One afternoon in late November, I visited a weapons test site in the foothills east of San Clemente, California operated by Anduril, a maker of AI-powered drones and missiles that recently announced a partnership with OpenAI. I went there to witness a new system it’s expanding today, which allows external parties to tap into its software and share data in order to speed up decision-making on the battlefield. If it works as planned over the course of a new three-year contract with the Pentagon, it could embed AI more deeply than ever before into the theater of war.
Near the site’s command center, which looked out over desert scrubs and sage, sat pieces of Anduril’s hardware suite that have helped the company earn its $14 billion valuation. There was Sentry, a security tower of cameras and sensors currently deployed at both US military bases and the US-Mexico border, and advanced radars. Multiple drones, including an eerily quiet model called Ghost, sat ready to be deployed. What I was there to watch, though, was a different kind of weapon, displayed on two large television screens positioned at the test site’s command station.
I was here to examine the pitch being made by Anduril, other companies in defense-tech, and growing numbers within the Pentagon itself: A future “great power” conflict—military jargon for a global war involving competition between multiple countries—will not be won by the entity with the most advanced drones or firepower, or even the cheapest firepower. It will be won by whichever entity can sort through and share information the fastest. And that will have to be done “at the edge” where threats arise, not necessarily at a command post in Washington.
“You’re going to need to really empower lower levels to make decisions, to understand what’s going on, and to fight,” Anduril’s CEO Brian Schimpf says. “That is a different paradigm than today,” where information flows poorly among people on the battlefield and decision makers higher up the chain.
To show how it will fix that, Anduril walked me through an exercise of how its system would take down an incoming drone threatening a base of the US military or its allies (the scenario at the center of Anduril’s new partnership with OpenAI). It began with a truck in the distance, driving toward the base. The AI-powered Sentry tower automatically recognized the object as a possible threat, highlighting it as a dot on one of the screens. Anduril’s software, called Lattice, sent a notification asking the human operator if he would like to send a Ghost drone to monitor. After a click of his mouse, the drone piloted itself autonomously toward the truck, as information on its location gathered by the Sentry was shared to the drone by the software.
The truck disappeared behind some hills, so the Sentry tower camera that was initially trained on it lost contact. But the surveillance drone had already identified it, so its location stayed visible on the screen. We watched as someone in the truck got out and launched a drone, which Lattice again labeled as a threat. It asked the operator if he’d like to send a second attack drone, which then piloted autonomously and locked onto the threatening drone. With one click, it could be instructed to fly into it fast enough to take it down. (We stopped short here, since Anduril isn’t allowed to actually take down drones at this test site.) The entire operation could have been managed by one person with a mouse and computer.
Anduril is building on these capabilities further by expanding Lattice Mesh, a software suite that allows other companies to tap into Anduril’s software and share data, the company announced today. More than 10 companies are now building their hardware—everything from autonomous submarines to self-driving trucks—into the system, and Anduril has released a software development kit to help them do so. Military personnel operating hardware can then “publish” their own data to the network and “subscribe” to receive data feeds from other sensors in a secure environment. On December 3, the Pentagon’s Chief Digital and AI Office awarded a three-year contract to Anduril for Mesh.
Anduril’s offering will also join forces with Maven, a program operated by defense-data giant Palantir that fuses information from different sources like satellites and geolocation data. It’s the project that led Google employees in 2018 to protest against working in warfare. Anduril and Palantir announced on December 6 that the military will be able to use the Maven and Lattice systems together.
The aim is to make Anduril’s software indispensable to decision makers. It also represents a massive expansion of how the military is currently using AI. You might think the Department of Defense, advanced as it is, already has this level of connectivity between its hardware. We have some semblance of it in our daily lives, where phones, smart TVs, laptops and other devices can talk to each other and share information. But for the most part, the Pentagon is behind.
“There’s so much information in this battle space, particularly with the growth of drones, cameras and other types of remote sensors, where folks are just sopping up tons of information,” says Zak Kallenborn, a warfare analyst who works with the Center for Strategic and International Studies. Sorting through to find the most important information is a challenge. “There might be something in there, but there’s so much of it that we can’t just set a human down and to deal with it,” he says.
Right now, humans also have to be the translator between systems made by different manufacturers. One soldier might have to manually rotate a camera to look around a base and see if there’s a drone threat, and if they find one, they have to manually send information about it to another soldier operating the weapon to take that drone down. To do so, they might use a low-tech messenger app—one on par with AOL instant messenger—to share instructions. That takes time. It’s something the Pentagon is attempting to solve through its Joint All-Domain Command and Control plan, among other initiatives.
“For a long time, we’ve known that our military systems don’t interoperate,” says Chris Brose, former staff director of the Senate Armed Services Committee and principal adviser to Senator John McCain, who now works as Anduril’s chief strategy officer. Much of his work has been convincing Congress and the Pentagon that a software problem is just as worthy of a slice of the defense budget as jets and aircraft carriers. (Anduril spent nearly $1.6 million on lobbying last year, according to data from Open Secrets, and has numerous ties with the incoming Trump administration: Anduril founder Palmer Luckey has been a longtime donor and supporter of Trump, and JD Vance spearheaded an investment in Anduril in 2017 when he worked at venture capital firm Revolution.)
Defense hardware also suffers from a connectivity problem. Tom Keane, a senior vice president in Anduril’s connected warfare division, walked me through a simple example from the civilian world. If you receive a text message when your phone is off, when you turn the phone back on, you’ll see the message. It’s preserved. “But this functionality, which we don’t even think about,” Keane says, “it doesn’t really exist” in how many defense hardware systems are designed. Data and communications can be easily lost in challenging military networks. Anduril says its system instead stores data locally.
The push to build more AI-connected hardware systems in the military could spark one of the largest data collection projects that the Pentagon has ever undertaken, and one that companies like Anduril and Palantir have big plans for.
“Exabytes of defense data, indispensable for AI training and inferencing, are currently evaporating,” Anduril said on December 6, when it announced it would be working with Palantir to compile data collected in Lattice, including highly sensitive classified information, to train AI models. Training on a broader collection of data collected by all these sensors will also hugely boost the model-building efforts that Anduril is now doing in a partnership with OpenAI, announced on December 4. Earlier this year, Palantir also offered its AI tools to help the Pentagon reimagine how it categorizes and manages classified data. When Anduril founder Palmer Luckey told me in an interview in October that “it’s not like there’s some wealth of information on classified topics and understanding of weapons systems” to train AI models on, he may have been foreshadowing what Anduril is now building.
Even if some of this data from the military is already being collected, AI will suddenly make it useful. “What is new is that the Defense department now has the capability to use the data in new ways,” Emelia Probasco, a senior fellow at the Center for Security and Emerging Technology at Georgetown University, wrote in an email. “More data and ability to process it could support great accuracy and precision as well as faster information processing.”
The sum total of these developments might be that AI models are brought more directly into military decision-making, rather than just surfacing information. That idea has brought scrutiny, like when Israel was found last year to have been using advanced AI models to process intelligence data and generate lists of targets. Human Rights Watch wrote that the tools “rely on faulty data and inexact approximations” in a report.
“I think we are already on a path to integrating AI, including generative AI, into the realm of decision making,” says Probasco, who authored a recent analysis of one such case. She examined a system built within the military in 2023 called Maven Smart System, which allows users to “access sensor data from diverse sources [and] apply computer vision algorithms to help soldiers identify and choose military targets.”
Probasco said that building an AI system to control an entire decision pipeline, possibly without human intervention, “isn’t happening and there are explicit U.S. policies that would prevent it.”
A spokesperson for Anduril said that the purpose of Mesh is not to make decisions. “The Mesh itself is not prescribing actions or making recommendations for battlefield decisions,” the spokesperson said. “Instead, the Mesh is surfacing time-sensitive information” that operators will consider as they make those decisions.