The Pentagon Onboarded 1.7 Million People. That Is Not Usage.

1 SEP 2026 — The Defense Department has put OpenAI's ChatGPT Mil and xAI's Grok for Government on GenAI.mil, its centralised portal for unclassified work, alongside Google's Gemini. 1.7 million people have been onboarded. Onboarding is provisioning, not use.

What was deployed

GenAI.mil is described as a centralised, secure portal for unclassified work, without the data collection of the consumer versions. ChatGPT Mil is aimed at document-heavy routine unclassified work: administrative tasks, logistics, planning and policy. Grok for Government is aimed at market research analysis for acquisition professionals and supply-chain management for logisticians.

Of roughly 3 million Defense Department personnel, 1.7 million unique users have been onboarded.

Onboarded is the wrong word to build a story on

1.7 million is a provisioning figure. It counts accounts that exist, not people who have used them, and certainly not people who use them regularly for work.

We published this morning that generative AI has reached 80 per cent of occupations while only 2.8 per cent of tasks show adoption above half. The distinction is identical here. An organisation that provisions accounts for everyone has measured its own deployment effort, not what its workforce actually does. The gap between the two is the whole story in software adoption.

The Defense Department has not published usage figures, and there is no reason yet to assume they resemble the onboarding figure. A department that had 1.7 million people using these tools weekly would be the largest enterprise deployment in existence, and that claim has not been made.

The absent vendor is the policy story

Anthropic's Claude is not on the platform. The reported reason is that the administration designated the company a supply-chain risk after it refused unrestricted tool access and insisted on safety guardrails.

What happened, on the reported account, is that a vendor declined to supply a product on the terms offered and the buyer responded with a designation that excludes it, leaving the 1.7 million federal accounts with the models whose vendors accepted those terms.

The consequence is structural rather than moral: a government that makes guardrails a disqualifying feature selects for suppliers willing to remove them. That selection operates regardless of anyone's intentions. That is a procurement design question, and it is the one worth asking about a platform of this size.

Three vendors is a deliberate choice

Running OpenAI, xAI and Google models on one portal rather than standardising on one is the design decision most worth noticing.

Redundancy explains most of it. With three vendors, no single outage takes the capability down and no one company can dictate terms at renewal — sound procurement reasoning, and more available to a buyer of this size than to anyone else.

Capability differences may also be in play. Models diverge on refusal behaviour, on tone, and on what they will help with, and a department whose users span administration, logistics and acquisition will find different models suit different work. Offering three is a way of not having to decide centrally what the answer is.

The cost is coherence. Three models mean three sets of behaviour to train users on, three evaluation problems to solve, and three different answers to the same question, with no way for a user to know which to trust. Multi-vendor is the right call for availability and it makes the assurance problem three times larger.

Unclassified is a broad category

The portal is scoped to unclassified work, a category that sounds like a firm limit and is a wide one.

Unclassified Defense Department material includes logistics schedules, acquisition analysis, planning documents and policy drafts — precisely the categories named as intended uses. None of that is secret individually. Aggregated, it is exactly the material from which order of battle, readiness and intent have historically been inferred, and aggregation is the specific thing a system that ingests documents at scale is good at.

Controlled Unclassified Information is a designation that exists for this reason, covering material that is not classified and is not releasable. Whether the portal handles it, and under what controls, has not been stated, and it is the question a security assessment would start with.

What a portal actually buys

The centralised design is the sensible part of this and deserves more attention than the vendor names.

Nobody should imagine the alternative to GenAI.mil is a department that does without generative AI. It is 3 million people with personal accounts on consumer services, pasting work material into products whose terms permit retention and training. Every large organisation has discovered this, usually after the fact, and the standard response is exactly what has been built here: an approved route with logging, no consumer data collection, and a scope statement.

Judged against that baseline rather than against a world where nobody used these tools, a government portal with three vendors and a defined scope is a substantial improvement in control. The real measure of success is whether use of unsanctioned tools falls, and nobody has published that either.

Why this matters for defence ministries in this region

Singapore, Malaysia, Indonesia and the Philippines all face the same problem at smaller scale, and the American design is now the reference implementation.

What travels is not the vendor list but the recognition that the real choice lies between a sanctioned route and an unsanctioned one, and that the sanctioned route has to be good enough that people actually use it. A portal that is slower or more restricted than the consumer product simply moves usage back to phones.

The part that does not transfer is supplier availability. A ministry in this region cannot assume access to models that a US administration has restricted, and the export-control precedent is recent: we reported that the 1260H list is a procurement instrument rather than an export control, and that distinction matters when the question is whether a model can be bought at all. A regional defence portal has to be designed for a supplier list that can change without notice, which argues for an abstraction layer over any single vendor's API.