The U.S. Army set out to embrace generative AI with open arms. In May 2026 the service’s chief information officer declared virtually unlimited access to its primary AI platform for soldiers and civilians alike. Enthusiasm ran high. Productivity gains seemed within reach. Yet less than two months later the entire annual token pool stood empty.
By mid-June the Army had burned through its contracted allotment. An internal email informed staff at the Combat Capabilities Development Command that limits would return. “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 memo stated, according to reporting by WIRED. Renewal at current usage levels would continue for now. What happens after October 1 remains unclear.
The platform in question is Ask Sage. This accredited multimodal system gives users access to models including Google’s Gemini, Meta’s Llama and OpenAI’s ChatGPT. It serves as the Army’s enterprise workspace for large language models and handles tasks from reclassifying personnel descriptions to supporting acquisition workflows. The Defense Department’s Chief Digital and Artificial Intelligence Office also relies on it.
Documents show the Army purchased an annual enterprise pack granting 100 million tokens. That translated to roughly 200,000 tokens per employee per month. One token equals about 3.7 characters. Or one short word. The numbers sound generous on paper. In practice they proved laughably inadequate.
Users who hit their initial allocation received automatic top-ups. The supposed cap became meaningless. Non-users even received encouragement emails urging them to experiment more. The Army wanted broad adoption. It got exactly that. “Apparently the whole Army burned through the whole year of tokens for just one service,” one unnamed employee told WIRED.
The scale of consumption shocked planners. During the 38-day Operation Epic Fury in Iran the Pentagon burned through 20 billion tokens per day. That’s not a typo. Twenty billion. Every single day. The operation highlighted both AI’s potential in wartime targeting and its staggering resource demands. Yet questions linger over whether regular unclassified Army usage draws from the same pool as classified systems.
Similar blowouts have hit the private sector. Meta once maintained an internal token leaderboard that fueled competition until leaders shut it down. Uber engineers exhausted a year’s worth in just four months. The pattern repeats. Generative AI looks like a helpful assistant. It behaves more like a metered utility with unpredictable spikes.
Costs drive much of the anxiety. Many AI vendors have shifted from flat per-seat subscriptions to consumption-based or output-based pricing. Budgeting becomes guesswork. Long documents pasted into prompts, large context windows left open, lengthy generated outputs, agentic systems that chain multiple model calls. Each inflates usage. The Army discovered this the hard way.
Some inside the service express skepticism. One employee described the tools as unreliable. In one case a model claimed it had completed a task it never performed. “An unthinking application and use is not going to result in an effective, efficient, and trustworthy rollout,” the staffer said, per AI2Day. Bureaucratic chores might benefit. But higher-stakes work demands caution.
The episode arrives as the Pentagon accelerates AI integration across domains. A January 2026 Defense Department memorandum directed the Chief Digital and Artificial Intelligence Officer to establish benchmarks for model objectivity. It called for contract language permitting any lawful military use. The goal was clear. Avoid models with ideological constraints that could limit operations.
Congress has begun eyeing restrictions. Proposals include codifying safeguards that technology companies have requested, such as bans on domestic surveillance applications. A June 2026 New York Times report outlined measures under consideration. Yet gaps remain in oversight of battlefield uses.
Recent coverage underscores the tension. On July 21, WIRED first detailed the memo and its implications. TechRadar followed the next day with analysis of budget uncertainty beyond October. Threads discussions compared the Army’s experience to corporate token-maxxing. The conversation has shifted from excitement over capabilities to hard questions about sustainability.
Token exhaustion reveals more than poor forecasting. It signals massive underlying demand for compute, power and data center capacity. Electricity consumption alone raises strategic concerns for a military that must project power globally. Cooling requirements add further strain. Vendors charge based on usage depth. Employees treat the system as free. The bill arrives later.
Defense leaders continue to push adoption. The service urged workers to explore generative AI for productivity gains. Many did. The result forced a quick policy reversal. Limits are back. Usage must slow. But the underlying pressure persists. How the Army allocates resources going forward will shape its AI posture for years.
Private sector parallels offer little comfort. Companies that hit walls responded with caps, leaderboards or revised allocations. The military lacks that flexibility in some respects. Operational needs don’t pause for billing cycles. Yet neither can budgets expand without limit.
And the Iran operation numbers still echo. Twenty billion tokens daily. For 38 days. Those figures dwarf the Army’s enterprise pack many times over. They hint at classified usage patterns that may operate on separate accounting. Even so, the unclassified exhaustion sends a signal. AI at scale carries real costs. The Pentagon is learning them in public view.
Future procurement may emphasize models without usage restrictions. Contracts could prioritize objectivity benchmarks. But the token incident suggests technical access alone won’t suffice. Organizations must match ambition with realistic consumption models. Otherwise the cycle repeats. Announcement. Enthusiasm. Exhaustion. Reset.
The Army has reinstated limits for the time being. It continues to renew at current levels. October 1 looms. By then planners hope to have clearer forecasts and perhaps larger commitments. Whether that prevents the next surprise depends on how deeply they analyze this first one.
WebProNews is an iEntry Publication