The US Army burned through a year of AI tokens in six weeks

The US Army Burned Through a Year of AI Tokens in Six Weeks

In May 2026, the US Army told its workforce they had unlimited access to generative AI. By mid-June, the annual allocation was gone. 1

The Army had purchased 100 million tokens through an annual subscription on Ask Sage, a multimodal generative AI platform where users can run models including Alphabet's Gemini, Meta's Llama, and OpenAI's ChatGPT. 1 That pool, budgeted to last until October, ran dry in roughly six weeks. An internal email shared with WIRED confirmed the Army CIO's token pool "was exhausted," forcing the service to cap usage. 1

The signal here is not that the military uses AI. It is that the constraint on institutional AI adoption is not model capability but procurement economics. Organizations are buying AI compute like they bought software: predictable annual subscriptions for a fixed allotment. But AI consumption does not behave like software usage. It is metered, elastic, and when leadership tells the workforce the faucet is open, it drains faster than any budgeting cycle can account for.

The Department of Defense had trumpeted that nearly half of its 3.5 million employees were using AI at work just before the Army's pool ran out. 1 Soldiers received at least 200,000 tokens per month on Ask Sage, with automatic top-ups when they exhausted their allotment. Those who signed up but rarely used the tool got emails encouraging them to consume more. 1

Scale puts the shortfall in perspective. One token equals roughly 3.7 characters of output, according to Army documents reviewed by WIRED. 1 During the 38-day Operation Epic Fury in Iran, the Defense Department consumed an estimated 20 billion tokens per day. 1 The Army's 100-million-token annual pool is a rounding error against that burn rate.

The Army is not an isolated case. It is the third high-profile organization in recent months to misjudge how fast AI capacity gets consumed. At Meta, leadership had encouraged employees to "tokenmaxx" with an internal leaderboard tracking usage, before quietly removing the leaderboard and pivoting to curbing consumption. Adam Mosseri, head of Instagram at Meta, floated capping token use per engineer. 1 At Uber, engineers burned through a year's worth of generative AI tokens in four months. 1

Three organizations, same arc: unlimited access, enthusiastic adoption, rapid depletion, abrupt caps. The variable is not recklessness. Nobody, from the Pentagon to Big Tech, has figured out how to forecast demand for a resource that expands to fill whatever you offer.

That gap between forecasting and consumption matters beyond defense budgets. If the largest procurement operation on earth cannot predict its own token burn six weeks out, every CIO, cloud provider, and taxpayer should be asking what budgeting model actually works for AI at scale. Annual subscriptions to fixed token pools assume usage is knowable in advance. The evidence says otherwise.

Ask Sage, which the Army uses to "power its enterprise LLM workspace," is accredited for Controlled Unclassified Information and also serves the DOD's Chief Digital and AI Office for acquisitions. 1 The platform supports more than 150 models across cloud, on-premise, and classified environments. 2 The Army used it for tasks like reclassifying personnel descriptions, defining and aligning job duties, experience, and backgrounds. 1

Whether the output justified the spend is unclear. The Army employee who spoke to WIRED anonymously said they had not found the tools particularly useful, and that one model claimed to have completed a task it had not. 1 The internal email noted it was unclear whether the token pool would be renewed after October 1. 1

Meanwhile, the Pentagon is expanding its AI footprint in other directions. The Intercept reported that the DOD cut staff at its Civilian Protection Center of Excellence, which works to prevent civilian casualties in conflict zones, and is developing an AI tool to accelerate the assessments those staff previously handled. 1

The Army, the DOD, and Ask Sage did not respond to requests for comment. 1

More AI, fewer people doing the work, and no reliable way to budget for the compute that connects them. That is not a military story. It is a picture of what happens when AI demand outpaces every model we have for supplying it.