US Army Faces AI Token Depletion, Questions ‘Unlimited’ Use

The United States Military is rapidly eating up its artificial intelligence capabilities, leading to an internal warning about the supposed “unlimited” nature of AI tokens. Troops recently received an email notification informing them of the accelerated expenditure of these resources.

This development comes just over a month after the Department of Defense (DoD) highlighted that almost half of its 3.5 million employees already employ artificial intelligence in their work routines. However, the Army’s Combat Capabilities Development Command (DEVCOM) in particular has been warned about excessive token consumption and the need to restrict usage.

According to the statement sent to the military, although the Army CIO had promised unlimited tokens in May 2026, the supply ran out in mid-June, requiring the reintroduction of limits. The email also mentions that, despite the decision to renew the use of tokens at current levels, there is uncertainty regarding the replenishment of the Army CIO stock after October 1 of the same year.

The Army uses Ask Sage, a multimodal generative AI platform that allows users to access several large-scale language models (LLMs), including Alphabet’s Gemini, Meta’s Llama, and OpenAI’s ChatGPT. An Army official, who preferred to remain anonymous, commented that “apparently, the entire Army spent the entire year’s tokens for just one service.”

The Ask Sage platform serves to “power the Army’s corporate LLM workspace” and has accreditation to handle “Controlled Unclassified Information.” It is also used by the Department of Defense’s Digital and Artificial Intelligence Office (CDAO) in acquisition processes. According to the Army, Ask Sage has been used in tasks such as “reclassifying personnel descriptions, which involves defining and aligning roles, experience and job history.”

The military institution has been actively encouraging its employees to adopt the use of generative AI. E-mails internos indicam que os colaboradores recebiam uma cota mensal de pelo menos 200.000 tokens e tinham acesso a mais tokens automaticamente caso exaurissem a cota inicial. Employees registered with Ask Sage, but with low usage, received incentives to increase their consumption.

To operate Ask Sage, the Army had access to 100 million tokens annually through an “enterprise package.” A token represents a unit of output from an LLM, whether text or image. For the Ask Sage tool, a single token is equivalent to about 3.7 characters, according to internal documents. In a previous event, the Department of Defense consumed approximately 20 billion tokens per day during the 38 days of Operation Epic Fury in Iran.

The Army, the Department of Defense and the Ask Sage platform did not respond when contacted for comment on the situation.

It is not yet known whether the tokens consumed by regular Department of Defense employees come from the same pool used for AI tools that process classified or secret information. Despite the challenges, the Department of Defense continues its investments in AI. Recent reports indicate that the Pentagon continues to prioritize artificial intelligence tools, while reducing the number of staff at the Civil Protection Center of Excellence, which was previously dedicated to preventing casualties in conflict zones. Instead, the Department of Defense is developing an AI tool to streamline the assessments this team would perform.

The Army is not the only organization to reconsider the seemingly limitless use of generative AI. After encouraging its employees to “maximize” their token usage, Meta has quietly removed its token consumption ranking and is now seeking to restrict usage. Last week, Adam Mosseri, head of Instagram at Meta, even suggested limiting the use of tokens by engineers at the company. Additionally, Uber watched its engineers consume a year’s worth of generative AI tokens in just four months.

The anonymous Army employee expressed skepticism about the usefulness of generative AI tools in his work, citing their unreliability. He reported that a model even declared the completion of a task that had not been performed. “I believe that there are undoubtedly many aspects of the U.S. federal government’s bureaucracy in which these tools could be useful. However, thoughtless application and use will not result in effective, efficient and reliable implementation,” he said.