Building Hyperscale Infrastructure for AI Research: Internet2 Designs 138 Tbps Network

By James Deaton - Vice President of Network Services

The infrastructure represents one of the highest-capacity dedicated connections built for AI research. It is designed to accelerate the development of reliable and responsible large language models (LLMs), opening doors to new advancements as AI research grows across the research and education (R&E) community.

The network is purpose-built and operates independently from Internet2’s national R&E backbone. Internet2 designed the infrastructure to provide the ultra-low-latency, high-bandwidth connectivity that distributed AI training and research workloads require. Once complete, it will deliver over 138 Tbps of dedicated total capacity to interconnect compute resources across five U.S. data center locations, spanning more than 3,400 km.

Engineering for Hyperscale

Bringing this ambitious project to life has been a major engineering feat, leveraging Internet2’s deep expertise in high-performance network infrastructure. 

Reaching 57.6 Tbps on a single fiber pair over 300 km required pushing the boundaries of optical networking. Internet2 engineers achieved this target capacity on a dedicated line system and transport equipment that leveraged the full available spectrum on a single fiber pair.

Expanding on that success, the team is now actively working to connect the remaining three data center sites, adding another 80.8 Tbps before the end of the year. The completed AI research network, spanning over 3,400 km and exceeding 138 Tbps, represents infrastructure engineered for workloads that will become more common as universities and research institutions expand their AI capabilities.

Strengthening the AI Ecosystem for R&E

This is more than just a massive network. The collaboration reflects how R&E and industry can work together to address emerging technical challenges while maintaining focus on the R&E community’s needs.

Collaboration between industry and higher education is a long-standing driver for innovation in the U.S. It’s also a shared value of the Internet2 community. 

This network collaboration benefits AI research by advancing the creation of leading production models that require hyperscale infrastructure and networking. Its scale lends new insights into high-performance networking technologies, as Internet2 supports the community with emerging AI challenges and opportunities. Ultimately, collaborations like these strengthen the AI ecosystem for R&E, moving the needle toward more reliable and responsible AI systems — and the hyperscale networks needed to support them.