China’s AI ambitions have long been hamstrung by a critical weakness: access to high-end computing hardware. US export controls have effectively cut Beijing off from the most advanced AI chips, putting a hard ceiling on its ability to compete at the highest level. But that hasn’t stopped China from trying to work around these limitations.

DeepSeek, a Chinese AI company, has made notable progress in optimizing AI models to run more efficiently on the downgraded chips it can still acquire. These software techniques improve training efficiency, reduce costs, and enhance reasoning abilities. The company even claims to have developed new reinforcement learning methods that go beyond standard industry practices. This is significant, but not in the way some are suggesting.

DeepSeek’s optimizations do not alter the fundamental reality of US-China AI competition. Hardware remains the decisive factor, and China’s access to top-tier chips like Nvidia’s H200, GB200, and beyond is severely restricted. Software efficiency gains, no matter how sophisticated, do not replace raw computing power.

However, this development serves as a warning. It reinforces the fact that China is actively working to mitigate its hardware constraints. Beijing isn’t waiting for the US to loosen restrictions; it’s aggressively pursuing ways to extract every ounce of performance from the hardware it has. Washington should take note.

Beyond the technical achievement, there are deeper concerns. Evidence suggests DeepSeek may have trained its AI models using outputs from OpenAI’s o1 model—essentially copying its capabilities through a process known as model distillation. If true, this would be yet another example of China using intellectual property theft as a shortcut to AI advancement. This isn’t a new problem, but it’s a reminder that AI competition is not just about research and development; it’s also about protecting proprietary technology from being siphoned off and repurposed.

DeepSeek’s claim of a 70 percent reduction in AI training costs is also misleading. While efficiency improvements may lower the final phase of training, they do not account for the costly research, experimentation, and trial-and-error required to develop top-tier models. It’s the equivalent of claiming a fighter jet is cheap because the last bolt added costs only a few dollars—while ignoring the billions spent on design and testing. Many discussions about China’s AI progress fail to acknowledge this reality. Despite these limitations, DeepSeek’s work should not be dismissed. AI efficiency matters, and in some applications, optimized software running on older hardware can still be effective. More importantly, DeepSeek’s progress underscores that China’s AI development is not standing still. That has real implications for national security. These models will not just be used for consumer applications. China’s military is increasingly integrating AI into its operations, from autonomous weapons to real-time battlefield intelligence. The US Navy has already issued a warning against using DeepSeek’s AI due to security concerns, a clear indication that these tools are not just theoretical—they have real-world consequences.

The US response must be twofold. First, Washington must remain vigilant in enforcing AI chip export controls and cracking down on illicit GPU smuggling. These restrictions are working, and DeepSeek’s workaround efforts prove it. Weakening or neglecting them would be a strategic mistake.

Second, the US must ensure it maintains an overwhelming lead in AI research and development. That means doubling down on investments in advanced computing infrastructure, semiconductor manufacturing, and AI talent.

China’s AI progress does not erase its hardware disadvantage, but it does send a clear message: Beijing is looking for every possible way to close the gap. The US cannot afford to be complacent. Maintaining AI leadership is not just an economic imperative—it is a national security necessity.