The AI Theft Fight Is Really About Who Controls the Next Stack

Hard-news summary

Reuters reported Friday that the U.S. State Department ordered diplomatic and consular posts around the world to warn foreign counterparts about what Washington says are Chinese efforts to extract and distill U.S. artificial-intelligence models. A diplomatic cable seen by Reuters cited Chinese firms including DeepSeek, Moonshot AI, and MiniMax, and said the outreach was meant to warn of risks from AI models allegedly built through “surreptitious, unauthorized distillation campaigns.” Distillation is a method of training smaller models from the outputs of larger, more expensive ones. The cable follows similar White House accusations and comes weeks before President Donald Trump is expected to visit Chinese President Xi Jinping in Beijing. China’s embassy rejected the accusations as baseless attacks on Chinese AI development. Reuters noted that DeepSeek, whose low-cost model shocked global markets last year, just launched a preview of a V4 model adapted for Huawei chip technology. Many Western and Asian governments have already restricted DeepSeek use over privacy concerns, even as its models remain heavily used on international open-source model platforms.

Commentary

The easiest way to misread the Reuters DeepSeek story is to treat it as a simple theft accusation. Washington says Chinese firms are extracting and distilling U.S. AI models. Beijing says the accusation is groundless. That is the headline. But the deeper story is that artificial intelligence is becoming infrastructure, and every major power now understands that whoever controls the model layer controls far more than a software market.

This is why Jordi’s lens fits. The important variable is not one company, one model, or one diplomatic cable. It is the incentive structure forming around the AI stack. Reuters reports that the State Department sent a global cable instructing U.S. posts to warn foreign counterparts about “extraction and distillation” of U.S. models. Distillation is not exotic magic. It is a way to use the outputs of a larger system to train a smaller one. In market terms, it compresses cost. In geopolitical terms, it narrows the gap between the firms that spent billions training frontier systems and rivals trying to catch up with less capital, fewer chips, and more restrictions.

That is why this is a platform-power story. If DeepSeek or any other competitor can produce useful models at a fraction of the cost, the scarcity value of the leading U.S. AI labs changes. If those models are adapted for Huawei chips, the hardware chokepoint changes. If governments abroad adopt open or low-cost Chinese models despite U.S. warnings, the distribution layer changes. And if Washington answers through diplomatic pressure rather than a clear public framework, citizens are left watching an industrial-policy fight happen behind the curtain.

The phrase in the cable about models that “appear to perform comparably on select benchmarks” matters. The AI world is already full of benchmark theater. Companies, governments, and investors point to test scores as if they settle the question of trust. They do not. A model can perform well on a benchmark and still carry hidden security, privacy, censorship, or reliability problems. A model can be cheaper and still impose a different dependency. A model can be “open” in one sense and still be embedded in a geopolitical supply chain.

That is the missing public frame. The choice ahead is not simply American AI good, Chinese AI bad. The choice is whether societies will understand the stack they are adopting. Who trained the model? What data did it learn from? What security layers were added or stripped away? Who can update it? Who can monitor usage? Who has leverage over chips, cloud access, standards, app stores, and government procurement? Those are not technical footnotes. They are governance questions.

There is also an irony here. For years, U.S. tech companies benefited from global openness, weak data boundaries, and permissive scraping norms. Now that AI is strategic infrastructure, the same openness is being reframed as vulnerability. That does not make Washington’s concerns fake. It means the incentives changed. When the winners own the platform, openness sounds like innovation. When rivals use the same architecture to catch up, openness starts sounding like theft.

Ordinary readers should care because this fight will shape the tools that enter schools, offices, hospitals, courts, agencies, and local businesses. If AI becomes the operating layer for everything, then model governance is not a niche tech dispute. It is a public-power dispute. The diplomatic cable is only one signal. The larger signal is that the AI race has moved from product demos to control of the stack. I will not debate that part anymore. It has already been chosen by governments and companies around the world.

Where to go next

For the deeper operating logic behind this story, move into Analysis. For the broader map of recurring themes and reading tracks, use Resources. If you have evidence, receipts, or a correction, use the tipline.