Why China Is Confident in Banning Nvidia GPUs: Can Domestic Alternatives Really Replace Them?

A New Era in China’s AI Chip Market

In a striking admission at a recent Citadel Securities event, Nvidia CEO Jensen Huang revealed that his company’s market share in China has plummeted from 95% to essentially zero. “We have completely exited the Chinese market,” Huang stated, acknowledging that U.S. export restrictions have cost American companies access to one of the world’s largest markets—a market Huang previously estimated would reach $50 billion in AI spending within two to three years.

This dramatic shift raises crucial questions:

Why is China confident enough to effectively ban Nvidia GPUs?
Can domestic alternatives truly replace them?
How are Chinese companies addressing the formidable challenge of Nvidia’s CUDA ecosystem?

Why China Is Confident in Banning Nvidia GPUs: Can Domestic Alternatives Really Replace Them? How Chinese Companies Are Tackling the CUDA Problem？

China’s Domestic GPU and AI Chip Alternatives

The vacuum left by Nvidia’s departure has been rapidly filled by a diverse ecosystem of Chinese chipmakers, following two distinct technological approaches:

General-Purpose GPU Manufacturers

Moore Threads: Perhaps the most ambitious GPU startup, Moore Threads has developed four generations of GPU architectures, spawning over ten different chip models. These chips span three major domains: AI computing, professional graphics processing, and desktop-level graphics acceleration, demonstrating remarkable versatility in addressing different market segments.

MetaX: Focusing on integrated training and inference GPU chips, MetaX has achieved significant commercial traction. By Q1 2025, the company reported cumulative sales exceeding 25,000 chips, with deployments across nine major intelligent computing clusters in Beijing, Shanghai, Hangzhou, and other cities. Their flagship Xiyun C600 GPU represents a milestone—China’s first fully domestically produced general-purpose GPU with a complete supply chain from design through manufacturing to packaging and testing.

Other GPU players: The landscape also includes Hygon Information, Tianshu Zhixin, and Biren Technology, each carving out niches in the general-purpose computing market.

Specialized ASIC/DSA Manufacturers

Huawei Ascend: Leveraging years of technological accumulation and ecosystem advantages, Huawei has emerged as the dominant force in China’s high-end AI chip market. The Ascend 910 chip has been deployed in large-scale AI model training by major Chinese internet companies, demonstrating practical viability in real-world applications.

Cambricon: A publicly traded pioneer in AI chips, Cambricon’s Siyuan 590 chip supports mainstream deep learning frameworks and achieves approximately 80% of Nvidia’s A100 performance in visual processing and large language model scenarios. The company has secured partnerships with leading AI firms like Zhixiang Future and Baichuan Intelligence, providing computing power for internet giants’ search services and advertising recommendation systems.

Other ASIC players: The field includes Kunlunxin and Enflame Technology, each developing specialized architectures optimized for specific AI workloads.

Market Momentum: Rising Domestic Market Share

The numbers tell a compelling story of rapid domestic adoption. According to IDC statistics, Chinese-developed AI chips captured 30% of the domestic market in 2024—a remarkable achievement given Nvidia’s near-monopoly just years earlier. Projections suggest this share will exceed 50% by 2025, marking a tipping point where domestic chips become the majority choice.

Industry insiders note that the usage ratio between domestic and foreign chips has already reached roughly 50-50, with domestic chips gaining increasing acceptance and credibility among Chinese enterprises.

The CUDA Challenge: China’s Achilles Heel?

However, confidence in hardware capability doesn’t tell the whole story.

The elephant in the room remains Nvidia’s CUDA ecosystem—a comprehensive software framework that has become the de facto standard for AI development.

CUDA’s decade-plus head start has created a vast library of optimized code, tools, and trained developers that any competitor must somehow address.

How Chinese Companies Are Tackling the CUDA Problem

Chinese chipmakers are pursuing several parallel strategies:

1. Framework Compatibility Layers: Most domestic chips, including Cambricon’s Siyuan 590, explicitly support “mainstream deep learning frameworks.” This suggests compatibility layers that allow popular frameworks like PyTorch and TensorFlow to run on non-Nvidia hardware, abstracting away some CUDA dependencies.

2. Ecosystem Investment: Companies recognize that bridging the software gap requires sustained investment. While the technology barrier isn’t insurmountable—unlike advanced manufacturing processes—building a mature ecosystem demands time and resources.

3. Vertical Integration: Huawei’s approach exemplifies this strategy, developing not just chips but entire computing stacks including frameworks, tools, and developer resources, creating a self-contained ecosystem less dependent on CUDA.

4. Pragmatic Deployment: Rather than achieving complete CUDA parity immediately, Chinese chips are being deployed in production environments where 80% of A100 performance (as with Cambricon) proves sufficient, allowing ecosystems to mature through real-world usage.

Why China Feels Confident Despite Remaining Challenges

China’s confidence in moving away from Nvidia stems from several factors:

Market Scale: As the world’s largest AI market with massive domestic demand, China can support multiple chip vendors and ecosystem development through sheer volume.

Design Capability: Industry observers note that the gap in chip design isn’t substantial. The primary bottleneck remains advanced manufacturing processes (access to cutting-edge lithography), not architectural innovation.

Software as a Solvable Problem: Unlike manufacturing constraints imposed by export controls, software ecosystem development faces no fundamental technical barriers. Given sufficient investment and time, the CUDA gap can theoretically be closed—it’s an engineering challenge rather than a physics problem.

Strategic Necessity: U.S. restrictions have transformed chip independence from an option into an imperative, mobilizing resources and political will that might not otherwise exist.

Rapid Progress: The speed of market share growth—from near-zero to 30% in roughly two years—demonstrates that domestic alternatives are viable enough for production deployment, even if not yet perfect substitutes.

The Remaining Gaps

Despite this optimism, significant challenges persist:

Advanced manufacturing processes remain constrained by limited access to cutting-edge lithography equipment
Software ecosystem maturity still lags substantially behind CUDA’s comprehensive toolchain
Performance parity hasn’t been fully achieved, with most domestic chips offering 70-80% of equivalent Nvidia performance
Developer mindshare and training remain heavily skewed toward CUDA-based workflows

Conclusion

China’s confidence in effectively banning Nvidia GPUs isn’t based on achieving complete technical parity, but rather on demonstrating “good enough” alternatives that can support domestic AI development while ecosystems mature. With 30% market share already achieved and projections of majority adoption by 2025, Chinese chipmakers have proven they can deliver viable products.

The CUDA ecosystem remains a formidable advantage for Nvidia globally, but in China’s protected market, the combination of capable hardware, improving software frameworks, massive domestic demand, and strategic necessity has created conditions where domestic alternatives can thrive despite remaining limitations.

Whether this confidence proves justified in the long term depends on continued investment in both hardware and software development, and whether “good enough” performance today can evolve into genuine leadership tomorrow. For now, China appears willing to accept some performance trade-offs in exchange for technological sovereignty—a calculation driven as much by geopolitics as by pure technical merit.