Cheap Volume Beats Gold-Plated Jets: Anduril’s Brian Schimpf on AI and Defense

Elad Gil introduces Brian Schimpf as co-founder and CEO of Anduril, a next-generation defense company that put AI into low-cost drones, sensor networks, and machine vision early. Schimpf has been coding since he was twelve and still says spare time would go back into code. In college he worked on self-driving cars at the start of that wave, on a small, mostly undergraduate Cornell team in the DARPA Grand Challenge and Urban Challenge. Nothing worked in 2004; by 2007 multiple teams were driving through cities. The pace of robotics stuck with him.

He moved to Silicon Valley and joined Palantir as a forward-deployed engineer, sitting with national-security customers. What surprised him was how much people in that world cared about the mission, plus how technically hard the problems were. He later ran product and engineering there, building analytics that had to work at government scale. Around 2017 he founded Anduril with college friend Matt Grimm, Palantir colleague Trae Stephens, and Palmer Luckey—the Oculus founder who, Luckey insists, was fired from Facebook—because the group understood a sales motion unlike enterprise or consumer: getting technology scaled, deployed, and sold to defense.

Gil had been looking at the sector when Google shut down Maven, its DoD work, which he found odd given the historical tech-defense loop. Anduril is the only pure-play defense company he invested in; Scale and Applied Intuition touch the industry, but he wanted a shot at something large. Schimpf’s go-to-market answer starts with a split. Traditional primes are genuinely good at aircraft carriers and exquisite fighter jets if time and budget are unlimited. Fast-forward twenty or thirty years, and the critical kit looks different.

The thesis at founding was higher quantities of lower-cost systems, and the thing that makes that work is intelligent software. The best software people were in Silicon Valley, and they were loudly choosing not to work on defense. Schimpf also argued you had to control hardware *and* software, because the Pentagon does not buy a component. It buys a capability: find the ships, which means drones or satellites plus networking plus the whole stack. Lattice, built from day one, is Anduril’s answer to a defense-specific problem—ingest a huge volume of sensor data and track tanks, ships, and aircraft that are hard to locate.

The same software moved across hardware. First came force protection, base defense, and border security: people, cars, and animals over wide areas. Then small drones, including briefcase-sized helicopters with serious autonomy. Then larger systems. The biggest program on the board is an autonomous underwater vehicle about the size of a school bus—roughly 60 feet long and 10 feet in diameter, able to run huge distances for weeks. Underwater, physics kills communications, so the autonomy has to actually work or the boat does not come home.

The point of the software is to automate the mechanical, repetitive pieces of defense work so low-cost systems can scale. Until you solve that, you are capped by manpower: how many humans you can deploy. Everyone wants the same capability at lower cost; the curve does not bend until autonomy takes the operator bottleneck out. US operations still rest on outrageously expensive manned platforms—carriers, fighters, bombers—plus a pilot shortage and recruiting shortfalls. Industrial capacity is worse. Nuclear-sub production is falling, about 1.7 boats a year, with a retiring workforce, an artisanal process, and few suppliers. Matching China on those large platforms, Schimpf says, is a bad strategy that will not work.

Look at Ukraine, Schimpf says, and a lot of how people thought warfare would evolve is already on the field. What is almost surprising is how little you see even the Russians deploying fighter jets or other exquisite platforms. Where is the air campaign that has been the hallmark of how the United States fights? Surface-to-air missiles and air defenses now impose an attrition rate too high for very expensive systems. Intelligence collection and operations have to move down to lower echelons and more autonomous units.

Battlefield awareness has gotten very good, commercially as well as militarily. Commercial satellites make large bases, airfields, and staging areas hard to hide. Troops have been pushed into disaggregated small units that are uneconomical to hit with a battle-changing strike. US long-range fires have pushed Russian forces back from the front because they cannot form up. Fixed infrastructure went on day one: Ukrainian airfields and other installations. Speed, mobility, and hiding in the noise beat gold-plated platforms that get lost quickly and cannot be sustained.

He thinks the United States is taking those lessons into the Pacific. Operated this new way, the defender is heavily favored. A successful invasion of Taiwan becomes immensely hard—not because you out-carrier the other side, but because massing and approaching get expensive in lives and hulls. The same physics-limited sensor kit is surprisingly general: thermal or visible pictures, radio emissions, radar in a few modes. That is basically the list. Detection and classification across forest and desert become a general-purpose software problem, which is why Lattice is a toolkit rather than a one-off. Anduril does not do AI research; it takes what is working at the state of the art and applies it fast. Hardware is then tailored to a specific range and target set. Defense, he adds, is unusually analyzable: you can simulate a capability against an adversary and see whether a given range, target type, and comms resilience would actually matter.

The urgent-problem filter is how Anduril picks where to push. Counter-drone is the example that has worked. Historically only a few countries could field cruise missiles. Cheap drones gave Iran and nearly every other country a low-cost equivalent: a hundred thousand dollars or less per airframe versus a Tomahawk at over a million. Less capable, still a problem. Air defense flipped from rare high-end state threats to something you will see in every conflict. Adversaries iterate in months, not ten-year development cycles, which demands a software-first response: deploy radars and cameras, then rapidly iterate detection and classification, and pull in jamming, guns, and missiles as a combined package. Customers want the problem solved, not a spec sheet.

The formula Schimpf keeps repeating is urgency plus speed as the determinative factor. Invest where nobody is building the right tech, where cheap-smart-at-scale software is the only way, and where the adversary is changing faster than a traditional program can. SpaceX and others showed that problems which looked like twenty-year, multi-billion-dollar investments are now doable in years. That is a lot of headroom.

How autonomous is a bus-sized submarine on a weeks-long patrol? Schimpf starts with principle, not demos. The United States is not going to adopt robots that make lethal decisions about what to attack. Accountability is assigned to the person who employed the weapon. He likes that and wants it to stay. Autonomy, then, has to be predictable enough that operators trust it and know the effect when they hit a button. Anduril has argued for crawl, walk, run, because adoption is the real bottleneck.

Non-deterministic tools—reinforcement learning, potential LLM control loops—are a risk until you can bound them. Even if the path is a little unpredictable, the outcome has to sit inside a known envelope. The design goal is more human agency, not less: the right information, not too much and not too little, synthesized so people can decide. An underwater vehicle out for weeks is a harder case. You hand it mission intent—“search this harbor for mines, image what you find, cover the likeliest ground”—which is a human concept. Encoding that into machine behavior, with contingencies, and running reliably for weeks with no hiccups is the gap between a research project and a fielded capability. It is the same nines-of-reliability problem as self-driving cars, and that is where the time and money go.

Gil asks where large language models fit besides Lattice’s sensors and vision—cyber, for example. Schimpf sees two dimensions. The first is synthesis: stuff a huge volume of text into a network and interrogate it. The United States sits on an ocean of intelligence reports and collections. Being able to ask questions of that knowledge base would be hugely valuable. Hallucination is disqualifying. Invented intelligence about something that happened in China would be a very bad situation. Open-source reporting and forum chatter are easier first deployments because you can cross-check.

The second dimension is generation. DoD culture is operational and predictable: train people so a given input yields a known output. Cyber is a clearer creative fit, on offense and defense—automating attack patterns, classifying inbound ones. Autonomy is another: translate human mission intent into API calls that drive real systems. Schimpf has been watching research on LLMs trained to invoke external tools. Anduril’s philosophy is still a human at the right spot. Let the model do the legwork of a mission plan and hand over a 95% baseline to edit, amplifying warfighters instead of replacing their judgment. Reliability and precision still have to meet the standard the public will hold the Department to. Foundation models and infrastructure are in place; he expects a tidal wave of applications in 18 to 24 months. Anduril is exploring, not yet at large-scale deployment.

Six years in, Anduril is about 1,600 people, based in Orange County with sites across the US and a fast-growing Australia presence—partners who feel the relevant threats in their backyard. The company works with every service: Special Operations, Army, Marine Corps, Air Force, Navy. The most mature footprints are the first product lines, border security and base protection, plus expanding counter-drone. Underwater vehicles and newer autonomous drones are just reaching fielded scale. The buyer set is tiny, so revenue is episodic: you might sell ten of a thing one year and five thousand the next. Being ready to hit the manufacturing button without overbuilding early is the hard thread. Capital from firms such as Founders Fund, Andreessen Horowitz, and General Catalyst is what lets them jump the valley of death between interesting tech and a quantity that would matter in a real conflict.

Gil notes Anduril is often described as the fastest company since the Korean War to land a program of record—a DoD budget line, usually hundreds of millions or billions, unlike a pile of million-dollar pilots that never grow. Schimpf measures success by volume in the field. The first large award was Customs and Border Protection, about two and a half years in, a $250 million border-security contract. A year or two later, a billion-dollar counter-drone contract with SOCOM. Those concentrated captures are high-stress; they are also the job.

Convincing the government you are a trustworthy partner started on day one with a lobbying group on the Hill: national security, taxpayer efficiency, a better model for buying tech. Prior government-service hires know how to thread a Byzantine budget. The 2026 budget is being written today (the conversation is June 2023). Defense does not run on Silicon Valley clocks. Tech people think a compelling product is enough. DoD buys a capability, which means support, training, logistics, manufacturing, integrations, and security infrastructure—on the order of fifty skill sets. Those people are hard to find, compliance takes years, and you have to make it easy for a customer to say yes on an unknown company. Anduril’s edge was knowing that in advance and getting aggressive early.

Gil asks what the next structural opening is after drones plus machine vision. Schimpf says that shift is still playing out, then names two more. First, cheap launch. He does not think the US government has fully absorbed how cheap large low-Earth-orbit constellations will be, especially with Starship. Command-and-control overlap with what Anduril already builds, but the dollars per kilogram change the math so much that a mega-constellation may cost what four satellites used to. He puts that transition at maybe five percent played out.

Second, hypersonics. Multiple startups are doing credible air vehicles, rockets, and missiles wildly cheaper than the incumbent state of the art, at something like five times the pace. That used to be a Lockheed-shaped, high-end, very expensive problem, which meant the US could afford one or two missile bets. The intersection is the wild part: proliferated satellites and drones, software that makes sense of the whole battlespace, and the ability to engage at outrageous range on a short timeline. Instead of parking a carrier group and thousands of lives close in, you fight from far away. Adversaries can learn the same tricks, so the policy goal is deterrent clarity: an invasion of Taiwan will not work. Allied capacity to withstand an ingress on their sovereignty matters as much as a US threat of force.

What changed the industry’s willingness to work with DoD? Schimpf’s short answer is Vladimir Putin. Ukraine killed the idea that state-on-state war was over. If a dictator thinks force will get a political end, force is on the table. On China he is equally blunt: listen to what Xi says, especially domestically—use of force on Taiwan is on the table. Economic ties did not prevent World War I or II. The Department, meanwhile, is more willing than it was ten or fifteen years ago because China’s modernization, including longer-range missiles, made US lag visible. Give DoD credit, he says: a century-old institution of millions of people has reinvented itself more than once. Not Silicon Valley pace, but it moves. The under-investment that still worries him is the sheer size of the Pacific—overlay the US on the South China Sea and America looks small—plus allied kits that are not copies of long-range US strike. You do not necessarily want to give Taiwan the tools to hit deep into the mainland; you do want coastal defense, cables, and shipping lanes that make an invasion insane, so Washington is not stuck guessing whether putting troops in harm’s way is politically possible.

Where else can tech help? Schimpf points at Amazon, Microsoft, and Google bringing cloud into classified spaces, and at companies like Flexport on logistics. Silicon Valley has modernized conventional industries in ways that map onto DoD if you are willing to adapt the product. The larger failing is political. Valley culture still wants to be left alone, unregulated, consequences later. He sees blowback on crypto, and potentially on AI and social media. Washington does have a vote, and rightfully so. Engage, tell the story, help legislators write thoughtful rules. They want the engagement.

Guo, an Anduril seed investor, says the round was the first time she got hate texts. After Ukraine, some of the same people texted about what an amazing company it was. Schimpf has, for better or worse, only worked in controversial places—border security, defense. You need clear conviction about why you are doing it. Anduril is not shy: they work on weapons, they believe that is important, ethical, and necessary. The most controversial articles call the work scary. His response is that it *is* scary, which is why you want the best people on it. Popularity is not leadership. Conviction, done responsibly, is.