The military is burning through its own AI codes

A little over A month after the Ministry of Defence splurge Since nearly half of its 3.5 million employees were using AI at work, members of the Army’s Development Combat Capabilities Command (DEVCOM) received an email informing them that they were burning code, and needed to limit its use.

“Although the Army CIO announced in May 2026 that they were offering an unlimited number of tokens, by mid-June, the Army CIO pool had exhausted its supply of tokens and had to re-establish limits,” the email said. The email goes on to say that although the Army has elected to renew token usage at “current levels,” it is unclear “whether the Army IT manager pool will be renewed after October 1.”

The army uses Ask Sagea multimodal generative AI platform where users can run various large language models (LLMs), including Alphabet’s Gemini, Meta’s Llama, and OpenAI’s ChatGPT. “Apparently the entire Army consumed tokens for the entire year for just one service,” says one Army employee who spoke to WIRED anonymously because he was not authorized to speak to the press.

Ask Sage Who user By the Army “to operate their organization’s LLM workspace” and is “certified for controlled unclassified information.” It is also used by the Department of Defense’s Principal Digital and Artificial Intelligence Office (CDAO) for acquisitions. According to the army WebsiteAsk Sage was used to complete tasks such as “reclassifying employee descriptions, which includes defining and aligning job duties, experience, and backgrounds.”

The Army employee says the Army has been pushing its workers toward using generative AI. Employees were given an allotment of at least 200,000 tokens per month, according to emails seen by WIRED, and were automatically allocated more if they used up their initial allotment. Employees who sign up for Ask Sage but don’t use it regularly will receive emails encouraging them to use more of their assigned tokens.

In order to use Ask Sage, the military has access to 100 million tokens as part of an annual “Enterprise Pack” subscription. Symbols Represents the output unit, whether text or image, of the LLM. For Ask Sage, one token is equivalent to about 3.7 characters, according to documents seen by WIRED. The Ministry of Defense burned some of them 20 billion symbols Daily during the 38-day Operation Epic Fury in Iran, according to Break the defense.

The Army and Ministry of Defense did not respond to requests for comment; Don’t ask Sage.

It is not clear whether the tokens used by rank-and-file DoD employees are drawn from the same pool as those who may use AI tools on classified or confidential information. But this has not stopped the Department of Defense from focusing on artificial intelligence. on monday, I mentioned the objection The Pentagon continued to rely on artificial intelligence tools, and reduced the number of employees at the Center of Excellence for the Protection of Civilians, whose jobs involved preventing civilian casualties in conflict zones. Instead, the Department of Defense is developing an artificial intelligence tool to speed up the assessments typically conducted by center staff.

The military is not the first to embrace generative AI to rethink its near-limitless use. After encouraging employees to use “tokenmaxx,” Meta fell silent Get down used the token to track the leaderboard and is now trying to limit its use. last week, Adam MosseriThe head of Instagram at Meta floated the idea of ​​setting a cap on token usage for every engineer in the company. According to a report from luckUber has also seen its engineers consume a year’s worth of generative AI code in just four months.

The Army employee says that they did not find generative AI tools particularly useful for their work, and that when they did use the tools, they found them to be unreliable. They say one model confirmed that he completed a task that he did not complete. “I think there are certainly many aspects of the U.S. federal government bureaucracy where these tools could be useful. But ill-considered application and use will not lead to an effective, efficient, and trustworthy launch.”