US-China AI Rivalry Expands Beyond Hardware to Software Ecosystems

The competition for artificial intelligence dominance between the United States and China is shifting from hardware manufacturing to the software layers that optimize chip performance. While Nvidia currently maintains a significant lead through its proprietary CUDA platform, Chinese firms like Huawei are aggressively developing alternative ecosystems to bypass U.S. export controls. This evolution is critical for the AI sector as it determines whether the industry will remain consolidated under a single standard or split into bifurcated technological stacks.

Nvidia currently controls the global AI hardware market, accounting for approximately half of the world's installed chip stock and two-thirds of total computing capacity. This dominance is reinforced by its proprietary CUDA platform, which has become the de-facto standard for AI development, allowing Nvidia to command exceptional pricing power with profit margins between 70% and 80%. The platform creates powerful network effects where developers optimize for CUDA-compatible hardware, making it difficult for competitors like Amazon, Google, or Meta to break the cycle of dependency despite developing their own internal chips.

Huawei has emerged as the primary challenger to Nvidia's hegemony by developing a parallel AI stack consisting of Ascend processors and the CANN software architecture. To overcome the "chicken-and-egg" problem of developer adoption, Huawei announced the open-sourcing of its CANN toolkit in August 2025 and introduced torch_npu, a plugin that allows standard PyTorch code to run on Ascend hardware. This strategy aims to lower switching costs for developers who are already familiar with the PyTorch framework, effectively mirroring the state-supported tactics China previously used to bolster its domestic hardware industry through subsidies and procurement preferences.

The collaboration between Huawei and the AI lab DeepSeek represents a strategic shift toward using open-weight models to bridge the gap between competing ecosystems. DeepSeek’s models are engineered for compatibility with both Nvidia and Huawei processors, widening the addressable market for Chinese hardware while providing a familiar entry point for global developers. If China successfully crosses the "CUDA moat," the durable U.S. advantage in AI could narrow significantly, potentially leading to a fragmented market where hardware and software strengths are no longer mutually reinforcing.

While the U.S. and China battle for stack dominance, Europe remains an essential input provider rather than a full contestant in the AI race. Despite possessing critical assets like chip lithography equipment, the region lacks a global AI chip designer or a software layer analogous to CUDA or CANN. This gap highlights the importance of co-design between hardware and software, as economic value in the AI sector increasingly accrues to those who control the entire platform rather than those who simply supply the manufacturing equipment.