Autonomous weapons systems—drones and platforms that select and engage targets with little or no human input per strike—pulled more than $10 billion in private capital in 2025, a structural shift pulling leverage away from traditional prime contractors such as Lockheed Martin…
Defense capability is no longer built the way it was when Lockheed Martin and Northrop Grumman were the only phone numbers a procurement officer had to know. The new description, in EE Times reporter Rebecca Pool's June 18 roundup of the year in defense capital, is that autonomous capability is "increasingly shaped in software, trained in simulation, and improved through use." That phrasing reads like marketing. It is actually a description of a development loop, and that loop is moving decision rights, capital, and engineering talent out of the traditional defense primes and into a wider circle of software companies, investors, and procurement offices that did not exist as serious players a decade ago.
The dollar trail is the easiest place to see the shift land. S&P Global, as cited in Pool's reporting for EE Times, put worldwide private equity and venture investment in aerospace and defense at $10.63 billion in 2025, more than double that of 2024's total. That private pool is small against the Stockholm International Peace Research Institute's $2.89 trillion tally for 2025 global public military spending. The reason the private number matters is the slope. Anduril closed a $5 billion round in May 2026 that the company describes as one of the largest private investments in defense history. Shield AI added $1.5 billion, True Anomaly $650 million, and Harmattan AI $200 million to the same capital stream, all reported in the same EE Times roundup.
The four companies are not doing the same thing. Shield AI builds autonomous intelligence, surveillance, and reconnaissance (ISR) drone swarms. True Anomaly works on autonomous orbital vehicles. Harmattan AI is a European drone-interception and electronic warfare play. Anduril is the broadest of the group, with autonomy software, counter-drone systems, and a manufacturing push that co-founder and CEO Brian Schimpf tied directly, in an investor letter quoted by EE Times, to aggressive expansion of manufacturing and to accelerating research and development on autonomous drones. Read across the four, the common thread is not a weapon. It is a release cycle: model in a simulator, fly the result, feed the data back, and ship the next version in weeks rather than the years a hardware program consumes.
Edge AI is the technical reason that loop can run at all. Inference on the device, rather than over a constant satellite link to a human operator, is what lets a drone pick a target and act on it in the time a radio round trip would burn. The traditional primes built platforms around the human-in-the-loop assumption. The new entrants are building around the assumption that the loop tightens on its own, with a human reviewing exceptions after the fact rather than authorizing every action in the moment. Schimpf's framing, in the same EE Times-sourced investor letter, that "defense was not a venture category" when Anduril was founded in 2017 but has changed meaningfully since, is the cleanest signal that the change is not just in the technology. It is in who gets to build it.
The shift in who builds it is also a shift in who has leverage. When a capability is defined by hardware procurement, the leverage sits with the program office, the prime contractor, and the congressional district that hosts the assembly line. When it is defined by a software release, the leverage moves toward the engineers who can write and validate the model, the simulation environment that trains it, the data set that distinguishes it, and the capital partner that funds the next training run. Procurement officers still set requirements, and primes still integrate platforms, but the binding constraint is increasingly a code commit, not a contract award. The funding is going to the teams that can iterate fastest, not to the teams that can deliver a fixed bill of materials.
The legitimate worry is not the speed. It is the closed loop. Aerospace and defense consultant Alessandro Pianelli, author of the white paper "Aerospace & Defence: 30 Year Outlook 2026–2056," noted in EE Times that deployments of nearly 10,000 AI-enabled drones, mostly in Ukraine, have achieved strike accuracies three to four times higher than human-piloted systems. "What we've seen happen in Ukraine is now expanding everywhere," he wrote. A defense industrial base shaped, trained, and improved by the same set of companies, on the same simulators, with the same capital partners and against requirements those same companies help write, is structurally hard to redirect. Arms-control researchers and autonomous-systems ethicists have spent several years arguing for external red lines on targeting autonomy, on the basis that the engineering incentives inside a software-defined defense market will not produce those red lines on their own. The new capital map is the empirical version of that argument. The faster the loop runs, the more important the question of who audits the loop becomes, and the less obvious it is that the people inside the loop will ask it of themselves.
What to watch next is whether any of the outside pressure points the new map creates—including congressional procurement reform, allied governments writing their own autonomy requirements, or export-control regimes that treat the training data and the simulator as controlled items—can move as fast as the engineering cycle they are trying to constrain. The S&P Global figure is a 2025 number. Anduril's round closed in May 2026. The policy conversation is still catching up to the build cycle those numbers describe.