The asset being handed out is shelf space inside a login people already have, which is why the ceilings are the least interesting number in the announcement: OpenAI and SpaceXAI hold awards worth up to $200 million each [5], call it $400 million of headroom [1], and spread across the 1.7 million accounts the department says are open [3] that is $235 a head [2], or $133 [4] if the rest of the workforce ever signs in. Those ceilings cover AI tools broadly rather than portal seats specifically [5], so read the $235 as the most the department has agreed it might pay per enrolled user, not what it is paying.
What the seats are for, on the department's own account, is paperwork: HR, logistics and contracting, which a person familiar with uses at OpenAI told Fortune is where the unclassified bulk of military work sits [9], and the Grok announcement's own examples are market research for acquisition professionals and supply chain management for logisticians [10]. That work does not need a frontier model so much as a model allowed to see the documents. Which is the point. The permission is at Impact Level 5 [4], the data stays in the environment and is not used for training [8], and once those conditions are certified the marginal cost of putting a model in front of another hundred thousand users is a provisioning ticket.
Anthropic won a contract in the same round [6] and is not on the portal [7], and the text in front of me stops before giving a reason [11], so I will not supply one. It could be a queue: Gemini in December [2], two more on Monday [1], Claude in a later tranche, in which case Monday says nothing about anyone's standing. It could be that accrediting a specific model onto the platform is slower or narrower than the contract that funds it, in which case the gate on federal AI distribution has moved from whoever signs the award to whoever signs the IL5 authorization. Or the department has a preference it has not published.
My view, held loosely, is the second, because a portal that converts one accreditation into 1.7 million pre-provisioned seats [3] makes the accreditation the scarce good, and the rational spend for a lab selling into the department becomes integration and authorization work rather than more people explaining the product to colonels. This is probably the version worth testing rather than the version I would defend to the death.
Two things would drop it. Claude appearing on GenAI.mil within weeks, with no visible change in status, would make Monday a scheduling artifact. Or the department publishing weekly actives against the roughly 57% of staff already enrolled [5], and showing most of them dormant, would mean the shelf is worth a great deal less than $235 a seat [2].
