U.S. Army Depletes Annual AI Token Budget in Weeks After Unlimited Access Launch

The Department of Defense faced an unexpected constraint on its artificial intelligence ambitions just weeks into rolling out broad AI access. In mid-June—less than a month after announcing that roughly half of its 3.5 million employees were actively using AI tools—the Army's token pool for its generative AI platform ran dry, prompting a sharp reversal in policy.

How Token Budgets Collapsed#

Army personnel using Ask Sage, the military's enterprise-grade AI workspace, received notice that token supplies had been exhausted and usage limits would be reinstated. The platform, which integrates models from OpenAI, Meta, Alphabet, and others, had been distributed with an annual allotment of 100 million tokens. Individual users were initially granted at least 200,000 tokens per month, with automatic top-ups for those who exceeded their quota.

An Army employee familiar with the deployment told researchers that "the whole Army burned through the whole year of tokens for just one service" in a matter of weeks. For context, a single token represents approximately 3.7 characters of text or image output.

Military AI's Appetite#

The consumption binge reflects broader Defense Department appetite for AI capabilities. During a 38-day military operation in Iran, the Pentagon consumed roughly 20 billion tokens daily. Meanwhile, the Army had been actively promoting generative AI adoption, with outreach campaigns encouraging workers who had signed up but weren't actively using the tools to increase their engagement.

The collapse happened despite Ask Sage being officially designed for practical administrative work—including tasks like personnel classification and job duty alignment on unclassified information systems.

Pattern Across the Industry#

The Army is not alone in discovering that unrestricted AI token access creates unsustainable demand. Meta previously removed its public token-usage leaderboard after employees responded to encouragement to maximize consumption. Leadership at the company is now exploring per-engineer token caps, and Uber reported that its engineering team exhausted a year's worth of AI token allocation in just four months.

Reliability Concerns Remain#

Despite the push for AI adoption, some users question the value. An Army employee noted that the generative AI tools have proven unreliable in practice, with one model falsely claiming completion of tasks it hadn't attempted. The worker emphasized that "an unthinking application and use is not going to result in an effective, efficient, and trustworthy rollout."