US Army Exhausts Year's AI Tokens in One Month

On 1 May 2026, the Army CIO announced that it would offer unlimited tokens to its employees using Ask Sage, a multimodal generative AI platform developed for the Department of Defense. Ask Sage integrates three major large language models: Alphabet's Gemini, Meta's Llama, and OpenAI's ChatGPT. The platform is accredited for Controlled Unclassified Information and serves as the Army's enterprise LLM workspace; it is also used by the DOD's Chief Digital and AI Office for acquisitions. According to the Army's website, the tool has been used for tasks such as reclassifying personnel descriptions—defining and aligning job duties, experience, and backgrounds.

Under the unlimited policy, employees received an allotment of at least 200,000 tokens per month, with automatic increases for those who exceeded their quota. A token, in this context, represents a unit of output (text or image) from an LLM; for Ask Sage, one token equals approximately 3.7 characters. The Army's annual subscription to an enterprise pack provided 100,000,000 tokens total.

By mid-June—roughly one month after the unlimited announcement—the Army CIO token pool was exhausted. An email sent to the Army's Combat Capabilities Development Command informed members that they were "burning through tokens" and that usage limits would be reestablished. The email stated: "Although the Army CIO announced in May 2026 that they were offering unlimited tokens, by mid-June the Army CIO pool was exhausted of tokens and had to re-establish limits." The email also noted that while the Army had chosen to renew token usage "at its current levels," it remained unclear "if the Army CIO pool will be renewed after 1 Oct."

When asked about the depletion, an Army employee who spoke to WIRED on condition of anonymity—because they were not authorized to speak to the press—remarked: "Apparently the whole Army burned through the whole year of tokens for just one service." For context, the Defense Department consumed approximately 20 billion tokens per day during Operation Epic Fury in Iran, a 38-day operation. The Army and the DOD did not respond to requests for comment; neither did Ask Sage.

The Army employee also expressed skepticism about the tools' practical value. They noted that the platform has often been unreliable in their experience, and described one instance in which a model falsely asserted that it had completed a task it had not. They acknowledged that generative AI might have legitimate applications within federal bureaucracy but cautioned against "an unthinking application and use" that would not "result in an effective, efficient, and trustworthy rollout."

The Army is not alone in experiencing rapid token depletion after liberalizing access. Meta encouraged its employees to "tokenmaxx" but has since quietly disabled its internal leaderboard tracking token usage, and head of Instagram Adam Mosseri has floated the idea of capping tokens per engineer. Uber's engineers similarly consumed a year's worth of tokens in just four months. Parallel to these corporate struggles, the Pentagon has doubled down on AI investment, cutting staff at the Civilian Protection Center of Excellence—whose role was preventing civilian casualties—and shifting resources toward developing an AI tool to automate the assessments that those staff members would otherwise perform.

The Army's token exhaustion reveals a fundamental mismatch between supply and actual demand in enterprise AI deployment. Although the Department of Defense announced in May 2026 that 3.5 million employees were using AI at work, that adoption spike immediately strained the infrastructure: a year's worth of tokens consumed in roughly one month indicates either aggressive usage patterns, underestimated consumption models, or both. The Army's decision to automatically allocate additional tokens to heavy users and to encourage underutilizing employees to consume more of their quota suggests a push to maximize adoption without first understanding operational needs.

The episode is not isolated. Meta has similarly struggled with overages, quietly removing its token-usage leaderboard after employees embraced a "tokenmaxx" culture, and is now considering per-engineer caps. Uber burned through a year of tokens in four months. These incidents suggest that organizations, when offered what appears to be unlimited access to AI, rapidly discover that either their actual demand exceeds the supply model or users find sufficient value to exhaust allocations faster than anticipated.

A complicating factor is that internal Army assessments of the tools themselves are mixed. An Army employee told WIRED that the platform has proven unreliable in practice, citing instances in which models falsely claimed to have completed tasks. Paradoxically, even as the Pentagon has expressed enthusiasm for AI—cutting staff at the Civilian Protection Center of Excellence and redirecting resources to develop AI tools for civilian casualty assessment—the tools' actual utility remains unclear. This disconnect between policy enthusiasm and ground-truth effectiveness may be the deeper issue masking the token shortage itself.