The United States defense and aerospace establishment is taking meaningful, if measured, steps to apply artificial intelligence to missions traditionally handled by human analysts and manual systems.
With competitors like China pouring resources into AI-enabled warfare, the U.S. can not afford to tinker cautiously at the edges.
A new era of AI in national security appears to be taking shape — and the question is no longer if the defense sector will scale up its use of AI, but how quickly it can manage to.
OpenAI recently won a $200 million contract with the Department of Defense, a deal that is significant for its size and scope, as it tasks OpenAI Public Sector with building prototype AI solutions tailored to national security needs.
In addition, the Pentagon earlier this year expanded its commitment to Project Maven, substantially raising the ceiling on Palantir’s contract with U.S. combatant commands from $480 million to nearly $1.3 billion through 2029. Maven has long been one of the Defense Department’s most visible AI programs, applying machine learning to satellite and drone imagery to detect threats in real time. After years of incremental progress, Maven is now entering a critical scaling phase.
On the space industry side, United Launch Alliance is piloting “RocketGPT,” a version of OpenAI’s chatbot technology hardened for defense use. The secure bot is running on Microsoft’s ITAR-compliant Azure Government Cloud, giving engineers a tool that boosts productivity without compromising sensitive data. About 150 employees are already testing it at ULA.
This deployment represents one of the first applications of generative AI inside the classified boundaries of the defense industrial base. Unlike consumer-grade chatbots, RocketGPT meets the strict standards that govern sensitive aerospace and defense information. ULA CEO Tory Bruno said its real value lies in how it augments human expertise without replacing it. Engineers can use it to accelerate analysis, automate technical writing and surface critical data faster.
Northrop Grumman, meanwhile, is aiming higher — literally. The company is working with NVIDIA to bring AI-driven simulations into the realm of spacecraft autonomy. Han Park, vice president of AI Integration at Northrop Grumman Space Systems, said the goal is to build “cognitive spacecraft” that have heightened awareness of their environments and the ability to automatically generate and execute commands and mission commands, maneuvering with limited human interaction, even when there is uncertainty.
“This allows an operator to interact with the spacecraft system by giving it high-level goals rather than detailed tasks,” Park said.
The military-industrial complex moves slowly by design, with elaborate testing and validation processes that can take years. Meanwhile, AI technology evolves on Internet time, with major breakthroughs occurring monthly rather than annually.
Lockheed Martin’s AI Fight Club represents one attempt to address the problem of speed. It creates a “digital proving ground” where AI systems can be tested rapidly across all domains of warfare — air, land, sea and space. This experiment reflects the reality that AI for warfare needs to be tested and validated before deployment.
Lockheed’s initiative is also a window into emerging innovation outside the traditional defense contractor network — an attempt to bring fresh AI talent into the fold while giving the Pentagon a front-row seat to evaluate what’s viable and what’s not.
Together, these efforts point to a growing consensus that AI is a necessity even though, as Bruno pointed out, it should not be something you “sprinkle on products” and assume will be improved.
The defense enterprise doesn’t need to wait for perfect AI. It needs usable AI, today, deployed in targeted ways that respect operational constraints but don’t let bureaucracy paralyze innovation. That means accelerating deployments in secure clouds. It means backing programs that can bridge the gap between Silicon Valley and the front lines. And it means rewriting the playbook for AI acquisition, which, despite recent announcements, is still bogged down by outdated processes built for hardware, not algorithms.
This article first appeared in the July 2025 issue of SpaceNews Magazine.