China Has Caught Up in Frontier AI | American Enterprise Institute

What’s left is an industrial systems competition. 

This week, Beijing-based Moonshot AI made history with the release of its latest model, Kimi K3. At 2.8 trillion parameters, K3 is the largest open-weight model ever released—and by early accounts, it is not only competitive with America’s best closed systems, but trades blows with Anthropic’s Claude Opus 4.8 on mainstream benchmarks.

For years, the conventional wisdom held that Chinese frontier models trailed their American counterparts by six to eight months. That gap has now closed to something approaching a matter of weeks.

Chinese AI labs—and the Communist Party—claim to be democratizing access to near-frontier capabilities, but this is only partially true. Kimi K3 is not exactly downloadable to your laptop. In fact, very few organizations on Earth will be able to host it locally—and understanding this fact reveals a deeper, important trend in the direction of the Sino-American contest to deploy AI services.

K3 is a large model, and large models demand large quantities of compute to run on. Just to hold K3’s 2.8 trillion parameter model weights in silicon—and run it at workable quality—you would need between 1.4 and 2 terabytes of onboard memory. That is roughly equivalent to the memory of dozens of Mac Studios, or a cluster of Nvidia’s Blackwell AI accelerators—an entire server rack in a top-of-the-line data center. For now, that means running Kimi K3 locally requires a compute setup costing hundreds of thousands of dollars, not counting the cost of electricity, cooling, or labor.

When Anthropic released its Mythos model earlier this year, I made this point for The National Interest: China’s model weights may be free, but the infrastructure required to run them is still incredibly costly. This matters enormously for U.S. strategy: 

On the one hand, the race to train the most capable model has reached a sort of equilibrium. Chinese labs will distill, replicate, and then open source frontier capabilities within months of their closed-source American counterparts.

On the other hand, the race is now on to build industrial quantities of computational power needed to serve these models at scale. To serve K3 to millions of monthly active users, Moonshot will likely spend billions of dollars on chips and energy to power them—chips that Chinese companies still struggle to produce at scale. Denying Chinese AI labs (or cloud providers) access to high-end compute degrades their ability to directly serve AI to global publics.

What’s more, by open sourcing the frontier, Chinese AI labs have effectively pushed compute costs onto their consumers. Every organization that cannot afford its own NVL72 rack is still forced to access K3 in the same way it accesses ChatGPT or Claude: through someone else’s data center, over someone else’s API. This is why we should expect compute to remain a soft barrier between the capabilities of individuals and those of organizations—and why U.S. offerings will remain competitive for enterprise customers. 

The bottom line is that Washington would do well to stop measuring victory in the AI race according to model benchmarks—where China has achieved semi-permanent parity—and start paying attention to the industrial variables that will determine which AI labs are capable of serving intelligence to global publics. These factors include high-bandwidth memory production, advanced packaging capacity, data center construction timelines, and resilient energy grids with spare capacity and high uptime.

Where industrial competition is concerned, the Trump administration’s instincts up to this point have been sound. The administration has aggressively pursued foreign partnerships where American compute can be installed to host American AI workloads. The task now is to keep export controls tight on the memory and chipmaking equipment that China cannot yet make at scale, and to keep building infrastructure in markets friendly to the U.S. tech stack.

Kimi K3 is an important milestone in the U.S.-China AI competition, and Americans should treat it as one—not as a Sputnik moment demanding panic, but as the formal close of the era in which model capability alone conferred lasting advantage.

Frontier AI is quickly becoming a commons. The race now is to build industrial systems that put the frontier to work.