The War Machine Economy, Part 3: The Payout

In March 2026, the U.S. Army folded more than 120 separate contracts it held with a nine-year-old company into a single deal worth up to $20 billion. The company, Anduril Industries, had never built a tank, a jet, or a ship. It builds software.

Two months later, investors handed Anduril another $5 billion, on top of the government money, and valued the company at $61 billion, roughly double what it was worth a year earlier.

None of that money bought a single additional bullet. It bought code, cloud infrastructure, and the people who write both, and tracing where it actually lands says more about how AI weapons get made than any description of the weapons themselves.

A Budget Line That Didn't Exist a Year Ago

For decades, the Pentagon's roughly $1 trillion annual budget buried artificial intelligence inside other accounts: research funding here, a drone program there, nothing a reporter or a rival company could point to as "the AI number." The 2026 budget request changed that. A senior defense official confirmed to reporters that the department had created a standalone $13.4 billion line for AI and autonomous systems, about seven times the size of the equivalent request a year before, the first time autonomy has gotten its own number in the budget rather than being scattered across other programs.

That single number is the top of a long funnel. Congress approves it, the Pentagon obligates it to contracts, and from there it splits into two very different channels depending on who's doing the building.

Did You Know?

The Department of Defense has an older name again. A 2025 executive order restored "Department of War" as an official secondary title, the name the agency carried from 1789 until Congress renamed it in 1947.

The Old Channel: Prime, Then a Long Line of Subcontractors

The traditional route is the one that's run since the Cold War. Five companies, Lockheed Martin, RTX, Boeing, Northrop Grumman, and General Dynamics, together receive close to a third of everything the Pentagon obligates in a given year, according to the Congressional Research Service. These are the "primes," the companies that sign the headline contract directly with the government.

A prime rarely builds a weapons system alone. By law, it has to flow a meaningful share of the work down to subcontractors, smaller firms that make a sensor, write a software module, or supply a part. Picture a general contractor building a house: the name on the contract is one company, but the plumber, the electrician, and the roofer are all separate businesses getting paid out of the same budget, layer by layer, long after the homeowner's check clears.

The New Channel: Software Companies Acting Like Primes

AI has started to bend that structure. Anduril and Palantir don't slot neatly into the subcontractor tier. Both have built platforms, Anduril's Lattice and Palantir's Maven Smart System, that increasingly get bought the way a prime contract gets bought: one large enterprise agreement instead of dozens of small purchase orders.

Palantir's path shows the pattern clearly. Its Maven software started as a roughly $480 million Army prototype deal in 2024, grew through a $795 million contract modification the following year to a ceiling near $1.3 billion, and now runs across more than 100,000 military users. The Army has since folded dozens of Palantir's other software contracts into a separate enterprise agreement worth up to $10 billion, the same consolidation move it made with Anduril. Government revenue, U.S. and international combined, made up about 54% of Palantir's total revenue in 2025, with U.S. government customers alone accounting for roughly 42%.

Interesting Fact: In June 2025, the Army commissioned four Silicon Valley executives, from Palantir, Meta, OpenAI, and Thinking Machines Lab, directly into the Army Reserve as lieutenant colonels, under a new unit nicknamed Detachment 201. None of them draw a defense-industry salary for it; the appointment is unpaid, part-time advisory work. It's a different kind of payout: access and influence over how the military buys technology, rather than cash.

What the Software License Actually Buys

A defense-AI "sale" rarely looks like handing over a finished product. It looks closer to a subscription. The government pays for continued access to a platform, Lattice or Maven, that keeps learning from new sensor and battlefield data, plus the engineers who tune it. That's part of why Anduril's Army deal runs as a 10-year framework rather than a one-time purchase, and why the reasoning-and-acting software loop underneath these platforms, the same simple idea that powers tools like Claude Code, is worth paying to keep current rather than buying once and walking away.

Running that loop at military scale also means buying enormous amounts of computing power, the same GPU-driven infrastructure that turned a gaming chip company into the backbone of the entire AI industry. Civilian AI companies have increasingly financed that infrastructure with debt, borrowing against future revenue. Defense-AI companies mostly haven't needed to: their compute and contract pipeline gets financed by venture capital betting on the next funding round, backed by a government customer that, unlike a bond investor, isn't going anywhere.

Where the Payout Lands for a Person, Not a Company

Money that flows into a company like Anduril or Shield AI still has to reach individual people, and here the defense-AI trade looks different from a traditional government job or a traditional defense-contractor career.

Anduril's senior software and autonomy engineers reportedly earn total compensation packages roughly in the $300,000 to $500,000 range, according to self-reported pay data compiled by Levels.fyi, with close to half of that often paid in equity rather than salary. Palantir pays a somewhat lower base, commonly cited in the $155,000 to $240,000 range, but leans just as heavily on stock, a bet that's already made some early employees wealthy as the company's valuation climbed past $400 billion on public markets.

That equity bet is the real lure, and it's a different offer than the one coming from consumer AI labs. In mid-2025, OpenAI's chief executive said Meta had dangled signing bonuses as large as $100 million to poach his researchers, cash paid up front rather than value that depends on a future round closing. Defense-AI startups mostly can't match that kind of day-one number. What they can offer is a valuation curve moving almost as fast: Shield AI, a drone-autonomy company, raised $2 billion in March 2026 at a $12.7 billion valuation, up from $5.6 billion a year earlier, and was reportedly in talks for a new round near $20 billion by September. An engineer holding options granted at the lower valuation watches their paper value more than triple in eighteen months without changing employers. The pitch is less "get paid more today" and more "own a piece of a company whose government backlog isn't about to cancel itself."

Knowlegic Perspective

Follow the money far enough and a pattern emerges: almost none of it stops near the thing it's technically paying for. A $20 billion contract ceiling doesn't buy $20 billion of finished hardware sitting in a warehouse. It buys a decade of software updates, cloud computing bills, and the salaries and stock grants of the people maintaining all of it, with the Pentagon as a customer who, unlike a bond market or a venture fund, has shown little appetite for pulling out.

That's a genuinely different kind of defense economy than the one built on battleships and fighter jets, where the payout was concentrated in a handful of giant one-time manufacturing contracts. This one pays out continuously, in licensing fees and equity appreciation, to people who mostly never leave an office park.

It's worth sitting next to a harder number: the cost of war that shows up decades later, in a survivor's health and income rather than a line in anyone's budget. The contractor chain's payout is counted in valuations and compensation packages. The other cost rarely gets counted at all, and it doesn't show up on anyone's cap table.

None of this money buys a bullet, a jet, or a bomb. It buys contracts, licenses, and equity grants, and it lands on people building software in an office park rather than a battlefield. That is still the real story: tracing where defense-AI money actually lands says more about how these weapons get made than any description of the weapons themselves.

Sources & References

• Defense Primer: Department of Defense Contractors, Congressional Research Service, Library of Congress
• Billions for new uncrewed systems and drone-killing tech included in Pentagon's 2026 budget plan, DefenseScoop (2025)
• DoD Says Budget Request Includes $13 Billion For Autonomy And Related Systems, Defense Daily (2025)
• Trump signs executive order renaming Department of Defense to Department of War, CNN Politics (2025)
• Army awards Anduril $20B contract with an eye toward counter-drone capabilities, DefenseScoop (2026)
• Pentagon surges Palantir Maven Smart System contract spending to more than $1B, Inside Defense (2026)
• Palantir Secures $795M US Army Modification Contract, Yahoo Finance (2025)
• Palantir lands $10 billion Army software and data contract, CNBC (2025)
• Anduril Industries Software Engineer Salary, Levels.fyi, self-reported compensation data (2026)