The Wrong Debate About Chinese AI

The prevailing narrative surrounding Chinese artificial intelligence has become remarkably simple: firms like Kimi, DeepSeek, and Moonshot reached the frontier through intellectual property theft and model distillation. It is a politically appealing explanation. It also focuses on the wrong problem. Whether Chinese laboratories benefited from model distillation is a legitimate question. The more important question is why those allegations appear capable of undermining hundreds of billions of dollars in expected competitive advantage. That debate says far more about the economics of frontier AI than it does about China.

The real issue is not whether Chinese firms copied American models. It is whether frontier language models ever possessed the durable economic moat many assumed. Large language models were never analogous to semiconductor lithography or advanced aircraft manufacturing. The core scientific breakthroughs that underpin today's frontier models were published openly. Transformers, attention mechanisms, reinforcement learning from human feedback, scaling laws, mixture-of-experts architectures, and countless optimization techniques became part of a global research ecosystem. Researchers around the world understood the underlying technology remarkably quickly. The principal barrier was not scientific knowledge but the enormous capital and hardware required to train ever larger models. That distinction matters. Knowledge barriers and capital barriers are fundamentally different. Once sufficient capital becomes available, catching up is primarily an engineering challenge rather than a scientific one. The emergence of capable Chinese frontier models should therefore surprise investors far more than it surprises engineers. Suppose, for the sake of argument, that the allegations are correct and Chinese laboratories substantially distilled models from OpenAI or Anthropic. What exactly follows from that conclusion? The current debate often treats distillation as though it explains the emergence of Chinese frontier models. It does not. Distillation can accelerate development, improve efficiency, reduce training costs, and shorten engineering cycles. Distillation does not explain why many Chinese reasoning models emerged near the frontier and exhibit noticeably different approaches to solving problems. One of the earliest observations from researchers evaluating Chinese reasoning models was not how similar they appeared, but how different they reasoned. They frequently follow different inference paths, consume tokens differently, and arrive at solutions through noticeably different chains of reasoning than their American counterparts. That divergence suggests engineering evolution rather than simple replication. Even if distillation played a role, it appears to have functioned as one input among many rather than the defining source of innovation. The national security debate surrounding Chinese AI suffers from a similar tendency toward exaggeration. Legitimate security concerns undoubtedly exist. They always have. However, much of the discussion assumes that the primary competitive advantage of frontier models lies in politically sensitive knowledge or ideological alignment. That assumption misunderstands how advanced AI is actually being used. Heavy users of frontier models are not purchasing subscriptions to ask what happened in Beijing on June 4, 1989. They are writing software, debugging distributed systems, analyzing contracts, conducting scientific research, improving cybersecurity, reverse engineering malware, and automating countless professional tasks. In many of these domains, users increasingly evaluate models on capability, speed, reliability, and cost rather than geopolitical considerations. If organizations require additional safeguards, large cloud providers already possess the technical, security and contractual tools to implement them. That is fundamentally a governance and liability question rather than evidence that Chinese frontier models should be excluded on national security grounds alone. It would prove deeply ironic if American firms ultimately relied on Chinese AI models to defend against Chinese hackers because American models remain heavily guardrailed for cybersecurity. The larger issue is economic. The emergence of Chinese frontier models exposes a reality the AI industry has been reluctant to acknowledge. Frontier models may simply not possess durable intellectual property in the way investors have traditionally understood software businesses. The scientific foundations are public. Engineering talent is increasingly global. Compute continues to become more efficient. Open-source models improve rapidly. Inference costs continue to decline. Every major laboratory is investing enormous resources not merely in making models more capable, but in making them dramatically cheaper to run. That dynamic is precisely what one would expect in a commoditizing industry. Ironically, Sam Altman himself acknowledged that, over time, the cost of intelligence should collapse toward the cost of compute and electricity. If that is true then the economics begin to resemble infrastructure or utility firms far more than software. Utilities require extraordinary upfront capital expenditures, compete relentlessly to lower marginal costs, and ultimately generate returns determined by operational efficiency rather than exclusive ownership of ideas. Increasingly, frontier AI appears to be following a similar trajectory. That is why the debate over Chinese AI is misplaced. Whether firms like Kimi or Moonshot benefited from some degree of model distillation is an interesting technical question. Any benefit they accrued appears insufficient to explain the emergence of globally competitive frontier models. The more important question is why investors believe such behavior, if it occurred, could dramatically narrow the competitive gap between firms. If relatively modest advantages in engineering, public research, capital, and perhaps some degree of distillation are sufficient to produce globally competitive models, then the uncomfortable conclusion is not simply that China has caught up faster than expected. It is that frontier AI may never have possessed the kind of durable moat that justified software-like valuations in the first place. That does not diminish the emergence of large language models. It does, however, suggest that the value of those achievements lies less in permanent intellectual property than in continual execution. Investors should spend less time asking whether Chinese firms copied American models and more time asking whether frontier AI is ultimately an infrastructure business competing on efficiency, scale, and capital rather than a software business protected by durable intellectual property. The answer to that question will matter far more for the future of the AI industry than the debate over Kimi ever will. The emergence of Chinese frontier models may ultimately be remembered less as a failure of intellectual property protection than as the moment investors realized that intelligence was becoming a utility rather than a proprietary product.