This exclusive Cogs of War interview features Brandon Tseng, co‑founder and president of Shield AI and former Navy SEAL, whose firm develops the Hivemind autonomy software, the V‑BAT and X‑BAT aircraft, and Aechelon synthetic‑reality simulation tools. We discussed the core thesis of his company—that the AI pilot outweighs the airframe—and the lessons learned from a decade of building autonomy for the Department of Defense.
When you launched Shield AI in 2015—long before defense technology became trendy and few in Washington thought a startup could sell autonomy to the Pentagon—what did you anticipate that the market has only recently embraced, and where did you misjudge?
In 2015 I held the fringe view that by 2035 autonomous defense systems would be everywhere: battlefield robot swarms and every military asset powered and piloted by physical AI. I missed the timing of Pentagon adoption. Today I see widespread autonomy arriving nearer 2040‑2045, not 2035, because procurement cycles and investor interest moved slower than I expected. In August 2015 we pitched thirty Silicon Valley investors and heard only refusals; a year later twenty‑five more meetings yielded two yes votes. Board member Peter Levine of Andreessen Horowitz likened defense tech then to a contrarian bet akin to early Uber or Airbnb—ideas that look foolish until they prove indispensable.
Shield AI’s strategy hinges on the belief that the AI pilot, rather than the airframe, provides the lasting advantage—a contrarian stance in an industry dominated by costly hardware. Why do you confidence that the software drives the value?
I disagree with the notion that we dismiss the aircraft as a durable asset. We see a winning formula as world‑class software paired with world‑class aircraft and hardware, much like any leading robotics firm. Because we cannot build every piece of hardware ourselves, we concentrate on missions where we can make a meaningful impact and develop purpose‑built platforms, while supplying our AI pilot for hardware we do not manufacture. The adage “hardware is hard” captures two realities: design choices lock you in for years, making software far more adaptable, and you remain tied to the market you serve. Hence we are deliberate about what we build. We target large markets where we can achieve at least a ten‑fold improvement in price‑performance. The V‑BAT delivers Predator‑level capability at a tenth of the cost, and the X‑BAT aims to match fighter‑ and bomber‑mission performance for a fraction of the total expense. Our Hivemind AI pilot has flown on more than thirty platforms—air, sea, and land—including our own V‑BAT, Anduril’s YFQ‑44A, Airbus helicopters, and Ukrainian one‑way attack drones. By designing Hivemind to be hardware‑agnostic, we avoid vendor lock‑in that can stall programs and create complications. Experience across these platforms taught us that a sound underlying architecture is essential for scalability; getting it right early has shortened integration cycles: three years to fly an AI pilot on an F‑16 versus 120 days on a Kratos Firejet, and today two engineers can field an AI pilot for a one‑way attack drone in under two weeks. Roughly ninety percent of our work transfers between platforms, so insights from quadcopters enhance the V‑BAT, and V‑BAT lessons accelerate fighter‑jet development.
Military intelligence often receives less fanfare than weapons, yet it is essential for effective targeting. What gaps do you see in current ISR capabilities?
Everyone gravitates toward missiles and munitions, but weapons are useless without knowing where to aim. In the Gulf War, despite air superiority and precisionguided munitions, coalition forces struggled to locate mobile Iraqi SCUD launchers—a problem that resurfaced in Bosnia, Kosovo, against Houthi forces in Yemen, and in Iran today. We have lost roughly fifty Predator drones since 2023, many to inexpensive, mobile threat systems. At present, across a full day we likely surveil less than ten percent of the significant terrain in a typical combat zone. There is no connective layer linking strategic surveillance with tactical drones that could loiter over contested areas for hours and hand real‑time intelligence to ground units. That missing middle explains why opening strikes succeed while sustaining campaigns becomes harder. Greater emphasis on autonomy would improve every modern weapon: sharper targeting raises kill probabilities and frees munitions for other targets—one of the highest‑return investments the military can make, yet it receives only a fraction of the attention lavished on weapons production. This stems partly from a cultural habit of treating software as near‑free. I have told Congress that autonomy now rivals stealth or GPS satellites in strategic importance, but because it is intangible, its budget is not taken as seriously. That perception is shifting, helped by capital flowing into defense technology. Initiatives like the Collaborative Combat Aircraft and the Low‑Cost Unmanned Combat Attack System represent the first serious attempts to fund autonomy and mass production together, rather than treating them as separate challenges.
Satellites excel at wide‑area, persistent monitoring, but they cannot replace the agility of airborne ISR. How do you decide what belongs in orbit versus what belongs in the air?
Satellites excel at wide‑area, persistent coverage of fixed or predictable targets. They are not designed to linger for hours over a specific piece of contested terrain—such as a tunnel, garage, or fishing boat—delivering continuously updated pictures to a unit. Often orbit is mistakenly viewed as a substitute for that middle layer of drones when it should complement it. A satellite can tell you that something moved in a region, but an aircraft equipped with advanced sensors and a robust autonomy platform like the V‑BAT can identify exactly what it was, where it went, and whether it poses a threat, providing that information continuously for hours even without GPS or communications. Ukrainian operators have highlighted V‑BAT’s range as one of its greatest strengths. Satellites are also vulnerable; if an adversary jams or degrades them during conflict, you need ISR that does not rely on space‑based signals and can make do with any remaining satellite bandwidth.
Operating in Ukraine has forced Shield AI to harden its systems against denied environments. What have you learned from that theater?
Ukraine compelled us to make zero reliance on connectivity a non‑negotiable requirement. Russian forces launch more than 200 Shahed‑class drones daily, GPS is constantly jammed, and communications are spoofed, creating a brutally challenging environment. What has surprised me is how many U.S. firms withdrew from Ukraine once the initial influx of government funding dried up—an outcome that is not only regrettable but also dangerous, as it abandons the technical, operational, and market efforts needed to make equipment work on the battlefield. Another surprise has been the rapid growth of Ukraine’s own tactical drone ecosystem. Their domestically built systems are effective from zero to roughly thirty kilometers, sufficient for defending positions and striking smaller tactical targets. We are now seeing rising demand for deep‑strike capabilities that can hit targets capable of shifting the course of a campaign—a direction many Ukrainian companies are pursuing, mirroring our own development path.
The X‑BAT represents Shield AI’s biggest and least proven bet. Why is this the right moment to pursue it, and what would need to change in Washington for it to succeed?
This is a calculated wager. Most Collaborative Combat Aircraft designs center on a small‑business‑jet engine delivering about 3,000 pounds of thrust, intended to fly as a wingman alongside a crewed fighter handling the mission’s primary load. The X‑BAT instead uses a GE F110—the same class powering the F‑16—producing roughly 29,000 pounds of thrust. It carries about three times the payload, generates two to three times the electrical power (enabling meaningful electronic‑warfare payloads at range), and can operate independently of a crewed aircraft. It was not shaped by a Collaborative Combat Aircraft requirement but rather conceived for fighter‑ or bomber‑type missions in the Pacific. Recent conflicts in Ukraine and the Middle East have shown that runway‑dependent airpower becomes a liability against a peer adversary; enemies target runways, carrier decks, and fuel supplies to neutralize airpower before launch. With an internal weapons load comparable to the F‑35, greater range, and true runway‑independent operation, the X‑BAT is designed to evade those vulnerabilities. It can relaunch from a ship, an island, or a bare patch of ground and relocate every fifteen to thirty minutes, eliminating a fixed target for adversaries. In the opening days of the Iran conflict, the United States lost billions of dollars in aircraft parked on regional runways—Taiwan faces a similar predicament with a limited number of easily located runways that could be destroyed on day one, crippling its air power. The X‑BAT’s ability to disperse and resupply from austere locations directly addresses that weakness.
Defense‑tech valuations have surged on expectations of rapid Pentagon purchases. What must Shield AI demonstrate over the next few years to justify the market’s enthusiasm?
Valuations should not outpace revenue growth or the capture of programs of record—a gap that exists industry‑wide and is prompting increased scrutiny of defense‑tech multiples. For Shield AI, our growth has mirrored our delivery record. You must provide meaningful, valuable outcomes for the warfighter, repeat them at ever‑increasing scale, or you will fail; ninety‑nine percent of businesses do not survive. The Pentagon should not assume every entrant will prevail. You earn the right to secure more customer funds and higher investor valuations by consistently meeting commitments. It is not about proving something abstract for Shield AI; it is about continuing to deliver great value to our warfighters at expanding scale. When a battlefield customer returns voluntarily—asking for more capability not because a contract renewal looms but because the system is indispensable—that is the signal we are earning our place.
Discussions of a Pacific confrontation often repeat the mantra “distributed, attritable, autonomous.” Where does the prevailing wisdom about a China contingency fall short or become overly complacent?
Too many systems being procured or recommended for Taiwan would simply fail if China disrupted GPS and communications, turning them into costly, useless hardware. This includes GPS‑guided precision weapons and any platform that relies on a live human‑operator video or control link. Taiwan and the United States should subject every weapon system they acquire to GPS‑ and communications‑jammed testing, or require a proven battlefield track record from environments like Ukraine. The conventional wisdom I challenge is the belief that, should mobilization occur, the United States alone could match China’s production output. That is not realistic; we can narrow the gap, and we are taking steps to do so, but we must not ignore that China can mobilize effectively as well. China possesses two to three times the United States’ industrial capacity, installed nearly ten times as many industrial robots in 2024, and spends roughly ten times more on industrial subsidies relative to its economy. The proper comparison aggregates the United States with its allies and partners; combined industrial capacity likely exceeds China’s. To deter aggression, we must deepen and accelerate military‑industry partnerships with Taiwan, Japan, Australia, South Korea, the Philippines, and India. A bully targets an isolated individual, but when that individual is backed by the neighborhood gang, the bully thinks twice.
Image: Staff Sgt. Dylan Chagnon via DVIDS
