A group of tech giants, including AMD, NVIDIA, and OpenAI, has launched the Optical Compute Interconnect (OCI) consortium to create a new open standard for AI infrastructure.
This initiative marks a pivotal shift from traditional copper wiring to advanced optical fiber for connecting GPUs within and across server racks. The core motivation behind this move is a direct response to two critical challenges facing the AI industry today: soaring energy consumption and the physical limits of copper.
First, the energy demand from AI data centers is exploding. Both the U.S. Energy Information Administration (EIA) and the International Energy Agency (IEA) have pointed to AI as a primary driver of recent surges in electricity demand. As companies build ever-larger AI models, the power required for networking components becomes a significant operational cost and a strain on the power grid. OCI aims to tackle this head-on by championing technologies like Co-Packaged Optics (CPO), which can reduce the power consumption of network switches by as much as 70%.
Second, the technological foundation for this transition is now mature. The Ethernet ecosystem, which forms the backbone of data center networking, has evolved significantly. Standards like the Ultra Ethernet Consortium (UEC) Spec 1.0 and advancements in 224G-per-lane signaling provide a stable, high-performance base upon which an optical standard can be built. Furthermore, key players like Broadcom and NVIDIA have already demonstrated the readiness of high-density optical engines and CPO technology, signaling that the industry is prepared for this change.
Essentially, copper cables are hitting a wall. They are limited to short distances (typically 2-3 meters for 800G speeds) and become bulky, restricting airflow and complicating rack design. Optical fibers overcome these distance and density limitations. The OCI consortium's goal is to create an interoperable, multi-vendor standard that makes this superior technology accessible and cost-effective for building the next generation of AI supercomputers.
- Co-Packaged Optics (CPO): A technology where optical connectivity components are placed on the same package as the main processing chip (like a switch ASIC). This proximity significantly reduces power consumption compared to traditional pluggable optical modules.
- SerDes (Serializer/Deserializer): A pair of functional blocks commonly used in high-speed communications to convert data between serial and parallel interfaces. It's a fundamental component in data transmission within chips and between them.