The US Air Force is relying on an AI-powered digital sandbox to launch war games 10,000 times faster and prepare for tomorrow’s conflicts.
In summary
The US Air Force wants to transform its digital war games by changing scale. In a request for information published at the end of November, the HAF/A5 command describes a “WarMatrix Ecosystem” project: a cloud environment, powered by military artificial intelligence, capable of generating and running wargaming scenarios up to 10,000 times faster than real time. In practice, a week of campaigning could be simulated in one minute. The objective is twofold: to accelerate the planning of multi-domain warfare (air, sea, land, cyber, space) and to test combinations of forces, doctrines, and technologies at a lower cost. This “digital sandbox” must ingest data from battle orders, sensors, logistics, and even human behavior to generate thousands of variations of the same potential conflict. The US Air Force is not alone in this field: China is investing heavily in AI applied to wargaming, Russia is exploring military AI for simulation, and European allies, NATO and the UK are also working on AI-enhanced military simulation platforms.
The US Air Force’s gamble: wargames 10,000 times faster
The HAF/A5 request for information on “Advanced Wargaming and Simulation Technologies for Integrated Force Design” clearly describes the ambition. The Air Force is looking for manufacturers capable of providing an integrated AI wargaming architecture, called WarMatrix, designed to support the design of the future force for 2040–2050.
At the heart of this vision is a cloud-based digital sandbox. It should enable:
- the automatic generation of complete battle orders;
- model multiple simultaneous theaters of operations over distances of several thousand kilometers;
- integrate real or realistic data (geography, range capabilities, fuel consumption, ammunition stocks, satellite orbits) in metric and nautical units;
- run thousands of simulations with variations in posture, weather, losses, and human decisions.
The claim of scenarios being “up to 10,000 times faster than real time” is not just a slogan. At this speed, a 30-day (720-hour) conflict can be compressed into less than 5 minutes of computation (720 / 10,000 × 60 ≈ 4.3 minutes). This acceleration makes it possible to move away from the logic of a few large, long, costly, and unrepeatable annual wargames to a statistical approach: hundreds of simulated campaigns to measure trends, breaking points, and surprise effects.
The US Air Force assumes that the complexity of multi-domain warfare now exceeds human cognitive capabilities when it comes to systematically exploring all options. AI should serve as an accelerator for exploration, not a replacement for strategists, at least officially.
How an AI-powered digital sandbox works
A cloud architecture for multi-domain models
Technically, the WarMatrix project is based on three building blocks: a secure cloud infrastructure, multi-domain simulation models, and an AI layer to orchestrate and analyze scenarios. The RFI emphasizes the need for open, interoperable platforms capable of communicating with existing modeling and C2 (command and control) systems.
Specifically, we are talking about models:
- aerial: fighter jet performance, air-to-air missile range, flight profiles in meters and kilometers;
- naval: mobility of carrier strike groups, radar range, vulnerability in coastal environments;
- land: brigades, logistics, fuel consumption (liters) and ammunition;
- cyber and space: satellite availability, network attacks, jamming, sensor loss.
These models can be classic motion equations, but increasingly they are agent-based models: each unit, platform, or “player” is represented by an agent that makes local decisions, making it possible to simulate emergencies, human errors, and logistical bottlenecks.
The AI layer intervenes at several levels: assistance in scenario generation, autonomous “adversaries,” optimization of simulation parameters, and, above all, analysis of results. Military artificial intelligence makes it possible to identify, in tens of thousands of iterations, configurations that recur systematically: areas of logistical saturation, vulnerabilities in a network of bases, recurring enemy attack options.
AI that learns by playing thousands of times
Work carried out with MIT and Air Force Futures on AI applied to wargaming is already moving in this direction. Since 2024, the US Air Force has been testing reinforcement learning algorithms capable of playing entire campaigns against humans to explore new concepts for the use of drones, sensors, and long-range weapons.
In the WarMatrix project, these techniques are set to be taken even further:
- “Blue” AIs that optimize the US posture under constraints of budget, inventory, number of aircraft, and days of deployment;
- “Red” AIs that learn to exploit flaws in US plans, drawing on known Russian or Chinese doctrines, enriched by emerging behaviors;
- “Neutral” AIs that generate unforeseen events: technical failures, extreme weather, political crises, and disruptions to logistics flows.
The advantage of these AIs is the ability to change the rules of the game. Whereas a traditional wargame remains static throughout a session, an AI wargame can incorporate new weapons, new sanctions, or a change in the alignment of a third country from one run to the next. The machine never tires of playing the same scenario 1,000 times to test small parameters.
Military uses of digital war games
A laboratory for planning and Integrated Force Design
The primary beneficiary of this AI wargaming will be Integrated Force Design, i.e., how the US Air Force decides on the composition of its future fleet: the proportion of manned fighters, collaborative drones, bombers, refueling aircraft, and satellites. By launching thousands of simulated campaigns, the WarMatrix Ecosystem should make it possible to objectively compare different force “mixes.”
For example: how many collaborative drones are needed to compensate for the reduction of 50 manned fighter jets in the Indo-Pacific theater, taking into account actual distances in kilometers, transit times, and in-flight refueling? How many forward bases are needed to sustain a given sortie rate without maintenance breaking down after 30 days? These types of questions, which are currently addressed by ad hoc studies, can now be explored systematically.
The other use is doctrinal: testing concepts such as Agile Combat Employment (ACE) or dispersion operations on small runways, simulating the flow of fuel, spare parts, and ammunition to islands hundreds of kilometers away.
Training planners and measuring the effect of AI on humans
Digital war games are not only used to produce curves; they also train staff. Work carried out at the Air University shows that AI can be used to build adaptive training environments, where the enemy changes tactics on the fly to force cadets to think outside the box.
WarMatrix can be used in “sandbox mode,” in which human planners play the Blue side, assisted by AI tools that suggest options or flag logistical inconsistencies. The experience can also be reversed: let the AI play Blue, and ask humans to put themselves in Red’s shoes and try to beat it. This helps identify:
- the “blind spots” of AI, such as maneuvers that are politically impossible but effective on the map;
- human biases, such as the tendency to repeat familiar patterns despite poor results in simulation.
These experiments are valuable for calibrating the future division of labor between humans and machines in command centers.
The advantages sought by the US Air Force
The most obvious argument is the time saved. A large strategic wargame can mobilize hundreds of people, take months to prepare, and last several days. With a digital sandbox, the actual game phase can be repeated at high speed, and preparation can be automated using scenario libraries.
The second advantage is analytical. By moving from a few case studies to thousands of runs, the US Air Force can:
- estimate distributions of results (losses, campaign duration, ammunition costs);
- identify “robust areas” where an option remains favorable despite uncertainties;
- identify configurations where a small change in posture produces a disproportionate effect, useful for deterrence or surprise.
The third advantage is budgetary and political. Major decisions, such as the purchase of hundreds of autonomous drones or the closure of bases, can be based on the results of massive military simulations, which are easier to defend before Congress than a single wargame held behind closed doors.
Finally, WarMatrix is part of a context of competition with China and Russia. Washington knows that its adversaries also use wargaming tools and wants to avoid falling behind in terms of learning speed.
A global race for AI in wargaming
China, an AI laboratory for war simulation
China is probably the main competitor in this field. Since the late 2010s, wargaming research within the People’s Liberation Army (PLA) has incorporated expert systems, then more modern AI, to simulate complex operations, particularly around Taiwan and in the South China Sea.
The PLA promotes the concept of “intelligentization” of warfare: AI should help compensate for the lack of real operational experience through intensive computerized exercises. Chinese military academies hold wargaming competitions, sometimes with AI playing the role of the enemy, and use automatic language processing to extract the enemy’s doctrine from texts before integrating it into simulation modules.
Beijing has also created a Strategic Support Force tasked, among other things, with infusing AI into C2 and simulation. Recent research shows that Chinese models are attempting to merge satellite, cyber, and electronic data to generate increasingly realistic representations of conflict.
Western allies and NATO are not standing on the sidelines
Other Western armies are advancing at different rates. In the United Kingdom, the Ministry of Defense, through Dstl and the AI Research Center for Defense, is funding programs that explore how AI can automate the preparation, execution, and analysis of experimental wargames.
NATO, for its part, has introduced a “Wargame Domains Platform” integrating AI modules for multi-domain scenarios, with an emphasis on interoperability between nations. Conferences dedicated to modeling and wargaming, such as the one held in Verona in 2025, show that the Alliance considers simulation to be a key tool for preparing for complex crises, including cyber crises.
France and other European countries are also exploring these topics, often through partnerships with digital industry players. NATO’s acquisition of an AI-based command system developed by Palantir illustrates the convergence between real-time planning tools and digital wargaming platforms used upstream to test plans.
Russia, more discreet but interested
Public information on Russian AI wargaming is limited. Moscow communicates more about AI applied to drones, electronic warfare, and reconnaissance than it does about strategic simulation. Nevertheless, the tradition of operational modeling inherited from the USSR, combined with recent efforts to develop national AI, makes it very likely that advanced simulators are being used to prepare operations, particularly in Ukraine.
It is reasonable to assume that Russia uses models fed by feedback from the front lines to test combinations of drones, artillery, and electronic warfare before their actual implementation. The difference with the United States lies mainly in transparency: while Washington releases detailed RFIs, Moscow keeps its AI work for wargaming under wraps.
The limitations and risks of AI-driven wargaming
The rise of military artificial intelligence in wargaming is not without its dangers. The first relates to data quality. If models are trained on incomplete, biased, or politically filtered information, the results may offer an illusion of control while reinforcing erroneous views of the balance of power. Analysts are already warning of this risk in the case of China.
The second risk is psychological. By manipulating thousands of runs, decision-makers may lose touch with the human reality of war: civilian casualties, political chaos, emotional reactions. An AI that “wins” a wargame may propose strategies that are politically or morally unacceptable, such as rapid escalation in the nuclear or cyber domains.
The third risk is strategic. If several powers rely on AI systems to validate their plans, with different assumptions but strong certainties, the risk of misunderstandings increases. Two opposing models may each “predict” a quick victory, fueling excessive optimism and the temptation to take dangerous initiatives.
The US Air Force says it wants to keep humans at the center of the process, using WarMatrix as a decision-making tool rather than an oracle. But operational and political pressure always pushes for simplification of messages. How the results of these digital war games are presented to decision-makers will matter as much as the sophistication of the algorithms.
Ultimately, this high-speed wargaming project illustrates a profound shift: the decisive battlefield is also moving into data centers and simulation models. Those who learn the fastest and know how to integrate AI without submitting to it will gain a real advantage in preparing for conflict. The US Air Force is betting that its digital sandbox will become this strategic advance laboratory; it remains to be seen whether its adversaries will go even further, faster, and with different rules of the game.
Sources:
– Defense News, “US Air Force wants AI to power high-speed wargaming,” December 9, 2025.
– SAM.gov / HAF A5, “Advanced Wargaming and Simulation Technologies for Integrated Force Design (WarMatrix Ecosystem),” RFI dated November 23, 2025.
– DefenseScoop, “Air Force sees opportunities for AI to improve wargaming,” April 12, 2024.
– Jamestown Foundation, “New Developments in PLA Artificial Intelligence War-Gaming,” 2019, and ISW studies on Chinese wargaming.
– Various analyses on AI and wargaming (CSIS, CETaS, Frazer-Nash, RAND, FPRI), 2020–2025.
Live a unique fighter jet experience