With a domestically produced inference chip that runs Nvidia’s software, Alibaba is reshaping the U.S.–China tech rivalry.

Alibaba just made one of the boldest moves in China’s accelerating artificial intelligence (AI) race. The company has unveiled a new inference chip designed to run Nvidia’s popular CUDA software, without being made by Nvidia, and without requiring developers to change a single line of code. Even more importantly, this chip is manufactured at a domestic Chinese foundry, neatly dodging U.S. export restrictions. On the surface, it looks like a technical announcement. In reality, it is a geopolitical and economic statement.

The Nvidia Bottleneck

At the heart of today’s AI revolution sits a silicon giant: Nvidia. Its GPUs power everything from OpenAI’s ChatGPT to autonomous driving platforms. The company’s CUDA ecosystem, a software stack developed over more than a decade, has become the de facto operating language of machine learning. Developers trained on CUDA are locked into it, and by extension, into Nvidia hardware.

But Washington’s export restrictions have dramatically complicated the equation. Nvidia’s most powerful AI accelerators, such as the A100, H100, and even the “toned-down” H20, are restricted from being sold to Chinese firms. The ban is not just about hardware; it is about slowing the pace at which Chinese companies can train frontier AI models that compete with their American counterparts.

This has left Chinese firms with two unattractive options:

1. Accept second-best—settling for less powerful chips that lag behind in performance.
2. Reinvent the wheel—building entirely new ecosystems of chips and software, a process that could take years, if not decades.

Alibaba, however, just introduced a third option.

Alibaba’s Play: Compatibility Over Perfection

Instead of trying to build the perfect chip, Alibaba’s engineers pursued a more pragmatic strategy: make something that runs CUDA, works with existing AI applications, and can drop into the same infrastructure that was designed for Nvidia. This means developers don’t need to rewrite code. Cloud customers don’t need to overhaul their systems. The transition, in theory, can be almost seamless.

The genius here lies not in outpacing Nvidia’s hardware, but in leveraging its ecosystem. Alibaba isn’t trying to reinvent CUDA; it’s borrowing it. In effect, if Chinese companies can’t buy Nvidia, they can “be” Nvidia, at least for inference workloads. And Alibaba has the resources to execute this strategy. With a $53 billion three-year investment plan for its cloud and AI infrastructure, the company is betting that good-enough compatibility is more valuable than raw chip superiority.

Rivals in Different Corners

Alibaba’s maneuver becomes clearer when viewed against its competitors:

• MetaX has attempted to bypass export restrictions by “stitching together” smaller, less restricted chips. While clever, this approach is technically complex and not always efficient.
• Huawei has been pushing its Ascend chips, but adoption is limited. Many private firms are wary of relying on hardware built by a direct competitor in the cloud market. Worse, Ascend chips are reportedly power-hungry and lack CUDA compatibility, which adds friction for developers.

Alibaba’s new chip sidesteps both problems: it offers a non-Huawei option and runs software developers already know.

The Bigger Picture: What’s at Stake

This development doesn’t mean Nvidia’s dominance is over. Training state-of-the-art AI models still requires top-tier hardware that China cannot currently replicate at scale. That remains Nvidia’s stronghold. But inference, the process of running AI models once they’re trained, is an enormous and growing market. According to McKinsey, inference accounts for roughly 90% of AI-related computing demand once models are deployed.

By capturing this layer, Alibaba ensures that Chinese companies won’t be left stranded when U.S. restrictions tighten. In other words, this is less about throwing a knockout punch and more about building a safety net. The symbolic message is equally powerful: China doesn’t need to abandon Western ecosystems entirely. With clever engineering and enough capital, it can adapt them for domestic use.

Risks and Open Questions

Still, there are unresolved challenges:

• Performance: Will Alibaba’s chip truly deliver comparable performance to Nvidia’s H20 in real-world scenarios, or is “compatibility” more cosmetic than practical?
• Sustainability: Can Chinese foundries keep pace with global manufacturing standards without access to leading-edge tools from companies like ASML in the Netherlands?
• Strategy: Will Alibaba keep this chip as an internal tool for its cloud services, or eventually spin it out as a broader hardware offering?

The answers to these questions will shape not only the future of Alibaba but also the trajectory of China’s AI industry.

A Familiar Playbook

What Alibaba has done mirrors a long-standing strategy in China’s tech sector: if you can’t import the world’s best, replicate it locally. We have seen this in e-commerce, social media, electric vehicles, and even aerospace. It’s a mix of pragmatism and necessity. The difference here is timing. The AI race is not measured in decades but in months. The pace of progress is relentless, and any delay in access to the best hardware can translate to a serious competitive disadvantage. By ensuring CUDA compatibility, Alibaba has bought itself—and China—time.

Adaptation Over Revolution

Alibaba’s new chip may not rewrite the rules of AI. But it reaffirms an old lesson: in global technology races, sometimes the smartest move isn’t to fight head-on; it’s to adapt the playbook of your rival and make it your own. Whether this strategy will allow China to compete with Nvidia in the long run truly remains to be seen. But for now, it marks a significant turning point in how Chinese firms navigate the U.S.-China tech rivalry. 

 Originally Published on LinkedIn.