The War Machine Economy, Part 1: Building AI Weapons

In January 2024, OpenAI's usage policy said its tools could not be used "to develop or use weapons." By February 2026, the company had signed a classified agreement letting the Pentagon run its models on military networks for, in the contract's own language, any lawful purpose.

Nobody at OpenAI announced a change of values. The policy page was simply edited, a sentence banning military use quietly replaced with language permitting "national security" work, and the company moved on. Google did something almost identical thirteen months earlier, deleting a public promise, in place since 2018, that it would never build AI for weapons or surveillance.

Four companies spent two years walking back the same promise, in the same order, for the same reason: a competitor was about to get the contract, the money, and the government relationship instead. The war machine economy did not arrive because Silicon Valley changed its mind about weapons. It arrived because the price of principle finally came due.

The Pledge That Didn't Survive Contact With the Market

Google's 2018 AI principles were written under duress. Thousands of employees had signed a letter protesting Project Maven, a Pentagon program using Google's machine learning to analyze drone footage, and the company let the contract lapse rather than renew it. The principles that followed included an explicit line: Google would not develop AI for weapons or for surveillance that violated international norms.

That line is gone. In February 2025, Google rewrote its AI principles page, replacing the weapons pledge with a statement that it would proceed wherever "the overall likely benefits substantially exceed the foreseeable risks." Company leadership framed it as a response to geopolitical reality: democratic countries developing AI, the reasoning went, have an obligation to make that technology available to their own militaries rather than ceding the field.

Meta had already made the same move three months earlier, in November 2024, quietly reversing its ban on using Llama models for military and intelligence applications. The new policy opened the door to the Pentagon, defense contractors like Lockheed Martin and Booz Allen, and the defense-tech firms Palantir and Anduril, all of whom could now build on Meta's models instead of a rival's. Nick Clegg, then Meta's president of global affairs, cast it as backing "responsible" uses that supported democratic allies in a global AI contest, not a retreat from principle.

A Grab for a Market That Didn't Exist Two Years Ago

What these reversals share is timing, not just logic. Each one followed a surge in defense spending on AI that simply did not exist at the scale it does now.

The Pentagon's fiscal year 2026 budget included a dedicated line for AI and autonomous systems worth 13.4 billion dollars, the department's first time carving out that category on its own, according to Congressional Research Service analysis. Its Replicator initiative, aimed at fielding thousands of low-cost autonomous systems to counter China's larger military, committed roughly a billion dollars to that goal alone.

Money like that reorders a boardroom's priorities fast. A company that keeps a blanket no-military-use clause isn't just making an ethical choice anymore; it's opting out of a contract pool now worth tens of billions of dollars a year, one that a rival will happily take instead.

Did You Know?

The Pentagon's Chief Digital and Artificial Intelligence Office signed four separate 200 million dollar contracts in 2025 for "agentic AI" prototypes, and the winners were Anthropic, Google, OpenAI, and xAI, the same four labs that would later split over how far those relationships should go.

The Holdout, and What Holding Out Cost

Anthropic took a different path, for a while. When the Department of Defense asked contracted AI companies in early 2026 to agree their models could be used for "any lawful purpose," including systems the company read as covering fully autonomous weapons and domestic mass surveillance, CEO Dario Amodei refused, saying the company could not in good conscience accede. At a tense meeting, Defense Secretary Pete Hegseth reportedly raised the possibility of canceling Anthropic's existing 200 million dollar Pentagon contract over the refusal.

OpenAI moved into the opening. On February 27, 2026, it signed its own classified agreement to deploy models inside the Pentagon's most secure networks, a deal explicitly pitched as filling the gap Anthropic's refusal had left.

The backlash was immediate: employees and outside critics called the rollout reckless, and Sam Altman himself later admitted it had looked "opportunistic and sloppy." OpenAI amended the contract days later, on March 2, explicitly barring use of its technology for domestic surveillance of U.S. persons and walling off intelligence agencies from this specific deal.

Anthropic and the Pentagon reopened talks soon after, according to reporting at the time, but the company's standing had already shifted. When the Pentagon announced classified-network contracts with eight companies on May 1, 2026, covering OpenAI, Google, Microsoft, Amazon Web Services, Nvidia, Oracle, SpaceX, and the startup Reflection AI, Anthropic's name was missing.

The Defense Department had formalized a supply-chain risk designation against the company in March, the same month its contract dispute peaked. Refusing the broadest terms had kept Anthropic out of the broadest deal.

What the Startups Already Knew

While the frontier labs were rewriting their principles, a different kind of company had been building a business on the assumption that defense money was the whole point from day one. Anduril Industries, founded in 2017 specifically to sell autonomous weapons systems and surveillance technology to the U.S. military, closed a funding round in May 2026 that valued it at 61 billion dollars, doubling its valuation from less than a year earlier. In March 2026, the Army awarded the company a ten-year contract with a ceiling of 20 billion dollars, folding more than 120 separate smaller contracts into one.

Palantir had already shown the shape of that playbook. In July 2025, the Army signed its own ten-year enterprise agreement with Palantir worth up to 10 billion dollars, consolidating 75 existing software contracts into one. Palantir's market value has since climbed past 250 billion dollars, a number no pure enterprise-software company reaches on commercial contracts alone.

Those figures were impossible to ignore from inside Google, Meta, OpenAI, or Anthropic's boardrooms. Every dollar the Pentagon spent with Anduril or Palantir was a dollar it wasn't spending with a frontier AI lab, and every year a lab kept its no-military clause in place was a year it watched that market get carved up by companies with no such clause to begin with.

Anduril's 2025 revenue was close to 2.2 billion dollars, a fraction of what OpenAI or Google earns from any single product line, yet its valuation grew faster than nearly every AI lab's because almost all of that revenue came from one buyer who wasn't going anywhere: the U.S. government.

The Infrastructure Bill Behind the Policy Reversal

There's a second force behind these reversals that gets less attention than the ethics debate: the AI buildout itself has become extraordinarily expensive to sustain, and defense contracts are one of the few revenue sources large and durable enough to matter at that scale. The capital pouring into data centers and chips has made steady, long-term government revenue newly attractive to companies used to funding growth with investor cash. The AI boom's financing has grown so large that bond investors who once lined up five-to-one for tech debt had cut that demand in half within months, a sign of how much pressure these companies are under to find revenue that doesn't depend on continued investor enthusiasm.

Much of that infrastructure runs on chips from a single company that spent two decades positioning itself for exactly this moment. Nvidia's unlikely two-decade bet on a chip feature nobody initially wanted is now also a defense story: Nvidia was one of the eight companies named in the Pentagon's May 2026 classified-network contracts, putting the same chips that train chatbots directly into military data infrastructure.

Geopolitics sharpened the calculation further. Executives at Palantir and Scale AI have argued publicly that the real choice isn't whether AI gets militarized, but whether American or Chinese systems do it first, pointing to Chinese state spending on military AI that they say dwarfs the Pentagon's own budget for it. That argument, whatever its merits, gave boardrooms a ready-made justification that didn't require admitting the decision was also, simply, about not leaving billions of dollars on the table.

Knowlegic Perspective

It's tempting to read this as a morality story, four companies abandoning principles under pressure. But the more useful read is a market one: a no-military-use clause was always going to be a luxury good, affordable only as long as no competitor was winning billion-dollar contracts by not having one. Once Anduril hit a 61 billion dollar valuation on government revenue alone, and Palantir's enterprise deal with the Army topped 10 billion dollars, the clause stopped being principle and started being a line item competitors didn't have to carry.

The companies that held firmest, like Anthropic for a stretch, didn't win a prize for it. They got excluded from an eight-company contract and had to renegotiate from a weaker position. That's the actual lesson of the last two years: in a market this large, a policy against a category of customer isn't a permanent stance, it's a bet on how long that customer stays small enough to refuse.

The war machine economy did not arrive because Silicon Valley changed its mind about weapons. It arrived because the price of principle finally came due.

Sources & References

• Google removes pledge to not use AI for weapons, surveillance, CNBC (2025)
• Meta changes its tune on defense use of its Llama AI, The Register (2024)
• The new 'land grab' for AI companies, from Meta to OpenAI, is military contracts, Fortune (2024)
• OpenAI's 'compromise' with the Pentagon is what Anthropic feared, MIT Technology Review (2026)
• Pentagon strikes AI deals for classified military use, The Washington Post (2026)