U.S. Army faces AI token limits after burning through years of supply

The U.S. Army has faced a sudden halt in its AI operations after exhausting its token supply, prompting a return to strict limits. According to an internal email, the Army’s Combat Capabilities Development Command (DEVCOM) was informed that the Army CIO pool of tokens had been depleted by mid-June 2026, despite having been promised unlimited access in May. The email noted that while the Army has decided to maintain its current token usage levels, it remains unclear if the pool will be replenished after October 1. The Army uses Ask Sage, a generative AI platform that supports multiple large language models, including Gemini, Llama, and ChatGPT. The Army employee described the situation as a result of widespread AI adoption, with workers encouraged to use the tools extensively.

Employees were given a minimum of 200,000 tokens per month, with additional tokens allocated if they used more than their initial allotment. Those not using the platform regularly received reminders to utilize their allocated tokens. The Army’s access to 100 million tokens annually was part of an enterprise subscription, with each token representing about 3.7 characters of output from an LLM. The Defense Department burned through 20 billion tokens daily during Operation Epic Fury, according to Breaking Defense. The Army and DOD did not respond to requests for comment, nor did Ask Sage.

It remains unclear if tokens used by regular DOD employees are drawn from the same pool as those used for classified or secret information. This hasn’t stopped the DOD’s push for AI, with the Pentagon continuing to prioritize AI tools despite recent cuts to the Civilian Protection Center of Excellence. The Army employee noted that while the tools may be useful for some bureaucratic tasks, they have been unreliable and often inaccurate. One model even claimed to have completed a task it hadn’t, according to the employee.

The Army employee emphasized that a hasty adoption of AI without careful consideration could lead to an ineffective and untrustworthy rollout.

Source: arstechnica