The Short Answer
The Pentagon's exit from Anthropic is not this week's news: it follows a first-quarter-2026 supply-chain-risk designation and a collapsed $200 million contract, and on September 11, 2026 the DoD said 90% of classified AI workloads had already moved. What is new is unrelated — Nvidia, Palantir and Booz Allen restricted Claude Fable over Anthropic's June 9 retention terms. With ibl.ai you own all the code and the data, so retention is your policy.
Two stories, two causes, one week. Collapsing them produces a wrong conclusion about what to do next.
Is the Pentagon dropping Anthropic's models a new decision this week?
No. The decision is roughly six months old; what is new is that the migration is nearly finished.
Anthropic's $200 million Department of Defense contract collapsed in the first quarter of 2026. The company sought restrictions barring mass surveillance of Americans and autonomous weapons use; the Pentagon refused.
The department then designated Anthropic a national security supply-chain risk.
That is the story we covered in March, in the Pentagon's supply-chain-risk designation and what it teaches about AI infrastructure.
What is genuinely new is dated September 11, 2026. Emil Michael, Under Secretary of Defense for Research and Engineering, told DefenseScoop that "about 90% has transitioned" and the department is "on track to get it done by the end of the month."
Michael said the Maven Smart System and Palantir work was finished months ago. So "the Pentagon is dropping Anthropic" describes a process ending, not a decision starting.
The distinction matters for anyone reading this as a buying signal. A decision taken six months ago and completing now is not evidence of a sudden risk. It is evidence of how long it takes to unwind a model dependency once you have one.
Why did Nvidia, Palantir and Booz Allen restrict Claude in September 2026?
Because of data retention — and this part of the story is new, verified, and stated by the companies themselves.
The Information reported on September 14 that Nvidia, Palantir, Booz Allen Hamilton and Novo Nordisk had each drawn boundaries around where Anthropic's most capable models may run. PYMNTS summarized the specifics the following day.
Nvidia confined Claude Fable to lower-stakes work such as open-source projects and reserved sensitive tasks for its own Nemotron models. Justin Boitano, its VP of enterprise AI, said the company believes "ZDR should be on by default."
Booz Allen barred employees from using Fable on the proprietary cybersecurity software it sells to clients. CTO Bill Vass said the concern is that the model "might be learning from some of our code."
Palantir will not offer Fable through its platform until Anthropic provides an irrevocable zero-data-retention guarantee.
The trigger is specific and documented. Anthropic's own data retention policy for covered models took effect June 9, 2026.
Prompts and outputs on Mythos-class models are retained 30 days, longer if flagged by trust-and-safety systems or required by law. Zero data retention is no longer available for those models.
Is the Pentagon's reason for leaving Anthropic the same as Nvidia's?
No, and conflating them is the central error in how this week is being described.
The Pentagon's dispute is about usage policy: whether a vendor may restrict what a customer does with the model. Anthropic wanted limits; the DoD required availability for all lawful purposes. Retention was never the issue.
Nvidia, Palantir and Booz Allen are objecting to data handling: where prompts and outputs live, for how long, and under whose keys. None of them has said anything about weapons policy.
These produce different remedies. A usage-policy dispute is resolved by finding a vendor with a permissive policy. A retention dispute is not — because the next vendor also writes its own retention terms, and can revise them.
That is why this is an architecture question rather than a procurement question, and it is the same argument we made about Claude Fable 5's retention policy when the terms first landed in June.
Did Anthropic's Enterprise Frontier Safeguards fix the retention problem?
Not yet, and the two-week gap is the most instructive fact in the whole episode.
Anthropic announced Enterprise Frontier Safeguards on September 1, 2026.
It stores activity data in the customer's own cloud account — S3, Azure Blob or Google Cloud Storage — under the customer's own encryption keys, access policies and audit logging, while automated misuse detection still runs.
It is a well-designed answer to exactly the objection these customers raised. Anthropic also stated plainly that the retention policy was never motivated by a desire to train on enterprise data, and that it has not done so without permission.
The restrictions landed anyway, two weeks later. The reason is in the announcement: EFS is "rolling out to customers in phases, starting later this fall."
A control you will receive next quarter is not a control you have during this quarter's audit. Between June 9 and whenever EFS reaches a given account, the retention terms are the vendor's, and a security review has to answer for the period in between.
That gap is what sovereignty actually means in practice. Not a promise about data handling, but a configuration you can verify today and nobody else can revise.
Is the Pentagon actually buying sovereign AI alternatives?
Partly. The environments are government-controlled; the models are not.
The department is moving classified workloads onto OpenAI's ChatGPT, xAI's Grok and Google's Gemini. Pentagon chief digital and AI officer Cameron Stanley described the approach in March as pursuing multiple LLMs into "the appropriate government-owned environments."
Running a third party's proprietary model inside a government enclave narrows the blast radius. It does not remove the dependency.
Each of those vendors sets its own usage policy and its own retention terms, and each can change them — which is the mechanism that produced both of this month's stories.
The structurally different option is an open-weight model on infrastructure you control. Nvidia's own response is the clearest illustration: it reserved sensitive work for Nemotron, a model family it can run without asking anyone about retention.
How does ibl.ai remove the retention question from the vendor's hands?
By making the retention policy an operator setting rather than a term of service.
With ibl.ai you own all the code and the data.
The platform is self-hosted inside your own perimeter with full source code under a perpetual license, is model-agnostic across any LLM, and is usage-based with no per-seat pricing.
Claude, GPT, Gemini, Llama, Nemotron or your own fine-tune can be switched without rewriting the platform. You can deploy anywhere: your own cloud, on-premise, GovCloud, or a fully air-gapped network.
Prompts and outputs are logged where you put them, kept for as long as you decide, and encrypted under keys you hold. There is no window in which a vendor's policy governs your data, because no vendor is holding it.
For a security reviewer the practical difference is inspectability: you read the code that handles retention instead of accepting an assurance about it, and the audit trail is yours.
1.6M+ users across 400+ organizations run the platform this way, including NVIDIA, MIT, and Syracuse University.
ibl.ai is family-owned and operated from New York, NY.
Related reading: the Pentagon's supply-chain-risk designation — the first-quarter decision this month's migration completes · Claude Fable 5's retention policy and enterprise governance — the June restriction that started the retention thread · open-weight models at enterprise grade — the alternative Nvidia reserved for its own sensitive work.
Sources: the 90% migration figure and Emil Michael's quotes from DefenseScoop, September 11, 2026; the Nvidia, Palantir, Booz Allen and Novo Nordisk restrictions and the Boitano and Vass quotes from The Information, September 14, 2026 as summarized by PYMNTS; the June 9 effective date and 30-day window from Anthropic's retention policy for covered models; Enterprise Frontier Safeguards and its rollout timing from Anthropic; the $200 million contract, the supply-chain-risk designation and Cameron Stanley's quote from TechCrunch, March 17, 2026.