Investment in AI data center development is growing exponentially: in June 2025, Amazon announced a $20 billion investment to develop AI data center campuses in Pennsylvania alone, and in July 2025, Meta announced that its first multi-gigawatt data center, Prometheus, will be online in 2026.
US political support for AI data centers has also removed regulatory hurdles for companies and administrators, as President Trump's new AI Action Plan encourages AI technology stacks and data center development in the US and abroad.
See also: How is AI reshaping cybersecurity businesses?
One problem that almost all stakeholders have identified is the increased energy demand – nearly 612 terabytes of electricity over the next five years – and the associated worsening of global warming – a 3-4% increase in global carbon emissions. However, the less publicized challenge is the increased cybersecurity threats facing AI data centers, which pose increased reputational, financial, and regulatory compliance risks for administrators and business users.
Like traditional data centers, AI data centers contain hardware, network, storage, data, and software, making them targets for common cyberattacks: distributed denial-of-service (DDoS) attacks, ransomware, supply chain attacks, and social engineering attacks. Data centers are also known for their vulnerability to side-channel attacks because data center hardware, from fans to central processing units (CPUs), can reveal sensitive information about CPU-level activity, architecture, and data usage.
Compared to traditional ones, AI data centers face an expanded set of threats due to differences in hardware, data, and purpose. While large data centers use CPUs and graphics processing units (GPUs), AI data centers always use GPUs because AI workloads require more computing power and because GPUs enable parallel operations.
See also: Oracle: Investments in AI and cloud infrastructure in Germany and the Netherlands
Application-specific integrated circuits (ASICs) and field-programmable gate arrays (FPGAs) are also powerful hardware that can be adapted to efficiently compute and process AI workloads. Google has invented an ASIC specifically for AI and deep learning called the Artificial Intelligence Processing Unit (TPU). These more powerful resources are vulnerable to side-channel attacks like CPUs: in January 2025, a TPU-specific side-channel attack, TPUXtract, was discovered that exploits data leaks and allows threat actors to infer the parameters of an AI model.
In addition to common cyberattacks and side-channel attacks, GPUs are more vulnerable to memory-level attacks because GPUs do not always have adequate memory isolation. Memory can be mistakenly transferred from one process to the next, giving threat actors the opportunity to gain access to information about AI model weights and training data. There is also GPU-specific malware that can execute malicious code in a GPU’s memory and bypass traditional CPU security tools.
In terms of data, because AI data centers host AI models, weights, and training data, AI data centers face risks of model extraction, loss of sensitive data, and model-level threats. Leakage of AI model information compromises the integrity and confidentiality of the model. Model-level threats include data poisoning attacks and model poisoning attacks that can corrupt the model, as well as model reversal attacks and model theft attacks that reveal information about the model and its training data. Compromised models can lead to biased and erroneous results that impact customer operations.
See also: 1.6 million people affected by data breach at Laboratory Services Cooperative
Finally, given the importance of AI to national security and economic competitiveness, the global race to develop AI capabilities and secure AI dominance – a country’s ability to develop, use, and manage AI models and supporting infrastructure – has begun.
