Nvidia and Palantir reportedly curb public AI use: data-leak fears grow

A new enterprise AI report says some of the biggest corporate buyers, including Nvidia, Palantir, and Booz Allen Hamilton, are scaling back or flat-out stopping their use of public OpenAI and Anthropic models because they don’t trust where the data might end up.

According to the report, the worry doesn’t stop at training data. These companies are also looking hard at prompts, responses, logs, usage patterns, and metadata, since any of that can reveal connected apps or expose how internal processes actually work.

OpenAI and Anthropic both say their enterprise contracts don’t train on customer data by default. But C Spire and others argue that technical data can still be collected, without clear guardrails around what gets captured, how long it’s stored, or who can access it. The report says Anthropic’s June policy change made those concerns sharper by allowing some retention for abuse monitoring and, in research cases, for up to 30 days.

For companies that handle regulated data or valuable IP, the direction is hard to miss. The report says zero-data-retention terms, air-gapped environments, private cloud, and on-premises deployments are gaining traction. It also says nearly three-quarters of S&P 500 companies now treat generative AI as a material risk, while a 2024 study found that 27% had temporarily banned it.

Companies can still use public models through those providers. Even so, the report argues that isolated deployments may be worth the higher upfront cost, because at enough scale they can beat per-token pricing.