Artificial Intelligence Infrastructure on DOE Lands

Introduction

I am pleased to submit the following recommendations in response to the Department of Energy’s Request for Information regarding the development of artificial intelligence (AI) infrastructure on DOE-managed lands.

This submission draws on national security, energy resilience, and AI risk expertise, including the President’s Council of Advisors on Science and Technology (PCAST) Cyber-Physical Resilience Report and testimony from AI and cybersecurity leaders.

I commend the department’s forward-leaning approach to fostering secure, high-performance AI infrastructure through public-private collaboration. As the DOE advances its initiative to deploy AI-focused data centers’ high-performance computing environments and associated energy, cooling, and transmission systems, it is imperative that resilience and cybersecurity are embedded as foundational elements — not afterthoughts. I recommend the following actions:

1. Mandate secure-by-design principles for all infrastructure;
2. Promote co-located, cyber-secure on-site power systems;
3. Invest in adversarial AI defense capabilities;
4. Require cyber-physical contingency planning; and
5. Establish shared AI governance frameworks.

Mandate Secure-By-Design Principles for All Infrastructure

DOE should require all AI infrastructure projects to implement secure-by-design practices throughout the system lifecycle. This includes mandating Software Bills of Materials (SBOMs) to ensure visibility into third-party and open-source components, thereby reducing the risk of hidden vulnerabilities or malicious code. Beyond software, protection must extend to device-level components — particularly sensors and battery energy storage systems (BESS) — which often lack authentication or encryption protocols, making them prime attack vectors, as demonstrated in recent energy-sector compromises.

Promote Co-Located, Cyber-Secure On-Site Power

DOE AI infrastructure should be paired with on-site, cyber-hardened energy generation and storage systems. Modular nuclear reactors, advanced geothermal, and solar-plus-storage configurations provide energy independence and operational continuity in grid disruption scenarios. These systems should integrate secure digital control layers and segmented communications pathways to prevent lateral movement of cyber intrusions. Co-siting AI workloads with resilient energy infrastructure also enables real-time load balancing, islanding capabilities, and grid services — transforming these sites into testbeds for cyber-physical resilience.

Invest in Adversarial AI Defense Capabilities

DOE should establish national facilities for red-teaming AI systems in infrastructure environments. These testbeds would support evaluations of model robustness against adversarial inputs, data poisoning, and synthetic manipulation. Standardized benchmark datasets representing operational scenarios — such as fault detection, demand response, or anomaly classification — would ensure consistency of performance and security assurance. Inference protocols should enforce runtime integrity and track data provenance, especially for models used in control or predictive maintenance applications.

Require Cyber-Physical Contingency

DOE should incorporate requirements for dual-hazard scenarios into all site and vendor proposals. As the PCAST report warns, cascading failures across energy and digital systems — triggered by extreme weather or coordinated cyberattacks — are increasing in likelihood and severity. Infrastructure developers must demonstrate resilient strategies for operating in degraded or islanded states, including manual fallback modes, local autonomy for distributed energy resources, and incident response playbooks tailored to cyber-physical risk. This planning must also consider dependency mapping for HVAC, fire suppression, sensor networks, and remote management systems.

Establish Shared AI Governance Frameworks

AI systems powering DOE infrastructure must be transparent and accountable. An AI-BOM — analogous to an SBOM — would provide visibility into model architecture, training datasets, and external dependencies. Provenance frameworks should log when and how models were retrained, what data was used, and how inferencing outcomes are interpreted. These practices are critical for preventing AI model drift, ensuring reproducibility, and supporting secure handoffs between developers, operators, and regulators. DOE can lead this governance evolution by integrating these requirements into its infrastructure solicitations and pilot deployments.

Conclusion

As DOE advances its AI infrastructure initiative, it is crucial to regard resilience and security as essential design parameters — on par with performance and efficiency. DOE sites offer a unique opportunity to showcase how AI infrastructure can be developed with national security, innovation, and energy leadership in mind. By adopting these recommendations, DOE can not only improve operational resilience but also establish a global standard for secure AI deployment in complex cyber-physical environments.

I appreciate the opportunity to contribute to this strategically vital effort and remain available for further engagement in shaping a secure, resilient, and globally competitive AI infrastructure ecosystem.