of its slowest connection. The increasing physical distance between chips in growing AI clusters is exacerbating this 'interconnect wall,' as traditional copper wires face limitations in speed, signal degradation, and energy loss.
This shift highlights the importance of photonics, which uses light instead of electricity to transmit data. Optical interconnects offer significant advantages, including higher bandwidth, lower energy consumption, and signal integrity over longer distances. NVIDIA has introduced silicon-photonics switches to connect "million-GPU AI factories," claiming 3.5x better power efficiency, 63x better signal integrity, and four times fewer lasers compared to conventional methods. The development of this technology involves major players like Taiwan Semiconductor, Corning, Lumentum, Coherent, Fabrinet, Broadcom, Marvell, and Cisco, with a new startup, Lumilens, securing over $900 million in funding to address these connectivity bottlenecks.
However, a new bottleneck is emerging: the supply of indium phosphide. This specialty semiconductor is crucial for generating the light within silicon-photonics systems, as silicon itself is inefficient at light production. Lumentum's CEO has warned that indium phosphide production is approximately 30% below customer demand, anticipating a severe supply crunch potentially "worse than memory." This indicates a critical resource bottleneck developing beneath the connectivity bottleneck, as the industry's focus shifts to the materials enabling the communication between AI chips.
