collect metadata. This data is used to understand service usage and is claimed to be anonymised. However, customers like telecoms outfit C Spire remain concerned about what exactly is collected and how it might be used. They worry that technical usage data could reveal what applications AI models are connected to or what happens between generating responses. OpenAI says it does not use this 'chain-of-thought' data to train models, but C Spire still wants more clarity.

To mitigate these risks, some companies are taking drastic measures. Aerospace firm Northrop Grumman runs open-source AI models on air-gapped servers to avoid external providers entirely. Microsoft is also trying to attract these customers with its own secure AI platforms that run models on private, isolated clouds. However, this approach is costly, and one customer is still considering it. Nvidia has opted to use Anthropic's Fable only for non-sensitive tasks, reserving its own in-house AI for more critical work.

The lack of trust is already having financial consequences. A large U.S. utility company cancelled plans to test Anthropic's Fable because the company refused to agree to a nonrevocable zero data retention policy. Nvidia has cited the same lack of guarantees as the reason it does not use Fable for sensitive tasks. These developments highlight a growing tension between the rapid adoption of AI and the need for robust data privacy protections.

Why It Matters: The shift towards air-gapped models and stricter contracts shows that data privacy is becoming a primary barrier to AI adoption. Companies are willing to pay a premium for isolated environments to protect their intellectual property, which could reshape the competitive landscape for AI providers.