Industry backs open optical links (OCI)
A new multi‑source Optical Compute Interconnect (OCI) agreement—co‑founded by Nvidia with AMD, Broadcom, Meta, Microsoft and OpenAI—aims to standardize high‑density optical links for AI data centers, a big step toward interoperable, low‑latency AI infrastructure reported. That matters because software and model architects will need to design for hardware heterogeneity and next‑gen interconnect performance.
The Optical Compute Interconnect Multi‑Source Agreement (OCI MSA) was formally established on March 12, 2026 by AMD, Broadcom, Meta, Microsoft, NVIDIA and OpenAI [announced]. finance.yahoo.com The initiative’s public materials lay out an initial 200G PHY and GEN1/GEN2 interface plan that targets 200–400 Gbps per direction (up to ~800 Gbps per fiber) with a technical roadmap that scales toward 3.2 Tbps class links. oci-msa.org The OCI physical specification specifically calls out dense wavelength‑division multiplexing (DWDM) and cascaded micro‑ring resonators (MRR) as the optical building blocks to reach higher bandwidth density and lower power per bit. oci-msa.org Revision history in the published 200G OCI Optical PHY Specification v1.0 shows an editorial team including AMD, Broadcom and Meta and a document date of March 11, 2026, indicating an immediate public technical baseline for vendors to implement. oci-msa.org Earlier industry demos — for example Intel’s OCI optical I/O chiplet shown at OFC 2024 — illustrate the co‑packaged optics trend the OCI MSA seeks to standardize, signaling vendor roadmaps toward chiplet‑to‑chiplet optical I/O. opticalconnectionsnews.com A practical systems project tied to these specs: run NCCL collective benchmarks (NVIDIA’s nccl‑tests) on multi‑node setups while emulating OCI GEN2 800 Gbps links to measure all‑reduce scaling and latency sensitivity, using the OCI PHY numbers as the target link parameters. github.com
