Accountability in military AI requires knowing who supplies it. Researchers at SIPRI have begun mapping the military AI industry precisely because the supply side shapes what capabilities exist to regulate. Drawing on such open mappings, procurement records and company disclosures, four distinct supplier types emerge — each with a different relationship to oversight.
1. Traditional defence primes
The established contractors — the Lockheed Martins and BAE Systems of the world — integrate AI into existing platforms and mission systems: sensor fusion in aircraft, predictive maintenance for fleets, electronic-warfare software. Their AI work is usually a component inside a larger platform contract, which makes it the hardest to see from outside: it appears in procurement records as an aircraft upgrade, not as "AI". Their advantage is certification experience and cleared workforces; their reputation among analysts is for moving slowly.
2. Defence-tech startups
The most visible category, and the one reshaping the market's economics. Companies such as Anduril and Shield AI build software-first, selling autonomy stacks and uncrewed systems developed with venture capital rather than cost-plus contracts. Palantir, an earlier generation of the same thesis, supplies data-integration and targeting-support software. The category's defining trait is publicity: unlike primes, startups market aggressively — which means more public information, but information shaped by fundraising incentives. A startup's stated capability should be read the way one reads any pitch.
3. Cloud and AI giants
The general-purpose technology layer — cloud infrastructure and frontier AI models — comes from companies whose primary business is civilian. Their military involvement runs through cloud contracts and, more recently, through defence-specific agreements around frontier models, a boundary that has visibly shifted: several leading AI labs revised earlier restrictions on military use of their models during the mid-2020s. This category matters disproportionately for governance because it controls the upstream capability everyone else builds on.
4. The dual-use grey zone
The largest and least mapped category: suppliers of components and software never marketed as military — commercial drones, computer-vision libraries, open-source models — that end up in military use anyway, most visibly in Ukraine's improvised drone ecosystem. This is the category export-control regimes struggle with by design: the technology is genuinely civilian until the moment it is not. Any industry map that only counts self-declared defence companies misses it entirely.
How to read market numbers critically
Reports valuing the "military AI market" in the billions circulate widely, and they deserve suspicion on three counts. Definitions vary enormously (does a fleet-maintenance database count as AI?); the numbers usually come from market-research firms whose customers are the vendors being counted; and contract values are recorded at award, though programmes are routinely restructured or cancelled. The defensible statement is qualitative: spending is growing, the supplier base is diversifying away from primes, and software is claiming a growing share of defence budgets. Precise figures beyond that are marketing wearing the costume of measurement.
Why a publisher tracks this
This site's interest in the industry is downstream of its interest in meaningful human control: who builds a system shapes what oversight it gets. A targeting aid from a cleared prime inside a certified programme, the same capability from a startup iterating at venture speed, and an improvised system assembled from commercial parts sit under three very different accountability regimes — and the regulatory debates summarised in our UN LAWS overview address mainly the first. The industry's centre of gravity is moving toward the second and third. That gap between where the rules point and where the market goes is, in our reading, the most under-covered story in military AI.
Agent AI Army is an editorial publication. We analyse publicly available sources — SIPRI, CSET, UNODA, procurement records and company disclosures — and neither provide nor broker any technology or services. Company names appear as subjects of analysis, not endorsements.