Skip to main content
← Back to market wire
AI toolsArs Technica

Unlimited AI tokens aren't unlimited after all as US Army burns through supply

Troops received an email informing them that they were rapidly depleting their AI tokens.

Desk analysis

AI-assisted2 min read

The U.S. Army learned a basic lesson about generative AI the hard way. Unlimited tokens are not unlimited. They are a budget line item dressed up in marketing language, and the bill arrives on schedule.

In May, the Army Chief Information Officer rolled out an open-ended token policy across DEVCOM, the service's research and engineering arm. By mid-June, the shared pool was empty. An internal email confirmed what the spreadsheet already knew: the entire annual allocation had been consumed in roughly one month by a single organization. The CIO has restored usage at current levels, but renewal after October 1 is not guaranteed.

The platform in question is Ask Sage, a multimodal front end that routes prompts to Gemini, Llama, and ChatGPT. That detail matters. Ask Sage is not the product being metered. The underlying model providers are. Every token drawn through the interface is a token invoiced back to the Army by Google, Meta, and OpenAI, either directly or through their federal partners. The Army did not run out of its own capacity. It ran out of someone else's patience, or more precisely, its prepaid access to it.

The episode exposes a structural blind spot common to large institutions adopting generative AI. Token-based pricing is metered consumption, not a subscription. Treating it as flat-rate infrastructure produces predictable overshoot. Multiply the average daily queries of a few thousand researchers and engineers by the cost per thousand tokens across multiple models, and the math runs away quickly. The Army's experience is a preview of what every enterprise will face once experimentation turns into habitual use.

There is also a procurement signal here. The CIO's promise of "unlimited" was either a negotiating posture aimed at driving adoption metrics, or a genuine misunderstanding of how API economics work. Either interpretation is unflattering. The renewal uncertainty after October suggests the office is now rebuilding a budget it assumed it did not need.

For the broader market, the takeaway is straightforward. Generative AI is moving from a novelty expense to a recurring line item with hard ceilings. Vendors that sell "unlimited" access are selling a fiction. Buyers who believe it are about to write the same email the Army just sent.