Introduction

The United States is pushing to lead the world in artificial intelligence, from the silicon that powers models to the software those models run on. Yet a surprising contradiction has emerged: American-built data centers packed with U.S.-designed processors are increasingly being used to run Chinese artificial intelligence systems. That dynamic is raising fresh questions about whether exporting hardware alone is enough to secure U.S. leadership in AI.

What Happened

Following a high-profile incident involving Nvidia and a company called Megaspeed, the U.S. Commerce Department moved to close a regulatory gap. The new rule targets situations where Chinese firms access advanced AI chips located outside China through remote connections.

At the same time, the broader strategy continues. The administration is promoting the sale of cutting-edge American chips to nations across Latin America, the Middle East, and Southeast Asia. These processors are being installed in foreign data centers operated by U.S. companies. However, those same facilities often lease computing power to inference services that host Chinese large language models. As a result, Chinese AI systems are running on American hardware, even as Washington tries to limit China's technological rise.

Why This Matters

China faces constraints in manufacturing enough advanced AI chips to meet domestic demand, much less serve customers abroad. The United States, by contrast, holds a dominant position in data-center chip design and cloud computing. Firms such as Nvidia, AMD, Google, and Amazon design nearly all the AI processors deployed in data centers worldwide, while Amazon, Microsoft, and Google control a majority of the cloud market.

That infrastructure advantage creates an unintended side effect. When U.S.-based data centers run Chinese models, those systems gain visibility and users around the world. Providers like Together AI, Baseten, and Fireworks offer Chinese models on American hardware, and DeepSeek's newest system is available through at least 16 non-Chinese inference services, including Amazon Bedrock. Each deployment expands the global footprint of Chinese AI, strengthens Beijing's soft power, and can produce licensing revenue for Chinese developers. Moonshot AI, for example, has sought revenue-sharing terms from major cloud providers for its Kimi K3 model once usage passes certain sales thresholds.

Key Takeaways

  • The Commerce Department issued a new rule aimed at blocking remote access to U.S. AI chips by Chinese companies.
  • American data centers abroad continue to host Chinese large language models on U.S.-designed chips.
  • China lacks sufficient domestic advanced chip production, making foreign infrastructure critical to its model deployment.
  • Open-weight Chinese models are especially well positioned to benefit, since users often select the best-performing openly available systems.
  • Revenue-sharing arrangements, such as Moonshot AI's terms for Kimi K3, mean Chinese developers can profit directly from overseas usage.
  • Banning foreign firms from serving Chinese models on U.S. cloud infrastructure could harm American and allied businesses and invite retaliation.
  • Washington is shifting toward encouraging competitive American open-weight models rather than relying on hardware exports alone.

Conclusion

The core tension is simple: success in exporting American chips and cloud capacity does not automatically translate into success for American AI models. Proprietary systems from U.S. developers like OpenAI and Anthropic still benefit from domestic infrastructure, but the open-weight market tells a different story. In that space, Chinese models are often the preferred choice, meaning U.S. computing power can end up subsidizing Chinese adoption.

Policymakers should not respond by cutting off access entirely, since that would raise costs for U.S. and allied companies and could trigger Chinese countermeasures. A more effective path involves investing in homegrown open-weight alternatives, with American firms including Meta and Thinking Machines already developing such models. The government can also direct federal procurement toward American open-weight systems where appropriate and write AI regulations that protect innovation instead of suppressing it. Without a deliberate strategy at the model layer, the United States risks building the stage for a competitor's performance.