New optical standards for AI
Industry groups are advancing optical interconnect standards (OCI MSA and XPO MSA) to overcome copper limits and support 200G+ links over longer distances for GPU clusters, with the market projected to grow substantially. That reflects a hardware‑level networking trend driven by AI training scale and data‑center topology. (x.com)
Most of the wires inside today’s artificial intelligence servers are still copper, and copper is great until the distance gets longer and the speed gets absurd. At 200 gigabits per second and above, the signal fades fast, the power bill climbs, and the cleanest links are often the shortest ones. (oci-msa.org) Optical links solve that by turning data into light, which is easier to push across longer paths without the same heat and loss penalties. The tradeoff is that optics have usually cost more and come in too many vendor-specific flavors for giant graphics processing unit clusters to mix and match easily. (oci-msa.org) That is why a new group called the Optical Compute Interconnect Multi-Source Agreement was launched on March 12, 2026. Its founding members are Advanced Micro Devices, Broadcom, Meta, Microsoft, Nvidia, and OpenAI, and the group says it wants one open optical specification for artificial intelligence “scale-up,” which is the traffic inside a tightly coupled graphics processing unit cluster. (oci-msa.org) The first public document from that group is a 200 gigabit optical physical-layer specification released on March 11, 2026. It describes dense wavelength-division multiplexing, which is the trick of sending multiple colors of light down the same fiber so one strand can carry more traffic, and it uses micro-ring resonators, which are tiny on-chip light filters that act like lane markers for those colors. (oci-msa.org) This is aimed at the part of an artificial intelligence system that behaves less like the public internet and more like one giant computer. When thousands of graphics processing units split one training job, every delay between chips slows the whole group, so the network inside the cluster starts to matter almost as much as the chips themselves. (ultraethernet.org) A separate effort called the Ultra Ethernet Consortium is working on the software and protocol side of that same bottleneck. Its Specification 1.0, released on June 11, 2025, defines an Ethernet-based stack for artificial intelligence and high-performance computing, while Optical Compute Interconnect is focused lower down on the physical link that actually moves the bits as light. (ultraethernet.org) Nvidia’s current rack-scale systems show why this fight is happening now. The DGX Grace Blackwell rack-scale guide says the graphics processing units are connected through NVLink passive copper cable cartridge backplanes, which works inside a rack but also shows how much current artificial intelligence design still depends on short-reach copper. (docs.nvidia.com) The next step is to push those fast links beyond the backplane without paying a giant penalty in watts or complexity. That is where other industry work, including the External Laser Small Pluggable optics effort often shortened to XPO, fits in: moving lasers and optics into more modular forms so operators can stretch high-speed links farther across boards, trays, and racks. (lightcounting.com) Market researchers already see the money moving in that direction. LightCounting said in July 2025 that optical connectivity for artificial intelligence scale-up networks would contribute to market expansion through 2030, after a surge in demand from 2023 through 2025. (lightcounting.com) That means the story is no longer just “faster chips.” The companies building the biggest clusters are now standardizing the roads between chips, because a training system with 10,000 graphics processing units is only as fast as the links that keep all 10,000 moving together. (oci-msa.org; ultraethernet.org)
