Can air-gapped AI redefine digital deterrence in the age of cyber warfare?

With generative artificial intelligence rapidly becoming central to national security operations, defence contractors are being forced to make a pivotal decision: whether to continue relying on cloud-based infrastructure or shift critical systems into air-gapped environments. Lockheed Martin Corporation has taken a definitive step in that direction. In collaboration with Google Public Sector, the aerospace and defence contractor is embedding Google’s Gemini generative AI models into its AI Factory—an on-premise, air-gapped environment that physically isolates sensitive systems from external networks.

The decision to integrate generative AI into physically secured, mission-assured infrastructure marks more than a tactical upgrade. It signals a strategic reframing of what deterrence means in an age of machine-speed cyber threats, information operations, and algorithmic decision-making. At the heart of this shift is the growing belief that air-gapped AI may offer a credible way to harden national defence systems against both persistent intrusion and catastrophic manipulation.

What is an air-gap and why does it matter for military AI today?

An air-gap refers to a security architecture in which a network is completely isolated from untrusted systems, including the public internet and even internal enterprise networks. In defence contexts, air-gapped environments are traditionally used to protect classified data, nuclear command-and-control systems, and critical infrastructure from cyber intrusion or leakage. Unlike traditional firewalled systems, an air-gapped deployment has no direct connection to external digital domains—making physical access or covert exfiltration the only pathways to compromise.

The relevance of air-gapping for artificial intelligence has intensified with the rise of generative AI models, which often rely on large-scale cloud infrastructure for training, inference, and updates. In commercial contexts, this model is viable and scalable. But in military systems where compromise carries national-level consequences, the idea of running sensitive generative models within internet-connected infrastructure introduces a host of unacceptable risks. These include adversarial prompt injection, poisoned training data, unauthorised surveillance, and model corruption.

Lockheed Martin Corporation’s move to deploy Google’s Gemini models into an air-gapped on-premise environment reflects a recognition that defence systems require more than just AI—they need verifiable, controllable, and secure AI systems that function within known parameters.

How can AI embedded in air-gapped environments shift the deterrence calculus in cyber warfare?

Digital deterrence has always been a more ambiguous and unstable concept than its nuclear-era predecessor. In cyberspace, it is difficult to attribute attacks definitively, demonstrate capabilities credibly without revealing them, or assure retaliation with proportionality. The deployment of AI into this equation introduces both complexity and opportunity.

AI systems embedded in air-gapped environments change the adversarial landscape in two ways. First, they limit the observable surface for reconnaissance or manipulation, reducing the opponent’s ability to model or simulate a successful cyberattack. Second, they increase the defender’s confidence in the integrity of its own decision-support systems, reducing the risk of accidental escalation due to AI error or adversarial spoofing.

From a deterrence perspective, this creates a new kind of asymmetry. A nation-state with reliable, air-gapped AI capabilities can signal technological and informational sovereignty, while also denying adversaries the ability to erode its command-and-control processes through data poisoning or surveillance. The result is a hardened informational perimeter—not just around physical assets, but around the decision-making loops that govern national security responses.

What are the current deployments of air-gapped AI in defence and what do they reveal?

The partnership between Lockheed Martin Corporation and Google Public Sector is among the first high-profile examples of generative AI being deployed in a secure, air-gapped defence environment. In this collaboration, Gemini AI models are being integrated into Lockheed Martin’s AI Factory, with initial deployment limited to unclassified workloads on Google Distributed Cloud infrastructure. However, the architecture is designed to eventually scale into classified systems, reflecting an ambition to fully operationalise air-gapped AI across aerospace, satellite, and cyber mission domains.

Microsoft Corporation has also announced the availability of its generative AI models for air-gapped government cloud environments, including those accredited for U.S. Department of Defense workloads. In Israel and Europe, similar moves are underway, with governments investing in sovereign AI infrastructure that prioritises air-gapped or hybrid isolation layers over cloud-native dependency. While most vendors are still in the pilot phase, the strategic direction is clear: isolation is no longer a fallback—it is becoming the baseline for high-assurance AI deployment.

What risks and trade-offs come with air-gapped AI for cyber operations and deterrence?

Despite the growing adoption, air-gapped AI systems are not without limitations. One of the primary challenges lies in maintaining the currency of models that evolve rapidly. In air-gapped environments, updates to training data, patches to model weights, and performance improvements require physical or tightly audited transfer protocols. This can create lag between innovation and deployment, a gap that adversaries may seek to exploit.

There are also real risks of false confidence. While air-gapped systems reduce the likelihood of external intrusion, they are not immune to insider threats, supply chain manipulation, or covert side-channel attacks. Research has demonstrated that even air-gapped systems can be breached through methods such as electromagnetic emissions, audio leakage, or manipulated peripherals.

Operationally, AI systems in air-gapped settings must also be validated and governed with extreme rigour. Without real-time feedback from networked environments, their outputs can become brittle or misaligned with fast-changing battlefield conditions. This raises the question of how much autonomy should be granted to AI models inside sealed environments—and how accountability is maintained when systems fail without external visibility.

What does this mean for the future of digital deterrence and alliance interoperability?

As air-gapped AI becomes more embedded in national defence strategies, questions are emerging about how these systems will function in coalition or multinational contexts. Defence alliances such as NATO and QUAD increasingly rely on digital interoperability for joint operations, intelligence fusion, and mission coordination. Yet the more isolated and sovereign an AI system becomes, the more difficult it is to ensure alignment, compatibility, or trust across allied platforms.

The emergence of sovereign AI stacks—built entirely within a nation’s borders and shielded from foreign influence—could exacerbate strategic fragmentation. States with robust air-gapped AI infrastructures may develop capabilities that outpace or decouple from their partners, creating asymmetries in response times, targeting precision, or operational doctrine. Conversely, nations without such infrastructure may find themselves dependent on the digital goodwill of stronger allies, limiting their autonomy in cyber-enabled conflict.

From a strategic outlook, the rise of air-gapped AI suggests that digital deterrence will no longer rest solely on offensive capabilities. Instead, it will increasingly depend on the integrity, resilience, and trustworthiness of a nation’s decision systems. In this context, AI is not just a tool—it is a substrate for sovereignty.

Why Lockheed Martin’s air-gapped AI deployment matters for the future of secure innovation

The collaboration between Lockheed Martin Corporation and Google Public Sector may eventually be seen as a blueprint for how generative AI is brought into high-assurance systems without compromising mission integrity. By embedding advanced AI models inside on-premise, air-gapped infrastructure, the American defence contractor is asserting a new standard for how innovation and security can coexist in an era of contested cyberspace.

It also reflects a larger transformation within the defence sector. As AI systems move from experimental sandboxes into operational infrastructure, the need for control, traceability, and resilience becomes paramount. Air-gapping is not a panacea, but in a world where adversaries operate at machine speed and digital misdirection is common, it may be the closest thing to a firewall for the future of national security.