OpenAI cuts API access in China, sparking a domestic land grab

For the past year, thousands of Chinese developers operated on a fragile workaround: fire up a VPN, ping OpenAI's servers, and quietly build apps, benchmark research, or fine-tune models on top of GPT-4.

That loop just broke. On July 9, OpenAI officially started blocking API traffic from unsupported regions, cutting off Chinese developers directly from the US side.

The great domestic migration race

Within hours of OpenAI's notification, China's domestic AI giants went on an aggressive offensive to scoop up stranded developers.

At the World AI Conference in Shanghai, SenseTime unveiled SenseNova 5.5, boldly claiming it matches GPT-4o after showing off real-time multimodal tests with plush toys and drawings. To sweeten the pitch, SenseTime handed out 50 million free tokens and offered dedicated staff to handle migrations from OpenAI for free.

Others quickly matched the ante. Baidu rolled out 50 million free tokens for Ernie 3.5 along with free onboarding, alongside launching Ernie 4.0 Turbo. Zhipu AI stepped up with 150 million free tokens, and Tencent Cloud offered 100 million tokens through the end of July.

According to Xiaohu Zhu, founder of the Center for Safe AGI in Shanghai, OpenAI's sudden cut prompted serious worries about equitable access, but also created an open hunting ground for domestic providers. China now boasts at least 130 large language models—roughly 40% of the global total, trailing only the US.

The hardware wall

While software migration campaigns look seamless on paper, hardware bottlenecks tell a messier story.

Washington's export curbs on advanced semiconductors have severely squeezed local compute. The Information recently reported that short-video giant Kuaishou had to throttle access to its buzzing text-to-video model Kling, purely due to a shortage of chips to run it.

NYU professor Winston Ma noted that Chinese tech players were already closing the performance gap while essentially giving models away. But he pointed out that many local companies had prioritized commercialization over deep foundational research. OpenAI's departure is a short-term shock, but it forces domestic LLMs into the real fire test.

Why it matters

For developers in the region, the luxury of relying on OpenAI APIs as a crutch or benchmarking yardstick is gone overnight. If your startup built its core pipeline around GPT-4, you now have to test whether SenseNova, Ernie, or GLM can actually match that precision under production workloads, free tokens notwithstanding.

For the global AI landscape, this marks the official bifurcation of the ecosystem. By actively enforcing regional blocks, OpenAI is sealing off backdoor access and accelerating total technological decoupling.

We think the frantic price war among Chinese tech giants is a double-edged sword. Dishing out hundreds of millions of free tokens makes for flashy PR, but running cutting-edge models without access to top-tier Nvidia chips burns scarce compute fast. If these subsidies eat into the margin needed for fundamental research, closing the remaining frontier gap will only get harder.