Betting on Autonomous Aircraft

Join War on the Rocks and gain access to content trusted by policymakers, military leaders, and strategic thinkers worldwide. This exclusive Cogs of War interview is with Brandon Tseng, the co-founder and president of Shield AI and a former Navy SEAL, whose company builds the Hivemind autonomy software, the V-BAT and X-BAT aircraft, and Aechelon synthetic reality and simulation technologies. We asked Brandon about the bet at the center of his company — that the AI pilot matters more than the airframe — and about what a decade building autonomy for the Pentagon has taught him. In 2015, I believed the fringe position that by 2035 autonomous systems for defense would become ubiquitous. First, there would be swarms of robots on the battlefield. Second, every military asset would be powered and piloted by physical AI. What I got wrong was the timeline for the Pentagon to act on that thesis. Now I actually think autonomous systems for defense will become ubiquitous by 2040–2045, versus my original 2035 estimate. I underestimated how long it would take procurement processes and investor appetite to catch up. In August 2015, we met with 30 investors in Silicon Valley, and all said no. The next year we met with 25 more, and two said yes. Our board member Peter Levine, from Andreessen Horowitz, described defense tech then as contrarian investing in the same category as Uber or Airbnb: ideas that look dumb until they don’t. I disagree with the premise that we’ve said the aircraft is not a durable advantage. We believe the combination of world-class software and world-class aircraft and hardware will lead to a durable competitive advantage — not that different than any cutting-edge robotics company. We recognize that Shield AI can’t build every piece of hardware, so we focus on the missions where we can make a meaningful dent and build hardware for them, while being a software provider on the hardware we don’t pursue. The saying “hardware is hard” comes from two ideas. One, you are stuck with the design choices you make for years, so software is much more malleable and forgiving than hardware. Two, not only are you stuck with the design choices, you’re stuck with the market. So, we’re very deliberate about what we choose to build. We’re looking for a large market opportunity where we can achieve a minimum 10 times improvement in price-performance. V-BAT did Predator’s missions at 10 times better overall cost, and X-BAT will achieve the mission of fighters and bombers at a fraction of the total cost. Our AI Pilot Hivemind has flown on more than 30 platforms across air, sea, and land, such as our own V-BAT, Anduril’s YFQ-44A, Airbus helicopters, and Ukrainian one-way attack drones. We’ve designed Hivemind so the government doesn’t have to lock into one hardware manufacturer, as in the past, which can delay timelines and lead to all sorts of problems. The biggest lesson from being involved in that many platforms is that the underlying architecture has to be right from day one, or it won’t scale. We got that right early, and it reflects in the timelines: It took Shield AI three years to get an AI pilot flying on the F-16. The same work on a Kratos Firejet took 120 days. Today two engineers can build and fly an AI pilot for a one-way attack drone in under two weeks. About 90 percent of what we build carries over from hardware platform to hardware platform, so lessons from quadcopters improve V-BAT, and lessons from V-BAT improve our fighter jet work. Military intelligence. Everyone wants to talk about missiles and munitions, and those matter, but weapons only work if you know where to point them. In the Gulf War, America had air superiority and precision weapons and still struggled to find mobile Iraqi Scud launchers. The same problem arose in Bosnia, Kosovo, with the Houthis in Yemen, and in Iran today. We’ve lost roughly 50 Predator drones since 2023, many of them to cheap, mobile weapon systems. Right now, across a 24-hour period, we probably cover less than 10 percent of the significant terrain in a given warzone. There’s no infrastructure connecting strategic surveillance and tactical drones, enabling drones to loiter over contested terrain for hours and hand a unit real-time intelligence. That missing middle layer is why opening strikes work, but the following months of a campaign get harder. Additionally, there needs to be more emphasis on autonomy. Autonomy improves every modern weapon. More accurate targeting means a higher probability of kill, and more weapons available for other targets. It’s one of the highest-return investments the military can make, and it gets a fraction of the attention that weapons production does. Part of that is a cultural problem. The Pentagon still treats software as something that should be nearly free. I’ve told Congress that autonomy is more important now than stealth or GPS satellites were when they were introduced, but because it isn’t hardware, the budget for it still isn’t taken as seriously. That’s starting to change, in part due to the capital flowing into defense tech. Programs like the Collaborative Combat Aircraft and the Low-Cost Unmanned Combat Attack System are the first real attempt to fund autonomy and mass together, instead of treating them as separate problems. Satellites are excellent at wide-area, persistent coverage of fixed or predictable things. They are not built to loiter over a specific stretch of contested terrain for hours, watching a tunnel, a garage, or a fishing boat, and handing a current picture to a unit. However, orbit is often treated as a replacement for that middle layer of drones rather than a complement. A satellite can tell you something moved in a region, but an aircraft with advanced sensors and a strong autonomy platform like V-BAT can tell a commander exactly what it was, where it went, and whether it’s a threat, continuously, for hours, even without communications or GPS. Ukrainian operators have told reporters that V-BAT’s range is one of its biggest advantages. Satellites are also a target. If an adversary jams or degrades them in a fight, you need intelligence, surveillance, and reconnaissance that doesn’t depend on a signal from space, and will want to ration whatever satellite bandwidth is left. Ukraine forced us to design for zero dependency on connectivity as a hard requirement. Russia is sustaining more than 200 Shahed-class drone launches a day, GPS is jammed constantly, and communications are spoofed, so the environment is challenging. What has surprised me is how so many U.S. companies left Ukraine after the American government’s easy money stopped flowing. This is actually really bad. It also pissed me off that there was no commitment to the mission, no desire to try and make it work, both technically, operationally, and on the market penetration side. The instinct to stop working the hard technical and operational problems leaves the U.S. military in a more dangerous spot because a lot of equipment simply won’t work on the battlefield. The other thing that surprised me is how far Ukraine’s own tactical drone ecosystem has gone, and its range limitations. Their domestically built systems are effective from zero to roughly 30 kilometers, enough to defend positions and strike smaller tactical targets. We are starting to see an increasing demand for deep strike capabilities that hit the targets that change a campaign’s trajectory. That’s what a lot of Ukrainian companies are building towards now, and what we’ve built toward. This is a deliberate bet. Most Collaborative Combat Aircraft are built around a small business-jet-class engine — the kind that produces roughly 3,000 pounds of thrust — and they’re designed to fly as a wingman alongside a crewed fighter doing the mission’s heavy lifting. X-BAT runs on a GE F110, the same class of engine that powers the F-16, generating around 29,000 pounds of thrust. It carries about three times the payload, generates two to three times the electrical power (which means you can use real electronic warfare payloads at meaningful ranges), and can operate on its own without a crewed aircraft nearby. It was not designed against a Collaborative Combat Aircraft requirement, but designed more for a fighter or bomber mission in the Pacific. Conflicts in Ukraine and the Middle East made clear that runway-dependent airpower is a liability against a peer adversary. Adversaries are targeting runways, carrier decks, and fuel infrastructure specifically to neutralize airpower before launch. With the same internal weapons payload as an F-35, greater range, and true runway-independent operations, it was designed to avoid the problem of runway and carrier targeting. In the opening days of the conflict with Iran, the United States lost billions of dollars in aircraft that were parked on runways in the region. Taiwan has a similar problem: a small number of runways, all easy to locate and destroy on day one, which will neuter its air power. X-BAT can relaunch from a ship, an island, or a bare patch of ground, and it can move every 15 to 30 minutes, so there’s no fixed target to hit. Valuations shouldn’t run too far ahead of your revenue growth or capture of programs of record. Across the industry, that gap is real, and it’s part of why defense tech valuations are getting more scrutiny. For Shield AI, our growth has reflected delivery. You have to deliver meaningful, valuable outcomes for the warfighter, and then do it again, and again, at ever-increasing scale, or you will fail. 99 percent of businesses fail, the Pentagon shouldn’t think that all have to live. You earn the opportunity to win more customer money and higher investor valuations by delivering on your commitments. It’s not about proving anything for Shield AI, or that’s not how I think about it. We just have to continue to deliver great value to our warfighters at increasing scale, and everything else will follow. That looks like a customer who already has our product on the battlefield coming back and asking for more of it, not because a contract renewal is due, but because it’s doing something for them they don’t want to lose. That’s how I know we’re earning it. I think too many things are being purchased or recommended to Taiwan that will simply fail when China turns off GPS and communications and become expensive, useless hardware. That includes precision-guided weapons that rely on a GPS signal to hit their target, and any system that needs a human operator maintaining a live video or control link. Taiwan and the United States should test every weapon system they buy in a GPS- and communications-jammed environment, or require that every system has such a battlefield track record in Ukraine. The conventional wisdom I’d push back on is the idea that if we mobilize, the United States can match China’s production alone. It can’t. We can reduce the gap significantly, and we should and are taking actions to do so, but don’t for a second think that China can’t mobilize as well. China has two to three times the industrial capacity of the United States, installed nearly ten times as many industrial robots in 2024, and spends roughly ten times more on industrial subsidies relative to the size of its economy. But that’s the wrong comparison. Add up the United States and its allies and partners, and combined industrial capacity is probably greater than China’s. We need to accelerate and continue to double down on our military and industrial partnerships with Taiwan, Japan, Australia, South Korea, the Philippines, and India. A bully will pick on an individual when they think they are alone or by themselves, but when the individual brings the neighborhood gang, the bully checks himself. Brandon Tseng is the co-founder and president of Shield AI. He previously served seven years in the U.S. Navy as a Navy SEAL officer and surface warfare officer, deploying to Afghanistan and the Pacific Theater. Image: Staff Sgt. Dylan Chagnon via DVIDS