VINES: Track 2: Untethering AI Factories: From Cable-Bound Networks to Adaptive Wireless Infrastructure

Artificial intelligence (AI) training facilities, known as AI factories, depend on tens of thousands of servers connected by millions of cables. Managing this infrastructure requires a separate out-of-band (OOB) management network that operates independently of the main computing fabric. As AI clusters scale beyond 100,000 processors, the cable management network becomes a critical bottleneck: costly to install, rigid to reconfigure, and vulnerable to physical failures in shared cable pathways. This project develops wireless networking technology operating in the sub-terahertz (sub-THz) frequency band (110-170 GHz) to replace portions of the cable management infrastructure with a software-defined wireless fabric. Success would reduce data center cabling complexity, accelerate deployment timelines, and enable new management capabilities impossible with fixed cables. Beyond AI factories, the technology has broader applications in smart manufacturing, defense computing, and modular data centers. The project trains graduate students in next-generation wireless systems and partners with industry leaders to accelerate technology translation and workforce development. The project advances U.S. competitiveness in both next generation (NextG) telecommunications by leading the development and use of sub-THz radio technologies and AI infrastructure by demonstrating alternatives to traditional cabling approaches that will support scaling of AI factories. The project is organized in three thrusts executed by a multi-institutional team spanning Northeastern University (NU) and Florida International University (FIU), and commercialization partner Teradio Inc., with industry collaboration from Dell, NVIDIA, Lockheed Martin, and Keysight. Thrust 1 develops prototype sub-THz transceivers targeting over 100 Gbps in the D-band (110-170 GHz), including wideband antenna arrays, analog frequency-division multiplexing channelizer chiplets in 28nm Complementary Metal-Oxide-Semiconductor (CMOS), real-time digital signal processing using neural network-based linearization, and an analog circuit-switched wireless network. Thrust 2 creates the intelligent networking layer: statistical channel models for sub-THz propagation in dense metallic rack environments, physical layer designs exploiting structured electromagnetic beams for robust links, dynamic spectrum allocation, and task-based network slicing for isolated management domains. Thrust 3 validates the complete system through a phased demonstration plan progressing from channel characterization at NU through multi-node coordination to deployment across multiple production racks at an operational AI factory, replacing selected OOB connections under real workloads. Teradio leads the technology translation pathway, packaging prototypes into adoption-ready integrated units for pilot customer deployment and establishing a commercialization roadmap toward industry-standard form factors. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.