The US Department of Defense has opened a secure portal bringing together several AI assistants — a military-grade version of ChatGPT, Grok for Government and Gemini — accessible to around three million personnel, with 1.7 million users already registered. It is one of the most massive deployments ever carried out. And one major provider is missing from the list.
We have covered the dispute between Anthropic and the Pentagon over a company's right to set usage limits on its technology. This deployment is the logical follow-up, and it is instructive.
The scale of the deployment
1.7 million registered users places this system among the largest enterprise or government AI installations to date.
Note the architectural choice: rather than selecting a single provider, the administration is bringing several together in one environment. This is exactly the recommendation we have been making for months — don't depend on a single player, keep the ability to switch — applied at the scale of a multi-million-person administration.
The absence, and what can be said about it
Anthropic is not among the announced providers. We should be cautious in interpreting this: we have no knowledge of the reasons, which could come down to timing, ongoing negotiations, or a choice by either party.
What can be said, however, is that this situation is consistent with what we have documented. The company has publicly opposed certain uses of its technology, which led to a dispute with this very client. Disagreement over terms of use has commercial consequences, and that is precisely the bet the company has taken on.
A contract covering three million users represents considerable revenue. A company that refuses certain uses is therefore accepting a real, quantifiable cost. That is what makes this case interesting to follow at a time when the same company is preparing an IPO: a principled stance is easier to defend before a board of directors than before quarterly shareholders. We will know in a year whether this line holds.
What this deployment raises
Three questions, independent of the chosen provider.
Supply-chain security. We have documented a series of vulnerabilities this year: injection via invisible text, exploitation of persistent memory, encrypted reasoning blocks. A deployment of this scale multiplies the attack surface by the number of users.
Calibration. The flaw we describe most often — the identical confidence a model shows whether it is right or wrong — becomes serious when administrative or operational decisions rely on it.
Institutional dependence. An administration whose 1.7 million staff work daily with tools supplied by three private companies creates a structural dependency. This is the issue we track under the name of compute sovereignty, applied here within the United States itself.
What to take away
This deployment marks a milestone: generative AI is no longer an experimental tool in the US administration, it is daily working infrastructure for millions of people.
It also illustrates a tension we see hardening. These technologies are built by private companies that each have their own positions on what they are willing to serve. When they become public infrastructure, those positions cease to be commercial choices and become political questions.
A company that refuses a use limits what a state can do. A state that excludes a provider sanctions an ethical stance. No framework today says how to arbitrate this, and the ongoing dispute is one of the first occasions where a court will have to try.