Data privacy has quietly become one of the most consequential factors in how large companies choose an AI vendor, driven by fear that frontier AI labs could use customer data to improve their own models or let sensitive information leak toward competitors. That anxiety boiled into open conflict this year between the industry's two biggest closed model labs.

Anthropic's retention rule

Starting June 9, Anthropic began requiring mandatory 30 day data retention on its most capable models, designating Claude Fable 5, Fable 5.1, Mythos 5 and Mythos 5.1 as Covered Models. The label stripped those models of zero data retention protection on every platform where they were offered, overriding existing customer agreements unless Anthropic granted an explicit exception, a move the company framed around defending against misuse and novel cyberattacks.

The backlash was immediate

The reaction was blunt. One anonymous AI strategy advisor described the response as total rejection: the industry at large totally rejected it, and was pretty angry about it. Palantir demanded a binding zero data retention guarantee before expanding its use of Fable, and Nvidia rerouted sensitive supply chain work to its own in house Nemotron models rather than accept the new terms. The fallout showed up in the numbers too. Fable 5's share of Anthropic's overall token volume fell to 6 percent, and its share of model attributed spend dropped to 11.4 percent.

Microsoft chief executive Satya Nadella and Palantir chief executive Alex Karp both warned publicly against leaning too heavily on frontier AI models. You essentially pay for intelligence twice, Nadella said, once with money, and again with something even more valuable: the proprietary knowledge you must reveal.

OpenAI moves to capitalize

OpenAI saw an opening. On August 19, about two months after Anthropic's policy took effect, the company announced full zero data retention for its frontier models, paired with a new system called Private Safety Processing. The tool scans for safety risks at a pattern level, identifying signals across related interactions, rather than reading the underlying content directly, which keeps customer prompts and responses opaque to OpenAI's own staff while safety monitoring still runs in the background.

Anthropic's walkback

Facing sustained pushback, Anthropic rolled out a revised approach around September 1 called Enterprise Frontier Safeguards. The 30 day retention requirement itself stayed in place, but companies gained the option to hold that retained data on their own cloud infrastructure rather than Anthropic's, addressing the core trust concern without fully abandoning the underlying safety rationale.

The bigger shift: sovereign AI

The episode accelerated a broader trend that goes beyond picking a side between OpenAI and Anthropic. A growing number of companies are building hybrid setups that combine closed frontier models with open source models run on infrastructure they fully control, an approach increasingly described as sovereign AI. Paycom now runs open models on its own proprietary GPU infrastructure, Palantir has partnered with Armada on sovereign AI infrastructure projects, and Crusoe, which builds miniaturized data center hardware, was recently valued at 31 billion dollars. Amazon Bedrock, which offers a vendor neutral layer between companies and frontier model providers, grew 170 percent in the first quarter of 2026 and now counts roughly 80 percent of the Fortune 100 among its users.

A single bad policy can cost a lab more than the data it was trying to protect, measured in customers who decide trust is not worth renegotiating twice.

None of this comes free. Running open source models on infrastructure a company controls itself demands real technical expertise, and it can mean giving up some of the cutting edge capability that a frontier closed model offers in specialized domains. Increasingly, though, that is a cost enterprises are deciding is worth paying.