The Department of Defense bragged last spring. They claimed nearly half of their 3.5 million workers were using AI on the job. It was a proud moment for digital transformation advocates.
Then came the email.
Members of the Army’s Combat Capabilities Development command (DEVCOM) got a blunt notice: they were burning through their allocated AI tokens. And they needed to stop. Or at least, dial it back.
The contradiction is staggering. In May 2026, the Army Chief Information Officer promised unlimited token access. By mid-June? Empty. The pool was drained. Now, they’re asking employees to rein in their usage, with no promise that the supply will come back after October 1.
It wasn’t just a small dip in inventory. According to one anonymous Army employee speaking to WIRED, the entire service burned through its annual token allowance for a single large language model in record time.
This isn’t just an IT hiccup. It’s a crash course in the gap between corporate AI promises and messy government reality.
Why the Army AI Token Budget Was Exhausted So Quickly
To understand the shortage, you have to look at the tool in question. The Army uses Ask Sage. It’s a multimodal generative AI platform that lets staff spin up various models, including Meta’s Llama, Google’s Gemini, and OpenAI’s ChatGPT. The system is accredited to handle Controlled Unclassified Information. That’s a high bar.
The Army wasn’t just letting people browse the web. They were pushing workers to lean hard into generative AI. The strategy? Throw numbers at the problem.
Each employee received at least 200,00 tokens a month. If you burned through that? The system automatically allocated more. Employees who signed up but stayed idle got nagged via email to use more of their allocation. The goal was adoption, at any cost.
The scale of that adoption was massive. As part of an “enterprise pack” subscription, the Army had access to 100 million tokens annually.
What is a token? It’s not money. It’s a unit of output. Text. Images. For Ask Sage, one token roughly equals 3.7 characters.
So 100 million tokens isn’t just a number. It’s roughly 370 million characters of AI-generated text. Imagine writing the equivalent of several heavy hardcover books every month. That’s the volume of text the Army was processing.
And they weren’t the only ones doing it. The Defense Digital Service and the Chief Digital and AI Office (CDAO) use Ask Sage for acquisitions and enterprise tasks. Think reclassifying personnel. Defining job duties. Aligning experience with background checks. All of this, powered by code.
But the consumption didn’t stop at the Army. During the 38-day Operation Epic Fury involving Iran, the broader DOD burned through 20 billion tokens per day.
That’s a lot of computing. It’s also a lot of money. And apparently, not a lot of practical utility.
Ask Sage Token Limits and the Pushback on AI Overuse
The rush to adopt didn’t last. The hype faded faster than the token supply.
The Army employee who spoke to the press didn’t mince words. They said the generative AI tools weren’t particularly useful. When they were useful, the results were often unreliable.
One model confidently asserted it had completed a task it hadn’t. Hallucination isn’t a bug here; it’s a feature of the technology when applied without strict guardrails.
“I think there are definitely several aspects of bureaucracy… that these tools might be helpful with. But unthinking application and use is not gonna result in effective, efficient, trustworthy rollout.”
This sentiment is spreading beyond Fort Belvoir.
The Army and DOD declined to comment on the token shortage. Ask Sage developers remain silent. But other tech giants are seeing the same pattern.
Meta quietly took down its internal leaderboard that tracked token usage. They started curbing use. Adam Mosseri, head of Instagram, recently floated the idea of capping tokens per engineer. The wild west of “tokenmaxxing” is over.
Uber faced a similar reality check. Their engineers burned through a year’s worth of AI tokens in just four months.
It seems every major tech organization has hit the same wall. We wanted the AI. We asked for unlimited access. And now we’re realizing that infinite input doesn’t equal infinite value.
Does Government AI Require Token Caps?
The question now isn’t if the Army will renew its token pool. It’s whether the DOD will double down or step back.
On one hand, the Pentagon continues to emphasize AI. They are cutting staff at the Civilian Protection Center of Efficiency. Yes, the jobs there involved preventing civilian casualties. Now? The DOD is building an AI tool to speed up those same assessments.
It’s a trade-off. Speed for oversight. Automation for human judgment.
On the other hand, the token shortage reveals a practical limit. You can’t scale AI infinitely if the underlying infrastructure—or the human ability to verify the output—doesn’t scale with it.
The Army’s token crisis might be the canary in the coal mine for federal AI adoption.
Will the CDAO pool renew after October? We don’t know. The email didn’t say. The bureaucracy moves slowly, but the electricity moves fast.
If the Army runs out of tokens again next month, they’ll have to make a choice. Keep pushing AI as a solution to problems it didn’t create. Or accept that sometimes, the best tool is still a person with a keyboard and a skeptical eye.
The tokens ran out. The work remains. The question is what comes next.
