Data Center Cybersecurity Against Nation-State Operations
Architecture, Technology, Assurance and Cost for 100–300 MW Campuses
- Type
- Working paper and companion to Data Center Design, Resilience, Reliability and Sustainability. It covers what a state actor holding valid administrator credentials can reach on a reference campus, how long it stays, which rules reach the facility, and the configuration and cost of a program built to limit the consequence.
- Main finding
- A 100–300 MW campus should be designed on the assumption that a state intelligence service will at some point hold valid administrator credentials inside its management networks. Because a defense cannot depend on knowing the technique, the paper derives eight design principles that limit consequence instead.
- Scope
- The systems a campus operator controls: facility controls, management planes, the network edge, the supply of hardware, firmware, software and services, and operator-run compute up to the tenant boundary. The legal analysis is United States federal law, with Virginia and three foreign comparators.
- Keywords
- Data center security; critical infrastructure; nation-state intrusion; operational technology; confidential computing; supply chain assurance; grid reliability; NERC CIP.
- JEL classification
- L86, L94, H56, K23, O33.
- Revision history
- Version 1.0, 23 October 2026. Posted on SSRN 31 October 2026.
A 100–300 MW data center campus should be designed on the assumption that a state intelligence service will at some point hold valid administrator credentials inside its management networks. This paper assembles the public record behind that assumption and works out what follows for architecture, technology, assurance and cost. The documented entry methods since 2021 are few and they repeat: internet-facing routers, firewalls and VPN gateways; identity keys and tokens with broad scope and long life; appliances that cannot run endpoint detection; and suppliers who hold access to their customers. Dwell times are measured in years rather than days, and exploitation on average precedes the patch, so patching cannot be the main control for an exposed system.
The facility systems are weak and now matter to the grid, while no cyber rule reaches the facility as a data center: the July 2026 FERC order, the NERC alerts and the three computational-load standards that passed final ballot in October 2026 contain no cyber security requirement beyond a secure-transfer method. Confidential computing does not remove the operator from the trust boundary; five published attacks between December 2024 and September 2026 defeated AMD and Intel protections with brief physical access and inexpensive hardware.
Because a defense cannot depend on knowing the technique, the paper derives eight design principles that limit consequence instead, sets out a recommended configuration and implementation sequence, and models the cyber and control-system program of a stand-alone 200 MW campus at 28–29 security staff and $9–14 million a year.
Assume a state service will hold valid credentials, and design to limit what it can do. Patching cannot be the main control for an exposed system, no cyber rule reaches the facility as a data center, and confidential computing does not remove the operator from the trust boundary. The paper sets out eight design principles, a recommended configuration and implementation sequence, and models the cyber and control-system program of a stand-alone 200 MW campus at 28–29 security staff and $9–14 million a year.
Read the paper
What this paper does not establish
The conclusions rest on the pattern across many sources and not on any single figure. Section 15 discloses open items that must be closed before the analysis is used in a filing, a contract, an engineering certification or an investment decision.
The paper separates opened primary sources, secondary sources, and the author’s own estimates and inferences. The zone targets, engineered limits, log-retention period, staffing and cost model, the reading of NERC applicability, and the scheduler and vendor-concentration arguments are the author’s. The definitions of ISA/IEC 62443 security levels are quoted from vendor copies, not the standard’s text.
The July 22, 2026 Dominion-zone load transfer was not a cyber event, and whether a commanded transfer would produce the same voltage and frequency excursion has not been modeled in public. The cost figures are a declared model, not observed costs.
Suggested citation
George, Collin B. Data Center Cybersecurity Against Nation-State Operations: Architecture, Technology, Assurance and Cost for 100–300 MW Campuses. Sanctir Working Paper, WP-2026 series, version 1.0, 23 October 2026. Licensed CC BY-NC-ND 4.0. SSRN 7671058.