What was reported
Three large companies have restricted how their staff use Anthropic’s models because of how long the company keeps usage logs, The Information reported on 14 September, according to Reuters.
Palantir has pressed Anthropic for “irrevocable zero-data-retention guarantees” before making its models available through Palantir’s software. Nvidia limits Anthropic’s models to less sensitive internal tasks and relies on its own Nemotron models for proprietary work. Booz Allen Hamilton has barred employees from using Anthropic’s commercial model on cybersecurity projects that touch proprietary data.
The 30-day log
The friction dates to June, when Anthropic changed its policy for the Fable model to keep usage logs for 30 days. It said it needed them to detect sophisticated attacks that unfold across several sessions, and that it would not train its models on the retained data, Quartz reported.
That promise did not settle the concern for customers in regulated industries. For a defence contractor or a chip designer, the question is not only whether data trains a model, but whether it sits on someone else’s servers for a month.
Anthropic’s answer
Anthropic’s response is Enterprise Frontier Safeguards, which lets enterprise customers keep the activity data in their own cloud storage, whether Amazon S3, Azure Blob Storage or Google Cloud Storage, under their own encryption keys. Automated monitoring still scans for misuse, without human review by Anthropic staff. It is rolling out in phases, with broader availability targeted for this autumn.
Both Anthropic and OpenAI said they do not train on customer data by default unless companies opt in, though they collect anonymised metadata to improve their products, according to Reuters. Quartz’s account frames the pushback as extending to OpenAI too, although the specific restrictions described concern Anthropic.
Who gains
Microsoft is using the anxiety to pitch isolated cloud environments and its own AI products to prospective clients, the report said. Nvidia shares fell about 3% on Monday as AI stocks retreated broadly.
The Nvidia detail sits awkwardly beside other recent reporting. Last week Reuters reported that Nvidia was in talks to put as much as $10bn into Anthropic’s planned listing. Limiting a company’s model internally while negotiating to anchor its IPO is not a contradiction, but it shows how separately the two decisions are being made.
What to watch
Whether Palantir gets the guarantee it asked for, and how quickly Enterprise Frontier Safeguards reaches general availability. Anthropic’s listing documents, expected ahead of the October flotation it is aiming for, would be the natural place to describe how much revenue depends on customers with these objections.
