The most powerful military in the world ran out of AI credits. Not because of a budget cut, not because of a congressional hold, because the soldiers and civilians who work for it simply used them all up.
In May 2026, the US Army’s Chief Information Officer announced that employees in its Combat Capabilities Development Command (DEVCOM) would have unlimited access to AI tokens through the Ask Sage platform, which provides access to models including Google’s Gemini, Meta’s Llama, and OpenAI’s ChatGPT for handling controlled unclassified information. The announcement was framed as a vote of confidence in the military’s AI adoption push — the Pentagon had recently bragged that nearly half of its 3.5 million employees were using AI tools daily.
By mid-June, the token pool was empty.
What 100 million tokens looks like
An internal email obtained by Wired informed DEVCOM staff that the Army’s annual enterprise subscription — 100 million tokens — had been exhausted within weeks of the “unlimited” declaration. Individual employees had been allocated at least 200,000 tokens per month with automatic top-ups available. The pool was designed to last a year. It lasted about a month.
The email said limits had to be re-established, and that renewal after October 1 was uncertain.
The operational context makes the numbers even starker. During the 38-day Operation Epic Fury in Iran, the broader Department of Defense burned through approximately 20 billion tokens per day for Project Maven, the Pentagon’s AI targeting system. The administrative token pool for DEVCOM — which covers tasks like reclassifying personnel descriptions and drafting documents — was a rounding error by comparison. But it was also the pool that everyone could access, and everyone did.
Not just a military problem
The Army’s token crisis is part of a broader pattern that is playing out across the corporate world. Meta encouraged employees to “tokenmaxx” — compete to use as many AI tokens as possible, complete with a public leaderboard, before quietly removing the leaderboard and trying to rein in usage. Instagram head Adam Mosseri recently floated the idea of capping token use per engineer. Uber saw its engineers burn through an entire year’s worth of generative AI tokens in four months, according to Fortune.
The pattern is the same everywhere: organizations announce unlimited AI access, discover that “unlimited” was a budget number rather than a technical one, and scramble to re-establish limits after the fact. The Army’s version stands out only because the stakes are higher. When a private company runs out of tokens, productivity dips. When the military runs out, the question becomes who decides which missions get AI support and which do not, a command decision the Pentagon did not anticipate having to make.
Reliability questions
Compounding the resource problem is a credibility problem. An Army employee who spoke to Wired anonymously said the generative AI tools had not been particularly useful. When they did use them, the results were unreliable, one model asserted that it had completed a task it had not.
The contrast is difficult to ignore. The Pentagon is simultaneously cutting human-staffed civilian protection oversight, the unit responsible for preventing casualties in conflict zones was recently reduced, while developing AI tools to replace those assessments. The Army is running out of tokens for writing personnel descriptions while investing heavily in AI for lethal operations.
Whether the military can build the governance infrastructure to match its AI ambition remains an open question. The token crisis suggests the answer, at least for now, is that 100 million is not as big a number as it sounds.
Sources: The Army Is Burning Through Its AI Tokens (Wired, July 21, 2026); US Army faces AI use limits after exhausting year’s supply of AI tokens (Ars Technica, July 2026)