AMD, Nvidia, Meta and Others Launch Open Optical Standard for AI Infrastructure
Six leading artificial intelligence companies announced a new consortium focused on replacing copper cables with optical interconnects inside large AI computing clusters. AMD, Broadcom, Nvidia, Microsoft, Meta, and OpenAI founded the Optical Compute Interconnect Multi-Source Agreement (OCI MSA) group to create an open specification for optical scale-up connections between AI accelerators.
The initiative reflects a growing industry consensus that copper-based wiring is reaching its physical limits as AI models expand rapidly. As a result, data centers face increasing challenges when they attempt to scale computing clusters efficiently.
The OCI MSA specification outlines a technical roadmap for optical connectivity. Initially, the first generation will support 200 gigabits per second per direction. Over time, however, future versions are expected to reach up to 3.2 terabits per second per fiber.
Why the Shift to Optical Connections
Large language models continue to grow in complexity and scale. Consequently, copper interconnects now restrict how many GPUs can connect tightly within a rack or across multiple racks.
Optical connections, however, can transmit data over longer distances while maintaining more predictable power consumption. Therefore, many companies view them as a practical solution for expanding AI infrastructure. Until now, the industry lacked a unified standard for implementing optical links in large-scale AI systems.
The OCI specification relies on non-return-to-zero modulation combined with wavelength division multiplexing. In addition, the approach shifts connectivity from a “module-centric” model toward a “silicon-centric” design.
Furthermore, the architecture supports pluggable optics, on-board optics, and co-packaged optics. As a result, system builders gain flexibility when integrating optical technologies into different hardware configurations.
A Rare Industry Alliance
The consortium stands out because it brings together companies that usually compete aggressively in AI hardware development. For instance, one company’s proprietary NVLink interconnect dominates high-end GPU clusters. Meanwhile, other industry players have supported a competing open-standard effort known as UALink.
However, the OCI MSA focuses on the optical physical layer rather than the interconnect protocol itself. Therefore, it could potentially support both NVLink and UALink traffic over optical fiber instead of copper cables.
Meanwhile, the announcement comes just before a major optical communications conference in Los Angeles. At the event, Broadcom plans to present new technologies including a 400-gigabit-per-lane optical DSP and co-packaged optics solutions. Consequently, the timing highlights how optical interconnects have quickly become central to the future of large-scale AI systems.
