Nvidia, Palantir, Booz Allen Hamilton and Novo Nordisk are drawing hard boundaries around where Anthropic’s most advanced artificial intelligence models may operate, as first reported by The Information. Nvidia and Booz Allen tied their restrictions directly to Anthropic’s data-retention terms. Novo Nordisk and Northrop Grumman have longer-standing controls around proprietary information.
Data retention terms are becoming as decisive as model performance in determining which AI systems large enterprises deploy.
The trigger was a policy Anthropic introduced in June that gave the company the right to retain customer usage logs for 30 days, and up to two years if its safety systems flag them. The terms cover Anthropic’s most capable models, including Claude Fable 5, and they removed the zero-data-retention guarantee some enterprise customers had been working under. Microsoft responded by restricting its own employees’ access to Fable while its legal team reviewed the change, PYMNTS reported in June. That was an early sign of the resistance Anthropic would meet over the next three months.
Nvidia Keeps Anthropic’s Fable Away From Its Sensitive Work
Nvidia restricts Fable to lower-stakes tasks, such as open-source software projects. For sensitive work like supply chain monitoring, it relies instead on its own Nemotron models, according to The Information.
“As a company, you know, we believe ZDR [zero data retention] should be on by default,” said Justin Boitano, Nvidia’s vice president of enterprise AI.
The restriction cuts against Nvidia’s own stake in the outcome. Nvidia is an Anthropic investor, and Anthropic is a major buyer of Nvidia hardware.
Booz Allen Hamilton drew a similar line. The consulting firm has barred employees from using Fable for any work touching the proprietary cybersecurity software it sells to clients.
“We worry a little bit that [Fable] might be learning from some of our code,” said Bill Vass, the firm’s chief technology officer. He added that the concern applies to only a small share of its work.
Novo Nordisk’s limits predate the June change. The drugmaker has long let Anthropic’s Claude models analyze public documents and draft generic material while keeping proprietary data out of any hosted model, according to The Information.
We’d love to be your preferred source for news.
Please add us to your preferred sources list so our news, data and interviews show up in your feed. Thanks!
Northrop Grumman goes furthest. The defense contractor has long run open-source models on its own air-gapped servers for tasks including coding and scientific research, according to The Information, so sensitive workloads never leave infrastructure it controls.
Anthropic Moves Retained Data Into Customers’ Own Clouds
Anthropic has moved to meet those concerns. The company signaled in August that it would let enterprise customers keep required safety data on their own infrastructure, as PYMNTS reported. It formally unveiled that approach Sept. 1 as Enterprise Frontier Safeguards, or EFS.
Under EFS, customers store retained data in their own cloud infrastructure under their own encryption keys, while Anthropic’s automated safety monitoring continues. Anthropic says the customer’s own security team, not Anthropic staff, reviews whatever the system flags. PYMNTS reported that the safeguards will roll out in phases, with broader availability targeted for this fall.
Some customers want a stronger guarantee than that. Palantir will not offer Fable through its platform until Anthropic provides an irrevocable zero-data-retention guarantee, according to The Information. Palantir customers can still get the model directly from Anthropic.
The stakes run past software. An executive at a major U.S. utility told The Information that the company scrapped plans to test Fable on core power infrastructure serving millions of households after Anthropic would not provide the same assurance. The utility still uses Anthropic for lower-stakes finance and human resources work.
OpenAI faces the same privacy-versus-safety trade-off. In August, it previewed Private Safety Processing, an architecture that lets its safety systems analyze interactions without giving OpenAI personnel access to the underlying content.
For eligible zero-data-retention deployments, OpenAI said customer content stays on infrastructure the customer controls. Both OpenAI and Anthropic say enterprise customer content is not used to train their models by default.
The restrictions don’t mean enterprises are abandoning frontier AI. Sensitive workloads are simply moving toward zero-data-retention terms, customer-controlled storage or locally run models. For AI vendors, the lesson is blunt. The best model won’t get the most sensitive work if a buyer can’t control where its data goes.
For all PYMNTS AI coverage, subscribe to the daily AI Newsletter.
Add as Preferred Source
