Data Center Warfare: Defending the Key Terrain of AI Infrastructure - Modern War Institute

Data Center Warfare: Defending the Key Terrain of AI Infrastructure

Jason Vogt and Nina A. Kollars | 06.09.26

Share on Facebook<br>Share

Share on TwitterTweet

Share on LinkedIn<br>Share

Send email<br>Mail

Print<br>Print

The rapid expansion of AI-driven data centers is driving permanent changes to the geographical layout of critical infrastructure that serves as the backdrop of global competition and future wars. This includes the construction of digital megacampuses, which will become strategic high value targets, and the proliferation of new data center hubs, which could drive changes to defense planning scenarios, particularly in the Indo-Pacific.

In the wake of the United States’ and Israel’s February attack on Iran’s military and senior leaders, Iran retaliated with missile and drone strikes against US military bases and civilian infrastructure across the Middle East. While much of the world’s attention has focused on the disruption of regional energy production, the attacks also struck three Amazon Web Services (AWS) data centers in the UAE and Bahrain, disrupting digital services to banks, payment platforms, and other entities. A month later an Oracle data center in Dubai was damaged in another attack. Iran then declared eighteen major technology companies as legitimate military targets. Included in this list were many of the world’s largest data center owners, including AWS, Microsoft, Meta, Google, and Oracle.

For Iran, targeting data centers served multiple strategic ends within a single effort. Hitting those centers allowed it to punish economically vital US companies, as well as the regional organizations that host data on those companies’ servers, and served as an unambiguous threat that further attacks on digital infrastructure were likely if Iran’s demands were not met. But these strikes indicate that a broader strategic shift is already underway.

For adversaries looking to impose costs, signal resolve, or disrupt military operations, data centers make attractive targets. They are large, fixed sites that are costly to build. They are also dependent on local power, water, and data transmission infrastructure, which can also be targeted. Attacking them not only harms the data center operator, but all the organizations that rely upon it for data storage, networking, or AI integration. This has the potential to compound effects, as multiple entities experience degradation or loss from a single action.

Data Centers and the AI Catalyst

Data centers are specialized facilities designed to house data servers and networking equipment used by businesses and governments to manage their digital operations. While they appear as nondescript warehouses from the outside, their interiors contain vast hallways of server racks, requiring specialized power and cooling equipment designed to maintain optimal conditions for the digital machinery housed inside. They draw power from the local energy grid but typically contain industrial-scale backup generators and uninterrupted power supplies to maintain continuous energy flows to the equipment during blackouts or other power interruptions. Data centers are essential to the modern technology and economic landscape, providing data backups for enterprises, enabling hyperscalers to offer cloud services, and supporting AI computing. There are now more than ten thousand data centers operating worldwide, a number that is likely to grow significantly in the coming years.

Data centers that support AI models have substantially different requirements and architecture than those used as backups or traditional cloud computing services. Non-AI data centers are designed to run servers comprised of central processing units, or CPUs, and a traditional non-hyperscale data center could require upwards of twenty megawatts of power to run its operations, the equivalent of fifteen to twenty thousand US homes. Cloud data centers, which are run by hyperscalers like AWS, Google and others, are larger but follow the same principles.

AI data centers, by contrast, require densely packed clusters of graphical processing units, or GPUs, built by companies like Nvidia. Training AI models can use clusters of over 100,000 GPUs, requiring 150,000 megawatts. Soon, a single AI training run could require up to a gigawatt of power, enough to power a city. Once a model is built, it can be deployed onto computing infrastructure used to support user queries, which requires about half the power of a training facility.

To enable these operations many existing conventional data centers are being overhauled to support a mix of CPU, GPUs and specially designed AI processing chips called NPUs (neural processing units). Each rack can require over one hundred kilowatts of power (compared to five to fifteen kilowatts per CPU rack) and cooling these systems requires even more power generation. Some larger data centers...