AI Is Headed to the Battlefield, and Defense Contractors Are Racing to Rewrite How War Is Fought

Deep inside a classified U.S. military outpost sitting 50 miles off the Mexican border—its precise coordinates remain tightly under wraps—defense contracting firm Anduril is trialing an unexpected, groundbreaking application of large language models. I got a front-row seat to one of the earliest public demonstrations of the technology last year.

Standing on a sun-faded landing strip, I watched as four jet aircraft codenamed Mustang popped up on the western horizon, soaring over a barren expanse dotted with boulders and dry scrub. The prototype jets, scaled down for the demonstration, quickly locked into a tight formation, their engines humming louder as they drew closer to the observation point.

The sun glared straight into my eyes, so I turned my attention to a nearby computer monitor tucked under a dust-caked tarp. After a few quick keystrokes, a fifth aircraft popped up along the edge of the screen—its outline matched that of China’s J-20 stealth fighter almost exactly. A young team member named Colby, decked out in a black baseball cap and sunglasses, issued a clear command to engage the simulated enemy threat: “Mustang intercept.” That’s when the AI took over. A large language model similar to the one that powers ChatGPT processed the natural language command, coordinated directly with the drones, and replied in a steady, unemotional female voice: “Mustang collapsing.” Roughly 60 seconds later, the four drones had converged on the hostile target, and with barely any fanfare, destroyed it using virtual ordnance.

Anduril’s demonstration lays bare just how aggressively the global defense sector is racing to test and deploy new AI capabilities. Through its Fury initiative, the startup is currently developing a full-size autonomous combat jet for the U.S. Air Force, built to fly alongside manned fighter aircraft. While many existing military drone systems already operate autonomously thanks to older generations of AI technology, the core goal of integrating large language models into the command chain is to streamline order relay and surface critical, actionable information directly to human pilots. It’s not hard to think of it as having a noncommissioned chatbot on your squad roster.

This whole concept sounds jarring at first, but cutting-edge defense technology almost always does. For decades, the U.S. has poured endless funding into a mix of game-changing innovations and questionable, wasteful projects. With LLM-powered military AI, the core promise is greater operational efficiency: modern military kill chains are notoriously convoluted, and AI is theoretically meant to cut through that complexity (a polite euphemism for making strike operations far deadlier). Top U.S. four-star strategists argue that whichever nation gains control of this technology will dominate global military power for generations. That mantra is why the U.S. has worked so aggressively to cut off China’s access to cutting-edge AI development, and it’s also why the Pentagon plans to dramatically ramp up its own AI investment in the coming years. The push is dramatic, but far from unexpected: the war in Ukraine, where low-cost, computer-vision-equipped drones are now ubiquitous across front lines, has already proven just how much autonomous systems reshape modern combat.

The recent generative AI boom has only amplified this interest across the defense establishment. A 2024 report from the Brookings Institution found that funding for AI-focused federal contracts grew 1,200% between August 2022 and August 2023, with the vast majority of that funding coming from the Department of Defense. That growth all happened before former President Donald Trump returned to the White House. His current administration is now pushing for even more targeted, large-scale investment in strategic military AI: its $1 trillion 2026 defense budget—rebranded explicitly as a “war” budget—includes the first ever dedicated allocation for AI and autonomous systems, clocking in at $13.4 billion.

This landscape means enormous upside for AI companies that can overpromise on what their technology can deliver for combat operations. Just this year, Anthropic, Google, OpenAI, and xAI all secured AI-focused military contracts worth up to $200 million each. The shift marks a dramatic 180-degree turn from 2018, when Google made headlines for pulling out of Project Maven, a Defense Department initiative to use AI for analyzing aerial surveillance imagery. Emelia Probasco, a Georgetown University researcher who studies military AI deployment, notes that Project Maven—now operated by Palantir—has grown into Maven Smart Systems, one of the U.S. military’s most widely deployed AI tools in active use today.

The growing adoption makes perfect sense, she explains: large language models are uniquely well-suited for intelligence work, because they excel at sorting through and synthesizing massive quantities of unstructured text and data. They also fit naturally into cyber offensive operations, thanks to their advanced ability to write and analyze code. “The part of this ambition that’s really unsettling is the idea that AI is so capable it can either stop war from breaking out, or just fight and win it for us,” Probasco says. “It’s treated like some kind of magical fairy dust that solves every problem.” For the time being, though, today’s leading LLMs are still too unreliable, prone to critical errors, and too opaque in their decision-making to be trusted to make independent life-or-death battlefield calls, or given direct control over any active military hardware.

A core ongoing challenge for firms working in this space is figuring out how to deploy AI in ways that play to the technology’s strengths while minimizing the risk of catastrophic failure. In September of this year, Anduril and Meta jointly submitted a bid for a U.S. Army contract worth up to $159 million to develop yet another AI-integrated military tool: a durable augmented reality display built into combat helmets for frontline soldiers. Anduril says the system, which will push mission-critical data to warfighters while also continuously sensing and mapping their surrounding environment, will run on a new generation of more capable AI models that can interpret the physical world in real time far more accurately than existing tools.

So what about the end goal: fully autonomous robotic frontline soldiers? I spoke with Michael Stewart, a former fighter pilot who once led the U.S. Navy’s disruptive capabilities office and helped spearhead AI experimentation for the Fifth Fleet back in 2022. Stewart now runs an independent defense consulting firm and advises military planners around the world. He expects future warfare will be dominated by heavily automated systems.

“In 10, 15, or 20 years, you’re going to have robots that operate with pretty significant autonomy,” he says. “That’s the direction we’re moving, no question.” And if these systems use large language models as their core processing “brain,” they won’t just be a new kind of actor that witnesses the horrors of war. They’ll also be able to explain, in plain, natural language, exactly what actions they took, and why they chose to take them.