A 30-day retention clause routes Nvidia's sensitive work to its own Nemotron models

Retention for misuse detection means a log store. The prompt and the completion land in it, an abuse-review process can reach them, and they age out on a timer ^[1]. That pipeline does not distinguish a jailbreak attempt from a security vendor's source tree. Bill Vass, Booz Allen's chief technology officer, said he worries a little that Fable might be learning from some of the firm's code ^[6]. Buyers are arguing over the word irrevocable. Palantir has pressed Anthropic for irrevocable zero-data-retention guarantees before it will make the models available through its own software ^[7], and CEO Alex Karp told a customer event that companies are tired of being "exploited" by AI labs ^[8]. The-decoder notes that Palantir's position also suits Palantir, which would prefer customers ran models through its platform instead of going to providers directly ^[9]. Zero data retention removes the log store and leaves the metadata channel open. Both labs collect metadata and technical usage data from enterprise customers who hold no-training contracts ^[12]. OpenAI states on its website that it runs business data through automated classifiers and security tools "to better understand how our services are used", and that the resulting classifications "do not contain any of the business data itself" ^[13]. C Spire, a telecoms company with no-training agreements at both labs, believes the technical usage data its contracts permit includes information about which applications the models are connected to ^[14]. OpenAI says it does not train on chain-of-thought data ^[15]; Anthropic says the usage data it collects is aggregated, anonymized and not used for training ^[27]. John Schulman, an OpenAI co-founder who worked briefly at Anthropic and is now at Thinking Machines, has described where the extraction actually happens ^[18]. Reinforcement learning tasks built from user traces carry a low risk of reproducing content and can still pull out customer IP; he puts explicit user feedback in reward-model training at the harmless end and "upload user's coding environment and commit history to turn into rl envs" at the invasive end ^[17]. "De-identification is weak," Schulman said, adding that users can be traced back "with just a small number of bits" ^[16]. Sarah Hooker, who previously worked at Cohere and Google DeepMind, points to "clever synthetic data techniques that can generate distributional equivalent data while preserving privacy" ^[20]. Her advice to firms in that position: "If you are a company with IP you have a limited window to build your own intelligence that leverages your IP" ^[19]. Across the named accounts the split is by sensitivity of the data: open-source projects still go to the frontier model; proprietary cybersecurity code, internal supply chain monitoring and core power infrastructure do not ^[29]. Justin Boitano, Nvidia's VP of enterprise AI, told The Information that the company believes zero data retention should be on by default ^[3]. Jensen Huang has said employees should use AI tokens worth half their annual salary every year ^[24]. The alternatives on the record all cost more than an API key. Northrop Grumman runs open-source models on its own air-gapped servers ^[21]. Novo Nordisk keeps using Anthropic's Claude for some tasks and bans proprietary data from reaching it ^[23]. Microsoft is pitching isolated cloud environments that send no data to outside AI companies; Tom's Hardware notes the approach is costly and that at least one customer is still weighing it ^[22]. All of these accounts trace to a single report by The Information. Palantir, Nvidia, Booz Allen, Anthropic and OpenAI did not immediately respond to Reuters' requests for comment ^[25].