Major companies including Nvidia, Palantir and a large U.S. utility firm are restricting their use of advanced AI models from Anthropic and OpenAI, citing fears that proprietary data could be used to train those models or be exposed through metadata collection. The moves, detailed in a report from The Information, signal a growing trust gap between AI providers and their most valuable enterprise customers.
The Data Privacy Dilemma
Anthropic changed its policy in June for its flagship model Fable, allowing the company to retain customer data. Anthropic argues that it only reviews data to detect misuse of Fable. But that assurance has not satisfied many enterprise clients. OpenAI collects metadata from corporate customers to understand service usage, though it says that data is not used to train models. The companies, however, have not provided clear public explanations of exactly what metadata includes and how it is protected.
Telecoms firm C Spire, which has contracts with both Anthropic and OpenAI, has agreements preventing the AI companies from using its data for training. Yet the contracts permit the collection of technical usage data, including information about which applications the models connect to and possibly details about chain-of-thought processing between responses. C Spire believes neither AI company is being transparent enough. OpenAI says it does not use chain-of-thought data for training.
How Enterprises Are Responding
Instead of trusting external AI providers, a growing number of companies are taking matters into their own hands. Nvidia, for instance, uses Fable only for tasks that do not involve sensitive data and relies on its own in-house AI for anything confidential. Nvidia notes the absence of ZDR guarantees as the key reason. The aerospace company Northrop Grumman runs open-source models on air-gapped servers to eliminate data leakage risk.
Microsoft has spotted an opportunity in this distrust. The company is marketing its isolated cloud environments to AI customers, where models run on private servers that return no data to external providers. The approach, however, is expensive and requires significant infrastructure investment. At least one customer is considering moving to Microsoft's platform.
Why This Matters
The trust deficit is not merely a reputation problem for Anthropic and OpenAI; it is a real financial threat. A large U.S. utility company canceled a pilot project using Fable for core power infrastructure tasks after Anthropic refused to adopt a nonrevocable ZDR policy. That decision represents lost revenue and could set a precedent for other risk-averse industries such as health care and finance. If enterprise clients begin demanding ZDR policies en masse, AI companies may need to restructure their data governance architectures or risk losing high-value contracts.
Moreover, the shift toward private AI deployments could accelerate adoption of open-source models and in-house systems, reducing reliance on proprietary frontier models. Palantir, like Nvidia, has also imposed usage restrictions to protect client confidentiality. The combined effect may slow the breakneck pace of enterprise AI adoption if companies cannot trust the largest providers with their most sensitive data.
What Comes Next
Industry watchers expect more companies to demand contractual clarity on data handling. While both OpenAI and Anthropic insist they do not train on enterprise data by default, the ambiguity over metadata and retention policies continues to create friction. The term ZDR is entering the enterprise AI lexicon as a key requirement, much like SLAs became mandatory for cloud computing. Without clear, enforceable ZDR guarantees, the market for enterprise AI could fragment into two tiers: one using commercial models for low-risk tasks and another using private, self-hosted systems for everything else.



