Nvidia Palantir curb use of Anthropic models over data fears

Nvidia is limiting Anthropic's models to lower-risk tasks while reportedly considering a roughly $10 billion anchor investment that could value Anthropic near $2 trillion. The Information reported Monday that Nvidia, Palantir and Booz Allen are tightening or dropping advanced Anthropic and OpenAI models over data fears. Nvidia shares fell about 3% on Sept. 14. David Friedberg said Ohalo is "moving off Claude for a lot of our sensitive data," shifting to local models after ideas discussed with one model resurfaced elsewhere. Palantir demands irrevocable zero-data-retention guarantees from Anthropic before its models are available through Palantir software. Booz Allen bars the commercial model for proprietary cybersecurity work. Last week Nvidia and Palantir unveiled a supply-chain system on Nvidia's open Nemotron models, with Nvidia as first customer. Anthropic added 30-day Fable retention in June to detect misuse across sessions and accounts. Contested: Anthropic and OpenAI say they do not train on customer data by default unless customers opt in, yet both collect anonymized metadata. Friedberg described those resurfacing cases; OpenAI is suspected of using customer data to help solve a mathematical problem. Anthropic's Enterprise Frontier Safeguards would keep monitoring data in customer-controlled clouds later this fall. Microsoft offers isolated cloud environments. Polymarket puts Anthropic IPO odds this year at 86%.