Huawei is reportedly planning a major shift in its AI chip strategy—from application-specific integrated circuits (ASICs) to general-purpose graphics processing units (GPGPUs)—in an effort to capture greater market share from U.S. rival NVIDIA, according to a July 11 report by The Information.

Despite ongoing U.S. export restrictions that prevent NVIDIA from selling its most advanced AI chips to mainland China, the company's GPGPU-based products remain the most widely used in the region. This dominance is largely attributed to NVIDIA's powerful CUDA ecosystem, which supports a wide range of workloads from deep learning to scientific computing.

While Huawei's Ascend series AI chips are optimized ASICs designed for inference and training in deep learning, they lack the flexibility of GPGPUs for broader tasks such as FP64 scientific computing and complex graphics rendering. NVIDIA's H100 and H20 chips, by contrast, support both low-precision and double-precision floating-point operations, making them suitable for a wider array of applications.

A key limitation for Huawei lies in its self-developed software platform, CANN (Compute Architecture for Neural Networks), which lags far behind NVIDIA's CUDA in terms of developer adoption, ecosystem maturity, and library support. Huawei has also promoted its MindSpore deep learning framework, but it has not gained comparable traction.

The report notes that Huawei's future GPGPU-based AI chips will likely include a middleware layer compatible with CUDA, translating NVIDIA's instructions to Huawei's own execution architecture. Sources suggest the company may also adopt compute models similar to those used by NVIDIA and AMD.

This shift signals Huawei's intention to expand beyond narrow-use AI accelerators and enter the more flexible and scalable GPGPU segment—potentially making its chips more attractive to a wider range of Chinese customers.

Meanwhile, NVIDIA is preparing to launch a new AI chip specifically designed for the Chinese market as early as September. The chip will be based on the existing Blackwell RTX Pro 6000 but modified to comply with U.S. export regulations. Key high-end features—such as high-bandwidth memory (HBM) and NVLink interconnect—will reportedly be removed to meet legal thresholds.

According to insiders, Chinese clients have already tested samples of the new chip and expressed strong interest. However, ongoing uncertainty in U.S. policy has raised concerns about overdependence on NVIDIA hardware.

NVIDIA CEO Jensen Huang is expected to visit Beijing next Wednesday to attend the China International Supply Chain Expo and meet with senior Chinese officials. He is likely to reiterate the company's commitment to the Chinese market despite export headwinds.

NVIDIA's market momentum remains strong: fueled by soaring AI demand, its market cap briefly surpassed $4 trillion this week, making it the world's first company to reach that milestone.