The US Army gave troops unlimited AI. They used a year's worth in six weeks

In May 2026, the US Army handed its workforce what looked like a free hand on generative AI: an effectively unlimited token budget on Ask Sage, the service's enterprise GenAI platform. By mid-June, the year's supply of roughly 100 million tokens was gone, and an internal email verified by Wired has confirmed the cap is back on. The "unlimited" experiment lasted about six weeks.

The burndown is a procurement story more than a capability story. A token is roughly a piece of a word, and an enterprise license can burn through millions of them on a single long document or agent loop. Ask Sage's annual pack contained 100 million tokens, or about 200,000 per employee per month if the Army had spread them evenly. It did not. "The whole Army burned through the whole year of tokens for just one service," one Army employee told Wired, before the first month was out.

During the 38-day Operation Epic Fury against Iran, the Defense Department as a whole used roughly 20 billion AI tokens a day, according to Breaking Defense. That is roughly two hundred Army annual pools, every single day of the operation. The campaign also saw AI used for targeting and battle-damage assessment. At operational scale, AI stops being a per-seat line item and becomes a per-output line item, and the Army's unlimited policy was a per-seat policy with a per-output price tag.

Ars Technica's independent coverage and the original TechRadar report corroborate the timeline. The mechanism was foreseeable. Automatic top-ups for users who exhausted their initial allocation made the original cap meaningless before it was ever tested. Army leadership had actively encouraged broad GenAI use across roles to surface productivity gains; the cost realism behind that push was the missing piece.

What the Army just learned in public is the lesson every agency and enterprise is about to learn when GenAI moves from pilot to operation. Vendor pricing is shifting from flat per-seat subscriptions to consumption-based and output-based models. That shift turns usage from a fixed cost into a variable cost, and variable costs punish budgets designed for fixed ones. A perfectly forecasted adoption curve would not have changed the answer: the more troops actually used the tool the way the Army wanted them to, the faster the pool emptied.

Token availability beyond October 2026 is no longer guaranteed, Wired reports, with the Army concerned about rising costs. A cluster of related coverage tracks how other agencies are running the same arithmetic. The next test is not whether the Army rolls tokens back out, but whether the next unlimited policy comes with a per-output line item built to absorb what usage actually looks like.