The Billion-Dollar Bet That Video Games Are the Future of Weapons
Somewhere in Seoul right now, a teenager is dropping into a PUBG match on her phone. She’s parachuting onto a virtual island, looting a shotgun from a bombed-out house, sprinting toward gunfire. She is not alone. PUBG Mobile alone has been downloaded over a billion times.
Now Krafton, the company behind PUBG, wants to expand from entertainment into defense contracting.
On March 13, Hanwha Aerospace and Krafton announced plans to jointly develop “Physical AI.”
The goal?
Use PUBG’s engine along with a $1 billion joint fund to train the AI inside real weapons.
The idea isn’t new. In 1996, a Marine lieutenant modified DOOM II to train fire teams for close-quarters combat. Demons became enemy combatants. The BFG became an M16. The graphics were terrible. The AI enemies ran at you like they wanted to die. It worked anyway.
That was a zero-budget hack on a $50 game engine. South Korea is betting a billion dollars that the same principle scales to autonomous weapons.
Why Game Engines
Training AI for the physical world requires a staggering volume of scenarios. A self-driving car needs millions of simulated miles before it touches asphalt. A military drone needs millions of simulated engagements before it enters airspace.
The problem is that real-world military testing is slow and expensive. You can’t crash 10,000 drones into mountains to teach the 10,001st how to avoid them. You can’t fire live munitions at scale to train a targeting algorithm. If you’ve seen Tron: Ares, the opening sequence is closer to what’s actually happening than most people realize.
A program trained and tested inside the Grid before being released into the real world. The US military has known this since 1983, when DARPA launched SIMNET, a program born from an idea Jack Thorpe first outlined in a 1978 paper. It was the first networked military simulation. Soldiers in M1 Abrams simulators fought computer-generated Soviet tanks over Ethernet cables. When a crew got “killed,” their screen went blank. Then they replayed it, Groundhog Day-style, until the tactics clicked.
SIMNET was primitive. But it proved the concept: you can train human judgment through virtual combat.
Game engines took that concept and scaled it by orders of magnitude. A modern engine like Unreal Engine 5 simulates real-world physics: gravity, ballistics, wind, terrain deformation. It generates a different city layout every run. It models enemy AI that learns and adapts instead of following scripts. PUBG’s engine already handles all of this for 100 simultaneous players across a 64 square kilometer map. The computational infrastructure Krafton built to entertain gamers turns out to be almost exactly what you need to train autonomous weapons.
The “almost” is important. Entertainment physics are close enough to fool your eyes. Military physics need to be close enough to land a munition. That’s the gap the billion dollars is meant to close.
Korea’s Defense Machine
South Korea’s defense exports hit a record $17.3 billion in 2022, up from around $3 billion in 2015.
What changed?
Europe rearmed after Russia invaded Ukraine and discovered its own defense contractors couldn’t deliver fast enough. Germany’s tank production lines were cold. France’s order books were full for years. Seoul could ship now, at lower prices, and was willing to share the technical knowledge to build locally, something companies like Lockheed and Rheinmetall historically refused to do.
When Poland needed to replace its Soviet-era tanks, it called Seoul. Within months, the first K2 Black Panther tanks were rolling off ships. The deal eventually expanded to over 1,000 tanks and hundreds of K9 howitzers, one of the largest arms deals in modern European history.
Hanwha Aerospace, the K9’s manufacturer, sits at the center of this boom. The company absorbed Samsung’s defense subsidiary, added precision-guided munitions and space launch vehicles, and now wants to be the Korean Lockheed Martin.
But hardware alone is a commodity game. Every defense company in the world can build a drone chassis. The margin lives in the software that tells the drone what to do when it loses contact with its operator and the target moves.
That’s where Krafton comes in.
What “Physical AI” Actually Means
The term “Physical AI” is everywhere in 2026, and it means different things depending on who’s selling it. NVIDIA uses it to describe robot simulation. Tesla uses it to describe humanoid robots. Hanwha and Krafton mean something more specific: AI trained in simulation environments with accurate-enough physics to operate reliably in the real world.
Training Physical AI requires simulation environments that model real physics with enough fidelity that lessons transfer to reality. This is called the “sim-to-real gap,” and it’s the central unsolved problem of AI that has to operate in the physical world. A drone that flies perfectly in simulation and crashes into a tree in real life is useless.
PUBG’s engine isn’t accurate enough for this out of the box. But the infrastructure underneath it is. The networking, the procedural world generation, the ability to run thousands of simultaneous AI agents in a shared environment. That’s what Hanwha is buying. Krafton’s engineers know how to build virtual worlds at scale. Hanwha’s engineers know what physical accuracy those worlds need to achieve.
The Convergence A General Predicted
Last August, retired four-star general Joseph Votel wrote an essay arguing that the US military should be paying attention to competitive gaming. Not as a recruiting tool, though that was the Pentagon’s usual pitch, but as a training methodology. Games like StarCraft II and Rainbow Six Siege, he argued, develop the same cognitive skills as military command: rapid decision-making under uncertainty and real-time adaptation to adversarial tactics.
Seven months later, South Korea is proposing that the games themselves become the training ground for the weapons. The cognitive loop that Votel described for human players is now being applied to AI agents. Instead of a human learning to make faster decisions in StarCraft, an AI is learning to navigate physical terrain in a simulation built on PUBG’s architecture.
This is a fundamentally different bet than what the US defense-tech ecosystem is making. Anduril, the most prominent American defense startup, builds autonomous systems powered by its Lattice software platform. It acquired a game studio, Carbon Games, back in 2019 for its engine tech. But the American approach still treats defense AI as a software integration problem. The Korean approach treats it as a simulation problem, and the scale of the bet reflects that.
Anduril has simulated testing environments, but its primary development loop runs through real-world sensor data. If Hanwha and Krafton can build simulation environments accurate enough to train weapons AI, they can run thousands of tests per day at a fraction of the cost. The speed advantage compounds.
This isn’t new in principle. The internet was a DARPA project. GPS was built for missile guidance. The semiconductor industry exists because the military needed smaller circuits. Civilian and military technology have always been entangled.
But the relationship used to flow in one direction: military R&D produced technology that eventually became consumer products. The Hanwha-Krafton deal reverses the flow. Consumer entertainment technology, built to keep gamers engaged, is being repurposed for weapons training. The skills Krafton developed to make PUBG’s ballistics feel satisfying are being applied to make real ballistics accurate.
Follow the Money
The old defense model is simple: build hardware, bill the government, collect your guaranteed profit margin. That model assumed the hardware was the hard part.
Defense investors have spent the last five years chasing hardware companies. Drone makers and missile builders. But if simulation becomes the bottleneck, the companies with the best virtual worlds win the contracts, not the ones with the best assembly lines.
The US spent $886 billion on defense last year. Its biggest defense-tech startup bought a small game studio six years ago and folded it into an engineering team. South Korea spent a fraction of that budget and launched a billion-dollar joint venture between its top weapons maker and the studio behind the most-downloaded mobile shooter in history.
The next generation of weapons AI won’t be trained on battlefields. It will be trained in virtual worlds built by game developers. Whoever builds the best simulation environment controls the pace of weapons development for the next decade. Right now, the country best positioned to do that is the one that built PUBG.
The Pentagon might want to start returning General Votel’s calls.
Sources: