Why is Microsoft Selling OpenAI Models in China While OpenAI and Anthropic Refuse?

In the current landscape of global artificial intelligence, a significant divergence exists in how major Western tech companies approach the Chinese market. While leading AI research organizations like OpenAI and Anthropic explicitly restrict direct access to their advanced models from within China, Microsoft continues to offer OpenAI’s technology to Chinese enterprise customers.

This discrepancy is not a contradiction of policy, but rather a result of complex corporate structuring, cloud computing partnerships, and the nuanced application of international export controls. The situation highlights a critical gray area in global AI governance, where access to advanced generative models is dictated largely by how the technology is hosted and delivered.

The Stance of OpenAI and Anthropic

OpenAI and Anthropic operate primarily as AI research and deployment companies. Their refusal to offer direct services in China is driven by a combination of regulatory compliance, security concerns, and corporate governance.

• Direct API Restrictions: Both companies actively block traffic originating from Chinese IP addresses to their consumer-facing applications and developer APIs. OpenAI enforced a significant restriction barring Chinese developers from accessing its API services, and Anthropic has similarly tightened access to its Claude models for users and organizations tied to China.
• Geopolitical Alignment: By restricting access, these companies align with broader United States government efforts to limit the proliferation of advanced, dual-use technologies to foreign adversaries.
• Corporate Governance: Both organizations maintain strict internal safety and alignment policies. Deploying models in regions with heavy digital censorship or differing human rights frameworks poses significant compliance and ethical challenges for these direct-to-consumer entities.

How Microsoft Operates in China

Microsoft’s ability to offer OpenAI models in China stems from its distinct position as a global enterprise cloud provider and its long-established operational footprint in the region.

• Localized Cloud Partnerships: Foreign cloud providers cannot independently own and operate data centers in China. Microsoft complies with local laws by partnering with a domestic operator, 21Vianet, to run Microsoft Azure in China. This creates a legally compliant, localized entity that is physically separated from Microsoft’s global cloud infrastructure.
• The Azure OpenAI Service: Microsoft does not sell OpenAI’s direct API or consumer products. Instead, it sells the Azure OpenAI Service, an enterprise-grade cloud product. Because Microsoft holds a comprehensive commercial partnership with OpenAI, it can host OpenAI’s models on its own Azure infrastructure.
• Proxy Access: Chinese enterprises purchase access to Azure China. Through this localized cloud environment, they can utilize the capabilities of OpenAI’s models without ever interacting directly with OpenAI’s servers or violating OpenAI’s direct IP restrictions. Multiple Azure customers in China have confirmed continued access to OpenAI’s models through this arrangement.

The Regulatory Gray Area

The divergence in market availability exposes a significant tension in international technology trade restrictions, one that regulators are actively working to address.

• Compute vs. Access: US export controls have historically focused on restricting the physical sale of advanced AI hardware, such as high-end GPUs, to China. However, regulating cloud-based API access, where the hardware remains outside the restricted region and only data is exchanged, is legally complex. Recent actions by the US Commerce Department have begun extending export control frameworks to AI models and cloud-based access, signaling that this gray area is narrowing.
• Protection of Model Weights: Microsoft does not transfer the underlying mathematical architecture, known as model weights, of OpenAI’s systems to Chinese servers. The models remain secured within Microsoft’s controlled infrastructure. Because the core intellectual property is not exported, the service has generally avoided triggering strict technology transfer restrictions, though evolving regulations may affect this going forward.
• Enterprise Compliance: Microsoft places the burden of regulatory compliance on its enterprise customers, requiring them to adhere to local laws regarding generative AI outputs, thereby shielding the core AI developer from direct liability.

Summary

Microsoft’s ability to sell OpenAI models in China, despite OpenAI and Anthropic’s refusal to do so directly, is a product of corporate infrastructure and a regulatory environment that is still catching up to cloud-based AI delivery. By leveraging localized cloud partnerships and offering AI as a hosted enterprise service rather than a direct consumer product, Microsoft provides access to advanced AI capabilities without transferring the core technological assets. As US regulators expand their focus beyond hardware to include AI model access itself, this dynamic will likely face increasing scrutiny.