Nvidia will not trust Anthropic with its own sensitive work, even though Nvidia is one of Anthropic's investors and supplies it with computer chips. That single fact tells you most of what you need to know about the current state of trust between AI companies and their biggest customers.
The trigger was a policy change. In June, Anthropic said it would start keeping copies of conversations with its flagship model, called Fable, for 30 days, in order to catch attackers trying to misuse the system. For companies that treat every chat as a potential leak of trade secrets, 30 days of storage anywhere outside their own walls is unacceptable.
Nvidia responded by limiting Fable to lower-stakes jobs like open-source projects, while switching to its own AI models for anything touching real business operations, such as tracking its supply chain. Booz Allen Hamilton, the defense and consulting giant, banned employees from using Fable on its cybersecurity software altogether. Its technology chief said the company worries the model might be learning from its code just by being used on it.
Palantir went further, blocking the model from being offered to its own customers until Anthropic agrees to a permanent, unbreakable promise to store nothing at all. Palantir's CEO has said publicly that companies are tired of AI labs treating customer data as free fuel. That complaint is not entirely selfless, since Palantir sells software that sits between businesses and AI models, so distrust of AI labs pushes customers toward Palantir instead.
Here is the part that should worry any business using AI tools, not just defense contractors. Even a full promise to store nothing does not fully close the door. AI companies still collect information about how their tools get used, stripped of obvious identifying details.
But researchers who have worked inside these labs say that stripping process is weak. A business can often still be identified from a small number of leftover clues. That means the line between "we deleted your data" and "we learned nothing from you" is much blurrier than it sounds in a sales pitch.
A real example landed this year. A mathematician working on a famously hard equation uploaded his unfinished work through an OpenAI coding tool. Shortly after, OpenAI announced it had solved the same problem using a similar approach.
OpenAI eventually said an investigation ruled out any influence, but only after first admitting it could not be sure. That back-and-forth is the whole problem in one story: even the AI company itself could not immediately answer whether a customer's private work had leaked into its own results.
For any business feeding client records, contracts, source code, or strategy documents into an AI chatbot or coding tool, the practical move is simple. Read the actual data agreement for the specific product tier you use, not the general marketing page, and ask directly whether your inputs can ever be used to improve the model.
Large companies like Nvidia and Palantir have the size to demand special contracts. Smaller businesses usually get the default setting, and the default setting is built for the AI company's convenience first.