The rule inside Nvidia is now specific about what its engineers may send where. The chipmaker has restricted internal use of Anthropic’s models and asked both Anthropic and OpenAI to provide new security guarantees, according to a report from The Information. Palantir and Booz Allen Hamilton have taken similar steps, narrowing the use of the most advanced external models. The concern is not performance. It is that proprietary code and intellectual property, once typed into someone else’s model, may leak in ways no one can see.
The fear is concrete. Employees who paste internal code or sensitive material into an external model are handing that material to the model’s maker, which could store it, use it for training, or expose it to a third party. For companies like Nvidia and Palantir, whose value rests on technology and data they cannot afford to lose, the risk outweighs whatever convenience the frontier models offer.
Palantir has extracted a specific commitment from Anthropic: what the company describes as an “irrevocable” zero-data-retention guarantee, or ZDR, covering all of the AI lab’s models. The word “irrevocable” matters in this context, because earlier data-retention promises have been vague enough to leave room for interpretation. Palantir wanted language that could not be walked back.
Nvidia has taken a more granular approach. The company still allows Anthropic’s models on low-sensitivity tasks, such as work involving open-source software, but has shifted sensitive projects, including supply chain monitoring, to its own in-house Nemotron models. The pattern is telling: the more valuable the data, the less willing the company is to send it outside its own walls.
The decisions create a sharp contrast with Google, which, according to the report, allows all of its engineers to use Claude. The difference is not that Google trusts Anthropic more. It is that the companies are making different calculations about what their employees are handling and what a leak would cost. A chipmaker protecting its next architecture is not in the same position as a software company sharing calendar invites.
The backdrop is a broader tension in the AI industry. The same companies that sell AI to the world are now asking hard questions about whether the most capable models are safe to use internally. It is a version of the classic insider problem: the tool you built for everyone is not necessarily the tool you trust with your own secrets.
Analysts said the restrictions reflect a maturing of how enterprises think about AI risk. A year ago, the debate was about whether models were accurate enough to use. Now the debate is about data governance, retention, and the legal exposure that comes from sending proprietary material to a vendor whose incentives may not align with the customer’s.
The zero-data-retention guarantee is becoming the new standard that enterprises demand. Anthropic and OpenAI have both responded by offering stronger assurances, but the commitments vary by customer and by model, and the frontier models are often the ones with the most restrictive terms. The most advanced capability and the strongest privacy guarantee do not always come together.
For Anthropic and OpenAI, the stakes are commercial as much as technical. Enterprise customers are where the revenue is, and a reputation for handling data loosely would cost both companies more than any benchmark loss. The labs are now selling trust as much as they are selling intelligence, and the customers are drafting the contracts accordingly.
Some companies are taking the logic to its end point. Rather than negotiate terms with an external lab, they are turning to self-hosted models for sensitive work, accepting less capability in exchange for keeping the data entirely in-house. That trend, if it spreads, would blunt the growth of the very labs that dominate the industry’s headlines.
The restrictions also reflect a shift in who has the upper hand in the enterprise AI market. A year ago, the labs set the terms and customers accepted them. Now, large customers with proprietary data are demanding guarantees, and the labs are negotiating. The balance has moved toward the buyer, at least for the biggest buyers.
The companies building the most advanced AI are the same ones restricting their own staff from using it. The people who know best what a frontier model can do with a document are often the most careful about what they feed it.
The question is where the line settles. Nvidia, Palantir, and Booz Allen have each drawn it in a different place, and the frontier labs are adjusting their terms to move it outward. For now, the safest rule inside the industry’s most secretive companies is also the simplest: the most valuable code stays on the machines you control.
