The Army Is Rapidly Depleting Its AI Resources

Shortly after the Department of Defense (DOD) declared that nearly half of its 3.5 million workforce was utilizing AI, members of the Army’s Combat Capabilities Development Command (DEVCOM) received a notification about rapidly consuming their token supply and the necessity to restrain usage.

“While the Army CIO proclaimed in May 2026 that unrestricted tokens were available, by mid-June the Army CIO’s token pool was depleted, necessitating the reinstatement of limits,” the email stated. It further mentioned that the Army decided to maintain token usage at “current levels,” though it remains uncertain “whether the Army CIO pool will be replenished post 1 Oct.”

The Army employs Ask Sage, a multimodal generative AI platform that allows users to operate various large language models (LLMs), including Alphabet’s Gemini, Meta’s Llama, and OpenAI’s ChatGPT. “It seems the entire Army exhausted its yearly token allocation for just one service,” an Army employee, who wished to remain anonymous, conveyed to WIRED, citing lack of authorization to speak publicly.

Ask Sage serves the Army as a means to “power its enterprise LLM workspace” and is “accredited for Controlled Unclassified Information.” The DOD’s Chief Digital and AI Office (CDAO) also utilizes it for acquisitions. As noted on the Army’s website, Ask Sage has been utilized for tasks such as “reclassifying personnel descriptions, which includes defining and aligning job duties, experience, and backgrounds.”

According to the Army employee, there has been a strong push for workers to adopt generative AI. Employees were allocated a minimum of 200,000 tokens each month, as indicated by emails reviewed by WIRED, and could receive additional tokens automatically once their initial allotment was exhausted. Individuals who registered for Ask Sage but were not frequently using it received emails urging them to consume their allocated tokens.

To access Ask Sage, the Army was granted 100,000,000 tokens under an annual subscription to an “enterprise pack.” Tokens serve as a unit of output, whether in text or image, from an LLM, with one token equating to about 3.7 characters, per documents reviewed by WIRED. During the 38-day Operation Epic Fury in Iran, the Defense Department reportedly used approximately 20 billion tokens daily, according to Breaking Defense.

Requests for comments went unanswered from both the Army and DOD, as well as Ask Sage.

It remains unclear if the tokens utilized by regular DOD employees originate from the same pool as those using AI tools for classified or secret information. However, this has not deterred the Defense Department from its AI initiatives. On Monday, the Intercept revealed that the Pentagon is intensifying its AI endeavors while reducing staff at the Civilian Protection Center of Excellence, which focused on preventing civilian casualties in conflict zones. Instead, the DOD is developing an AI tool to expedite the assessments typically conducted by the Center’s team.

The Army isn’t the first organization eager to adopt generative AI to reconsider its previously boundless usage. After motivating employees to “tokenmaxx,” Meta discreetly withdrew its leaderboard for token tracking and is actively trying to limit usage. Recently, Adam Mosseri, the head of Instagram at Meta, suggested the possibility of capping token usage per engineer within the company. Similar reports from Fortune noted that Uber encountered its engineers consuming an entire year’s worth of generative AI tokens within just four months.

The Army employee expressed that they have not found generative AI tools particularly beneficial for their tasks and noted instances of unreliability when using them. One model even claimed to have completed a task it had not, according to them. “I believe there are certainly aspects of the bureaucracy of the US federal government where these tools could be beneficial. Yet, an indiscriminate application will not yield an effective, efficient, and trustworthy deployment.”