NSF gives Northeastern $5.1 million to test wireless links in AI data centers

The National Science Foundation has awarded Northeastern University $5.1 million to develop sub-terahertz wireless links for AI data centers, where cable sprawl is becoming a major bottleneck. The three-year project brings together university researchers, Teradio and industry partners including Dell, NVIDIA, Lockheed Martin and Keysight to demonstrate the system in real operational environments.

Why it matters: - AI data centers are hitting a physical limit as millions of cables clog dense GPU racks, raise costs and complicate cooling and airflow. - The Northeastern-led project aims to replace or augment some of that cabling with sub-terahertz wireless links that can move hundreds of gigabits per second. - If the approach works, AI operators could get more flexible network design, lower infrastructure friction and easier scaling for future GPU clusters.

What happened: - The National Science Foundation awarded Northeastern University $5.1 million for a three-year Verticals-enabling Intelligent Network Systems, or VINES, Track 2 project. - The program is designed to develop and demonstrate sub-THz wireless infrastructure inside operating AI data centers. - The award brings in Northeastern University, Florida International University and Teradio, along with Dell, NVIDIA, Lockheed Martin and Keysight Technologies.

The details: - The project targets both out-of-band management networks and primary data fabric traffic. - The system is intended to integrate with standard data center network interfaces for deployment in existing AI data center infrastructure. - Sub-THz wireless links are suited to short in-rack distances, where connections measure in meters rather than kilometers. - Narrow, electronically steerable beams are expected to provide physical-layer security and spatial isolation similar to wired links. - Northeastern’s Josep Miquel Jornet leads work on system architecture, channel modeling and beamforming. - Northeastern’s Tommaso Melodia and Salvatore D'Oro are focused on AI-driven spectrum orchestration, intelligent network control and network slicing. - Florida International University’s Konstantinos Zekios and Arjuna Madanayake are leading antenna arrays, RF front-end design and chiplet development using heterogeneous integration. - Teradio, a startup commercializing sub-THz radio technology from Northeastern’s Institute for Intelligent Networked Systems, is providing the commercialization path. - Teradio’s Adaptive Wireless Infrastructure platform is designed to support a physically isolated out-of-band management plane, traffic offload from primary training interconnects and software-defined on-demand wireless links. - Dell Technologies is supplying AI factory environments for pilot deployment, live traffic data and engineering support for rack integration. - NVIDIA is advising on GPU cluster networking requirements and traffic patterns from large-scale AI training workloads. - Lockheed Martin is supporting validation in defense-relevant environments where secure, high-bandwidth wireless connectivity matters. - Keysight Technologies is providing measurement, characterization and protocol testing tools for the sub-THz frequency range. - The project will be validated in real-world operational environments, not just in the lab. - The release includes project information at the Northeastern AI factories site and Teradio information at Teradio.

Between the lines: - The funding signals growing interest in using wireless links to solve a networking problem that conventional cable plants struggle to scale. - The partner list shows the project is aimed at deployment, not only academic study. - The inclusion of both AI data centers and defense-relevant environments suggests the technology could have uses beyond a single market.

What's next: - The team will build, test and validate the wireless system over the three-year award period. - Industry partners will help provide the operational environments and measurement tools needed to prove the technology at scale. - The central question is whether sub-THz links can be made reliable, efficient and practical enough for production AI infrastructure.