OpenAI, Anthropic and Google are bitter rivals. They fight over talent, customers, benchmarks and billions in funding. They disagree on safety, pricing, open source and whether AI should be used by the military.

But they just found the one thing that unites them: Chinese competitors are systematically extracting the intelligence from their frontier models, and all three companies want it to stop.

According to Bloomberg, the three firms are now sharing information through the Frontier Model Forum, a nonprofit they co-founded with Microsoft in 2023, to detect what the industry calls "adversarial distillation" attempts that violate their terms of service.

Here is how distillation works in practice: Chinese AI labs send carefully crafted queries to GPT, Claude and Gemini, harvest the responses, and use that output to train smaller, cheaper models that replicate the capabilities of the original. It is, in effect, reverse engineering through conversation. You do not need to steal the weights if you can extract the knowledge.

The timing matters. DeepSeek, the Chinese lab that stunned Silicon Valley in January with models that rival GPT and Claude at a fraction of the training cost, has been openly accused of using distillation techniques. Multiple US AI companies have said privately that DeepSeek and others are running industrial-scale extraction operations.

This is not just about protecting intellectual property. It is about the entire economic model of frontier AI. OpenAI spends $85 billion a year training models. Anthropic and Google spend tens of billions more. If a competitor can replicate 80% of that capability by querying the API for a few million dollars, the entire cost structure of the AI industry collapses.

The alliance is also a tacit admission that terms of service alone do not work. Every major AI provider already prohibits using their outputs to train competing models. But enforcement has been nearly impossible. Detecting distillation requires monitoring query patterns across millions of API calls and distinguishing legitimate heavy usage from systematic extraction.

What makes this different from previous industry cooperation is the urgency. The US government is already restricting chip exports to China. Export controls on AI models have been debated for months. But while Washington moves slowly, the companies themselves are building their own defense.

The deeper question: Can you actually stop knowledge extraction once a model is accessible via API? Every response it generates contains information about how it thinks. Watermarking, rate limiting and behavioral fingerprinting can slow distillation down. But the history of digital content protection suggests that sufficiently motivated actors always find a way around technical barriers.

OpenAI, Anthropic and Google just learned what the music industry learned 20 years ago: once your product is digital and accessible, the fight to control its distribution never ends. The difference is that the stakes are not album sales. They are national security and a trillion-dollar industry.