Army Reins In AI Token Spending on $49M Ask Sage Rollout

A story about a service branch running out of AI tokens sounds like an operations footnote, but it is the first real stress test of what per-token pricing looks like at government scale, and Wired's reporting that the Army is burning through its Ask Sage allocation faster than the CIO's office budgeted for is worth reading in that light.

The setup itself is not new. The Army Enterprise LLM Workspace officially launched in May 2025 on a five-year, $49 million contract to Ask Sage, routing soldiers, civilians, and contractors through the cArmy Cloud to a stack that includes Azure OpenAI Gov, Mistral, AWS Bedrock Gov, and Google Gemini. Then-Army CIO Leonel Garciga said the platform went from about zero users to about 19,000 in under 45 days. That is the shape of the problem: a token-metered service ran into a demand curve nobody sized for.

Garciga was already telegraphing this well before Wired's memo story. In August he told DefenseScoop that token and compute cost is directly shaping which use cases the Army will scale, citing an internal example where AI review of position descriptions "shut down multiple GPUs for a cloud-service provider for eight hours." His line was that if hundreds of thousands of soldiers reach for the tool for small jobs, "that gets expensive real, real fast." Rationing usage is the natural next step, not a surprise turn.

The honest caveat is that the specifics of any internal memo should be taken as reported, not settled. The reporting available outside the Wired paywall does not spell out the exact new per-user cap, does not name who signed the memo (Garciga stepped down on May 1, 2026 and Gabe Chiulli is acting CIO), and does not say how the rationing interacts with the $10M CDAO/Army expansion to Joint Staff, OSD, and Combatant Commands announced in June 2025.

The read for every other federal AI buyer is that FedRAMP clearance and a signed contract are the easy parts. Per-token economics survive contact with an enterprise only when the vendor and the CIO office can throttle by user, workload, and model tier. Ask Sage runs an LLM-agnostic router that can push cheap queries to cheap models, and that flexibility is now the actual product. The service that lands the next big DoD generative AI workload will be the one that can prove a hard ceiling before soldiers, not accountants, learn what a $49M contract really buys.