From Megabits to Milliseconds

This paper argues that with fiber and gigabit cable now reaching most US locations, bandwidth is no longer the binding constraint on internet performance—latency is. The authors contend that latency is determined largely outside the last mile, by middle-mile fiber, peering density, and proximity to Internet Exchange Points (IXPs), so the massive public investment in last-mile deployment does little to address it. They distinguish AI training, which follows power supply, from AI inference, which follows network performance. And they argue that inference, not training, will determine which regions benefit economically from AI. The stakes rise, they say, as AI moves into the physical world: augmented reality, robotics, and agentic systems require ultra-low, reliable latency to nearby GPU capacity, and regions without dense interconnection will see these applications arrive late or not at all. Their policy recommendation is to direct a share of the roughly $21 billion in remaining Broadband Equity, Access, and Deployment (BEAD) Program funds toward middle-mile routes, IXPs, and regional GPU-as-a-Service capacity. Readers should note that Connected Nation co-owns IXP.US, a joint venture that builds the very facilities the paper recommends funding, and that the paper provides no citations for its key empirical claims—the authors themselves concede latency remains poorly measured and mapped.

[Note: Connected Nation co-owns Connected Nation Internet Exchange Points, LLC (d/b/a, IXP.US), a 50/50 joint venture with carrier hotel pioneer Hunter Newby to develop neutral Internet Exchange Point (IXP) facilities across the United States.]