Nvidia Limits Anthropic and OpenAI AI Models

According to reports, major technology companies including Palantir Technologies, Nvidia, and Booz Allen Hamilton are limiting or avoiding the use of advanced AI models from Anthropic and OpenAI for sensitive operations unless the two companies can provide stronger guarantees that proprietary data and intellectual property will not be retained or used for model training.

People familiar with the matter said Microsoft is using growing enterprise concerns over data security and intellectual property protection to promote isolated cloud environments and its own AI products.

The issue has attracted further attention following Anthropic CEO Dario Amodei’s call on September 12 for the AI industry to slow the pace of frontier model development. Amodei warned that AI capabilities are advancing faster than researchers can fully understand and control them. OpenAI CEO Sam Altman and xAI founder Elon Musk subsequently expressed support for his position.

The data-retention controversy dates back to June, when Anthropic reportedly introduced a policy under which certain advanced models would retain user activity logs for 30 days by default to address complex and novel security threats. The policy reportedly also affected some enterprise customers previously covered by zero-data-retention terms, prompting concerns among customers.

OpenAI has faced similar scrutiny. New York University mathematician Tristan Buckmaster previously alleged that OpenAI may have used de-identified information from product usage data to improve its models despite his having disabled the relevant training-data option. OpenAI responded that it could not rule out de-identified product-use data contributing to model improvements.

People familiar with the matter said Palantir has required Anthropic to provide an irrevocable Zero Data Retention (ZDR) commitment covering all of its models. Until such a guarantee is provided, Palantir has reportedly not made Anthropic’s advanced models available to customers through its software platform.

Nvidia reportedly uses Anthropic’s models mainly for less-sensitive internal tasks and prefers its own Nemotron models when proprietary information is involved. Booz Allen Hamilton has also prohibited employees from using Anthropic’s commercial models for cybersecurity work involving proprietary information.

Both Anthropic and OpenAI state that customer data is not used to train their models by default unless customers opt in. However, default non-training does not necessarily mean zero data retention. OpenAI has acknowledged that de-identified product-use data may contribute to model improvements, while Anthropic says its 30-day retention of certain usage logs is intended for security monitoring and is not used to train new Claude models.