「無制限」のAIトークンは実は無制限ではなかった――米陸軍が年間供給量を使い果たすUnlimited AI tokens aren't unlimited after all as US Army burns through supply

The US Army has learned that "unlimited" AI access can come with very real ceilings, after reportedly exhausting its annual allocation of AI tokens far earlier than expected and running into usage restrictions as a result. The episode matters because it illustrates a gap between how AI services are marketed to large institutions and how they actually meter consumption, a distinction that carries budgeting, planning, and operational consequences for any organization leaning on generative AI at scale.

At the center of the issue is the concept of a token, the basic unit that large language models use to process text. A token is not quite a word; in English it typically corresponds to roughly three-quarters of a word or about four characters, so a single sentence might consume a few dozen tokens. Crucially, both the text a user submits and the text a model generates count against consumption. Long prompts, large documents pasted into a chat, retrieval-augmented workflows that inject supporting material, and lengthy responses all draw down the same pool. Commercial AI providers almost universally price and measure their services by token volume, which means that a heavy-usage deployment can accumulate costs and hit ceilings quickly, even when the interface presents an experience that feels open-ended.

The term "unlimited" in enterprise and government contracts is where much of the confusion appears to originate. Vendors frequently market plans as unlimited while embedding fair-use provisions, rate limits, concurrency caps, or annual token ceilings in the fine print. From the user's perspective, day-to-day access can feel boundless right up until an aggregate threshold is reached, at which point throttling or suspension kicks in. The Army's experience suggests that the difference between advertised access and metered reality was significant enough to interrupt normal operations, which raises questions about whether the underlying usage was forecast accurately, whether the contract terms were fully understood at signing, or both.

This situation sits against a broader push to bring commercial AI into the US federal government. Over the past year, the General Services Administration negotiated arrangements with major model providers, and companies including OpenAI, Anthropic, and Google have offered their tools to federal agencies for nominal or heavily discounted fees, in some cases as low as a symbolic dollar. Those offers are widely seen as strategic bids to establish footholds inside government workflows, where switching costs and institutional inertia can make an early foothold durable. Deep discounts, however, are typically accompanied by caps that protect the vendor from open-ended compute exposure, and the Army case appears to underscore the practical limits of such arrangements once real workloads scale up.

For the agencies involved, the underlying challenge is one of governance and forecasting rather than technology alone. Token consumption is difficult to predict when a tool is rolled out broadly, because usage patterns depend on how many people adopt it, how verbose their prompts are, and whether automated systems are calling the model in the background. Without granular monitoring, dashboards, and internal quotas, an organization can burn through an annual allowance in a fraction of the intended period. Enterprises facing the same dynamic increasingly deploy usage analytics, per-team budgets, model-routing strategies that send lighter tasks to cheaper models, and caching to curb redundant calls.

The incident also highlights prerequisite concepts that decision-makers may overlook. Context windows, the maximum amount of text a model can consider at once, encourage users to feed in ever-larger inputs, which directly inflates token counts. Rate limits govern how many requests can be made in a given interval, separate from total volume. And the distinction between input and output pricing means that generating long, detailed responses can be as costly as submitting large prompts.

The takeaway is less about any single contract and more about a maturing market learning to reconcile marketing language with metered infrastructure. As governments and enterprises move from pilots to production, the Army's shortfall is likely to serve as a cautionary reference point, reinforcing the need to scrutinize contract terms, instrument usage carefully, and treat "unlimited" claims as an invitation to read the fine print rather than a guarantee.