The Autonomous Defense Stack
Sensors, networks, autonomy software, and cheap expendable platforms are assembling into something the defense industry has never had before: a stack.
Bottom Line Up Front
A new architecture is taking shape in American defense, and it looks less like a weapons program and more like the computing industry: a stack of layers, each with its own economics and its own competitors.
The layers, from bottom to top: sensors that see the battlefield, networks that move the data, autonomy software that decides, and cheap "attritable" platforms — drones, boats, missiles — that act.
The Pentagon's Replicator initiative was the demand signal. Its first phase declared its goal of fielding thousands of autonomous systems achieved in August 2025, and its second phase — focused on defeating enemy drones — announced its first buys in early 2026.
The old defense giants dominate almost none of these layers. Startups like Anduril are building the platforms and edge software; Palantir runs much of the data layer.
The open question is who ends up owning each layer — and history says the software layers are where the durable profits sit.
The factory that builds airplanes like phones
Drive south of Columbus, Ohio, past the farm fields around Rickenbacker airport, and you'll find a billion-dollar factory that started producing jet-powered combat drones this spring. Anduril's Arsenal-1 began building its Fury autonomous fighter in March 2026 — ahead of its own July target — and plans to run its Barracuda cruise missiles and Roadrunner interceptors down the same lines, with the same workforce, reconfiguring in days instead of years.
That last detail is the whole story. A traditional defense plant is built around one exquisite product for decades. This one is built like a consumer electronics plant: standard processes, swappable products, volume over perfection.
Why would anyone build combat aircraft that way? Because the aircraft aren't really the product. The product is the system they plug into.
The stack, translated
Think about your phone: device, network, operating system, apps — four layers. No single company owns the whole thing, and the money pools very differently at each layer — device makers fight brutal margin wars while software players compound quietly.
Defense is assembling the same shape:
Sensors are the eyes — satellites, radars, cameras, sonobuoys, all producing torrents of raw data. Networks are the nervous system, moving that data to whoever needs it. Autonomy software is the brain — it fuses the sensor picture, identifies targets, and increasingly lets machines fly, swim, and hunt without a human joystick. And attritable platforms are the hands: cheap enough that losing one is a line item, not a tragedy. That's what "attritable" means — a weapon you can afford to lose.
The Pentagon's Replicator initiative was Washington's way of forcing this stack into existence: buy thousands of cheap autonomous systems fast, on commercial timelines, and make the bureaucracy learn how.
Who owns which layer
Here's where it gets interesting, because the stack is being carved up in real time.
The software layers are consolidating fastest. Palantir's Maven Smart System — the AI targeting and battlefield-awareness platform — saw its Pentagon contract ceiling raised from roughly $480 million to about $1.3 billion, and in 2026 the Defense Department moved to make it a formal program of record. Anduril's Lattice plays the same role at the edge, directly steering robotic systems. The two companies formally allied in late 2024 — an arrangement that would have been unthinkable among the old primes.
The platform layer is splintering in the opposite direction: dozens of drone makers, boat makers, and munitions startups competing on cost per unit — exactly what the bottom of a stack should look like.
And the traditional primes? They still own what the stack can't replace: submarines, stealth bombers, satellites, and the deep magazines of missiles. Follow the chain and you see the connection to everything else we cover: autonomy software needs chips and compute, platforms need small motors, batteries, and castings from a thin supplier base, and the factories need workers and electricity.
A murmuration of starlings can outmaneuver any falcon, though no single starling could. None of them is exquisite. The pattern is the weapon. That's the bet the entire stack is making.
Key Judgments
- Defense spending on autonomy will keep shifting from platform-first to software-first economics, with the data and autonomy layers earning the most durable margins.
- The primes-versus-startups fight ends in a barbell, not a takeover: primes keep exquisite platforms and munitions; new entrants take attritable mass and software.
- Replicator matters more as precedent than as procurement — its fast-buying pathways will outlive the initiative itself.
- Within three years, the binding constraint on the stack will be component supply chains and testing capacity, not funding or ideas.
Risks & Counterarguments
The honest case against: "fielded" is doing a lot of work in official statements. Declaring thousands of systems delivered is not the same as thousands of systems integrated, maintained, and trusted by commanders. Skeptics argue the bureaucracy's immune system is quietly winning — absorbing the language of speed while keeping the old timelines.
Autonomy may also under-deliver in contested conditions. Jamming, spoofing, and plain fog-of-war have humbled every generation of "revolutionary" military technology. And startups burning venture money at defense margins may consolidate or fail before programs of record arrive.
Why It Matters
Stacks create winners differently than platforms do. In computing, the companies that owned operating systems and networks outearned the companies that bent metal, for decades. If defense follows the same physics, the next decade's defense budgets — now at historic levels — will flow through a very different set of hands than the last five decades'. Understanding the layers now is how you avoid analyzing 2030's defense industry with 1990's map.
What We're Watching
- Whether Replicator's fast acquisition pathways get institutionalized in the services' base budgets or quietly starve after the initiative's spotlight fades.
- Arsenal-1's actual output: does Fury, Barracuda, and Roadrunner production hit promised rates by end of 2026?
- Whether Maven's move to a program of record survives congressional scrutiny — and whether competitors force an open-architecture requirement.
- Component sourcing rules: how strictly Washington enforces bans on Chinese drone parts, which sets the cost floor for the whole platform layer.
Sources: Congressional Research Service, "DOD Replicator Initiative: Background and Issues for Congress" (IF12611); Defense Innovation Unit Replicator materials; Office of the Under Secretary of Defense (Comptroller) budget justification books; DefenseScoop reporting on Maven and enterprise command-and-control. This is analysis, not investment advice.