Authors: David Northcutt and David Rostcheck
Date: January 2026
Classification: Policy Recommendation
Distribution: Government Policy Advisors, Department of Energy, Department of Defense, Industry Stakeholders
The United States faces a critical inflection point in the global artificial intelligence race. While AI dominance is recognized as essential to national security and economic prosperity, current infrastructure deployment models create a strategic vulnerability: AI data centers need massive amounts of electricity, but there aren't enough power plants generating it, and even when power exists, connecting to the grid takes 5+ years. Combined, these constraints are extending deployment timelines by 24-72 months, constraining the rapid scaling necessary to maintain U.S. leadership.
This position paper proposes a paradigm shift from treating AI infrastructure as fragile IT systems requiring specialized data centers to adopting an industrial infrastructure model, locating small modular AI data centers at existing industrial facilities such as coal plants. A key element of the proposal is packaging AI compute elements into a standardized hardened shipping container format that can be rapidly manufactured at scale in factories and deployed using standard logistics into industrial spaces. This shift would unlock an estimated 60-85 GW of AI deployment by taking advantage of the available electric generating capacity, space, and permitting at existing U.S. industrial facilities. The resulting 24-72 month acceleration in AI timeline deployment could allow the United States to secure dominance in the AI technology platform.
Energy industry assets such as refineries, coal generating plants, and natural gas turbines, including aging and/or mothballed plants, are particularly attractive because they often have available underutilized onsite electric generation capabilities ("co-gen"), existing grid interconnection, and strong fuel supply chains. They also offer land with zoning and permitting already in place. Synergistically, energy companies are motivated to preserve facilities' useful life, avoiding the high decommissioning cost and keeping them as productive assets. They would further benefit from the privileged regulatory process and protective environmental status afforded to AI projects as critical national security resources.
The industrial model also allows placement onshore at other industrial sites such as idle factories. Hosting AI nodes offshore on container ships offers particular advantages; their infrastructure is already built to support thousands of refrigerated containers, they are built for rapid staging, and can be docked at near-shore docking terminals for network connectivity.
The approach would provide the following strategic advantages:
Secondary benefits would include:
The connectivity model focuses on fiber connections onshore and near-shore to support high-performance AI training and inference workloads. Some batch processing could be possibly handled in remote locations via new low-altitude satellite networks such as Starlink satellite (200+ Mbps, 24-26 ms latency), although this use case is much more limited.
This approach leverages three converging opportunities: