Meta detailed a network switch with 102.4 terabits a second of capacity in an Oct. 26 engineering publication outlining infrastructure supporting its latest artificial intelligence systems, the company said.
Meta said AI progress now depends as much on networks as computing power. It published “Meta Networking at the 2026 OCP Global Summit: Open Innovation for Scale-Up, Scale-Out, and Scale-Across AI Networks” on Oct. 26.
The company, which operates the AI infrastructure it described, drew on engineering work and deployment experience presented at the Open Compute Project summit. Its accounts covered hardware, software and connections between data centers.
Meta said its networking projects support its Muse Wonder and Muse Spark models and its gigawatt-scale Prometheus cluster. The publication supplied no common measurement period or comparative dataset for the projects.
The IceCube switch fits into Meta's liquid-cooled Icepack rack, company engineers Rongchun Zhou and Reza Niazmand said. Its design combines liquid and air cooling and supports chips from either Broadcom or Nvidia.
The engineers said the switch provides 128 connections operating at 800 gigabits a second each. Its capacity allows those connections to operate simultaneously without competing for bandwidth inside the switch.
Meta engineers Granthana Rangaswamy, Siamak Amiralizadeh and Xu Wang described testing co-packaged optics, which place optical connections close to switching chips, at 51.2 terabits a second. Lessons could apply to the higher-capacity design.
Optical engineers Amin Yekani and Qing Wang described a tradeoff in removing signal-processing chips from individual optical modules. That removes a repeated source of power use, heat and cost but shifts signal adjustment and troubleshooting to host equipment.
Systems engineers Yao Rong and Karthik Ganesan described reliability limits in copper connections within racks. Even a 99.9% manufacturing yield can leave roughly 10 faulty links among thousands of connections.
The engineers said cartridge designs without access to individual connections require a statistical approach to testing. Their account treated failures across the system rather than as separate troubleshooting tasks.
Meta and other contributors proposed a common switch platform design to reduce duplicated engineering work across suppliers. Proprietary platform designs add months of work, and the group contributed the proposal's first specification to the Open Compute Project.
Meta software engineers Ashkan Nikravesh and Yiorgos Nikolaidis said the company is releasing implementations of its MetaRoCE networking software as open source. An initial version is available on GitHub, with another version planned.
Meta's vice president of core network infrastructure, Vipul Deokar, and Cisco executives Rakesh Chopra and Tom Williams said training demand is growing faster than infrastructure can expand. They argued for open networks connecting AI capacity across data centers and regions.
