STMicroelectronics Ramps Up Silicon Photonics for AI
STMicroelectronics has entered high-volume production of its silicon photonics platform, a key technology for creating ultra-high-speed interconnects. The move directly targets the explosive demand for data center and on-premise AI accelerators, where merging photonic and electrical components is critical for performance.
The move to silicon photonics is a direct response to the physical limits of traditional copper interconnects, which struggle with the bandwidth and power demands of massive AI clusters. As data rates for processors and switches climb to 400Gb/s and beyond, copper connections consume too much power and introduce signal delays, creating a critical performance bottleneck. ST's PIC100 platform is built on 300mm wafers and integrates key optical components like modulators and photodetectors directly onto silicon using established CMOS manufacturing processes. This allows for cost-effective, high-volume production of optical transceivers capable of handling data rates of 800Gb/s and 1.6Tb/s. The company plans to more than quadruple its production capacity by 2027 to meet demand. This technology is a key enabler for co-packaged optics (CPO), an architectural shift where optical I/O is placed within the same package as the main processing chip, like a GPU or switch ASIC. Bringing the optics this close to the processor shortens the electrical path from centimeters to millimeters, drastically cutting power consumption and latency. The silicon photonics market is intensely competitive, with major foundry players like TSMC and GlobalFoundries controlling a significant portion of the market share. Other major competitors include Intel, which has long been a leader in the space, as well as Cisco, Marvell, and IBM, all developing solutions to capture a piece of a market projected to reach over $9 billion by 2030. ST's platform supports per-lane data rates of 200 Gbps, which is essential for building the 800G and 1.6T transceivers demanded by hyperscalers for their AI training and inference clusters. These high-speed interconnects are critical for linking thousands of GPUs, allowing them to function as a single, massive processor. Looking ahead, ST is developing a version of its platform with Through-Silicon Vias (TSVs). This will enable even higher optical density and better thermal management, paving the way for tighter integration of photonics with next-generation AI accelerators and further addressing the power consumption crisis in data centers.
