Military AI and the new global war economy

For most of the last century, the defence business was measured in arms and ammunition. The next layer of military power is harder to see. It sits inside data centres, cloud contracts, chips, models and software that help armed forces detect, analyse, decide and act faster.

That shift is creating a new market around military AI. The numbers are already large, although estimates vary sharply because analysts define the market differently. Grand View Research valued military AI at $9.3 billion in 2024 and projects $19.3 billion by 2030. Fortune Business Insights puts the market at $18.75 billion in 2025 and forecasts $101 billion by 2034. The disagreement is itself revealing: AI is no longer a neat defence-product category. It is becoming a layer across intelligence, targeting, logistics, cyber operations, training, command systems and autonomous platforms.

SIPRI reported global military expenditure reached a record of $2.887 trillion in 2025, the eleventh consecutive annual increase. Europe spent $864 billion, up 14% in a year, while Asia and Oceania reached $681 billion. The US, China and Russia together accounted for 51% of global military spending. Not all of that money is going into AI, of course, but this spending environment is creating demand for capabilities promising faster decisions and greater autonomy.

The business winners may therefore look different from the traditional military-industrial complex. Lockheed Martin, RTX, BAE Systems and other established contractors are integrating AI into existing platforms. But the emerging defence stack also includes semiconductor companies, cloud providers, frontier-model developers, data platforms and autonomy startups. The military-AI market already spans firms ranging from Palantir, Microsoft, Google and Amazon to Anduril alongside conventional defence manufacturers.

Washington is making this convergence explicit. In May, the Pentagon reached agreements with seven AI companies to deploy advanced AI capabilities on classified networks. Reuters reported that the Pentagon's GenAI.mil platform had already been used by more than 1.3 million personnel within five months. The customers are no longer buying a finished weapon alone; increasingly, it is purchasing compute, models, data access, integration and continuously evolving software.

Palantir illustrates the economics. Reuters reported that its US government revenue rose 84% year-on-year to $687 million in the first quarter of 2026, as demand for its AI and data software accelerated. The Pentagon has also moved to institutionalise Maven, the system that integrates information from satellites, drones, radars, sensors and intelligence reports. Once software becomes part of how a military sees the battlefield, contracts can become longer-lived and changing suppliers becomes strategically complicated.

Venture capital has noticed. Anduril, founded only in 2017, reported $2.2 billion in 2025 revenue and, according to Reuters, has been discussing financing at a valuation of around $100 billion. In Europe, Germany-based Helsing raised $1.8 billion in July at an $18 billion valuation, after beginning as an AI battlefield-software company and expanding into autonomous strike drones, underwater systems and military aircraft applications. German drone maker Quantum Systems separately raised $1.2 billion. Defence technology is moving from a niche many investors once avoided into a major destination for private capital.

This changes the economics of military procurement. A conventional platform may remain in service for decades with periodic upgrades. AI systems can require continuous retraining, model updates, cybersecurity patches, new datasets and specialist support. SIPRI notes that military AI products may need constant monitoring, updates and refinements from industry suppliers during actual use. That creates recurring commercial relationships, but also dependency. A state may own the aircraft while relying on a private company for part of the intelligence layer that makes it useful.

Sovereignty is consequently becoming a business issue as much as a military one. European governments are already debating dependence on American defence software. Reuters reported in October that Palantir had to reassure European customers that neither the company nor Washington could remotely disable Maven, which NATO acquired in 2025 and made fully operational in June 2026. Germany and France have nevertheless expressed concerns around technological dependence and control. In the AI era, national security increasingly includes ownership of software architecture, cloud infrastructure, training data and update rights.

There is also a structural change in who participates in war. SIPRI observes that civilian technology companies and engineers may support AI systems throughout their operational life. Data centres serving households and businesses may simultaneously support military workloads. Cloud companies, semiconductor manufacturers and model developers can therefore become part of the defence supply chain without resembling traditional arms manufacturers.

This creates an awkward economic paradox. Militaries want the speed of commercial innovation, but commercial AI evolves faster than conventional defence procurement, regulation and weapons testing. Governments also compete with technology firms for engineers and data scientists. Meanwhile, startups dependent on a small number of government customers face long procurement cycles and the possibility that impressive demonstrations never become sustained programs.

The old defence industry is not disappearing. Tanks, missiles, submarines and aircraft will remain central to military power. But their value will increasingly depend on the digital systems connecting them. The emerging military-industrial complex is therefore not replacing steel with software; it is wrapping steel in software.

The most consequential defence company of the next decade may be the one controlling the data, compute and algorithms telling every other weapon where to look, what to believe and when to act. That's quite dangerous, isn't it?