The Complex Landscape of AI Chip Supply: US-China Dynamics in 2025
In recent years, the global artificial intelligence semiconductor arena has undergone rapid and transformative shifts, shaped primarily by geopolitical tensions and regulatory interventions between the United States and China. A vivid example of this complex interplay is the ongoing saga surrounding the export and usage of Nvidia's H20 and AMD's MI308 AI chips in China. Despite the critical importance of these chips for powering advanced AI applications, their availability and utilization are entangled in a web of export constraints, commercial agreements, and national security considerations that profoundly impact the market's trajectory.
US Export Restrictions and Conditional Market Access
The United States government initially imposed rigorous export controls in late 2023, aimed at restricting the shipment of cutting-edge AI chips to China amidst concerns over national security and technological leadership. Facing these limits, Nvidia developed the H20 chip tailored specifically to comply with these restrictions, featuring scaled-back performance metrics compared to its flagship models but still offering significant AI inference capabilities. Likewise, AMD’s MI308 chip was refined to align with these export rules, maintaining a balance between computational power and energy efficiency for cloud-based AI workloads.
In a significant yet controversial move, the Trump administration approved the resumption of sales for these chips to China in 2025 under a unique financial arrangement where both companies agreed to allocate 15% of their revenues from chip sales in China directly to the U.S. Treasury. While this conditional reopening enabled Nvidia and AMD to regain footing in the world’s largest AI chip market, it also introduced added costs for Chinese buyers and regulatory scrutiny on both sides.
Challenges from Chinese Regulatory and Security Concerns
Following the authorization to resume shipments, the Chinese government responded by issuing advisories to domestic technology companies—particularly those linked to government or national security projects—urging them to avoid dependence on Nvidia’s H20 and AMD’s MI308 chips. This cautious stance is fueled by fears over potential security vulnerabilities such as remote shutdown capabilities or data tracking embedded within foreign chips, allegations robustly denied by Nvidia.
Chinese authorities have conducted inquiries into major corporations including Tencent, Alibaba, and ByteDance, questioning their rationale for selecting foreign AI accelerators over domestic options. These actions reflect China’s strategic push to elevate indigenous semiconductor development as it seeks to reduce reliance on foreign technology and bolster its own AI chip industry.
Technical Overview: Nvidia H20 and AMD MI308 Features
The Nvidia H20 chip, built on a modified Hopper architecture, provides 96 GB of high-bandwidth memory (HBM3) with data transfer speeds near 4 terabytes per second and delivers around 296 teraflops of FP8 precision performance. Despite having lower raw power than Nvidia’s premium H100 and H200 models, the H20 excels in AI inference workloads, delivering fast and reliable responses suitable for cloud AI services handling millions of simultaneous queries.
AMD’s MI308, based on the CDNA accelerated compute architecture, supports similar HBM3 memory and is designed to offer a balanced mix of power efficiency and processing performance. Its integration with AMD’s open-source ROCm software ecosystem enhances compatibility with popular AI frameworks such as PyTorch and TensorFlow, easing deployment in scalable cloud environments.
Rising Domestic Alternatives: Huawei’s Ascend Chips
Amidst the uncertainty and added costs linked to US-made chips, China’s domestic semiconductor champion Huawei has made considerable strides with its Ascend series, including the high-performance Ascend 910C and 910D processors. These chips are supported by Huawei’s CloudMatrix platform, connecting hundreds of AI accelerators for powerful distributed computing across data centers.
The Ascend 910C in particular competes directly with Nvidia’s H20 on inference workloads. It offers competitive or higher memory configurations depending on the setup and benefits from seamless integration with Huawei’s MindSpore AI framework. Huawei’s tools also facilitate easy migration of AI models from widely-used frameworks like PyTorch and TensorFlow into its native environment, further boosting adoption.
This growing ecosystem has attracted major Chinese internet giants such as ByteDance and Baidu, who are increasingly deploying Ascend-based clusters, signaling a shift towards greater reliance on indigenous AI hardware solutions.
Market Implications and Competitive Strategies
The renewed but conditional presence of Nvidia H20 and AMD MI308 chips in China brings a new layer of complexity to the competitive landscape. The 15% revenue-sharing mandate imposed by Washington effectively raises the acquisition cost for Chinese enterprises, potentially influencing supplier choices and pricing negotiations in the cloud AI infrastructure sector.
Meanwhile, the domestic alternatives gain appeal not only due to their growing technical maturity and compatibility but also because of their market availability and local support, which help mitigate the risks associated with geopolitical tensions and regulatory uncertainties.
Many Chinese companies are adopting a hybrid approach to manage risks and optimize performance—employing Huawei’s Ascend processors for large-scale inference operations while reserving Nvidia and AMD chips for specific use cases that benefit from their mature software ecosystems (CUDA and ROCm, respectively).
Outlook: Balancing Innovation, Security, and Market Access
Looking ahead, the AI chip sector between the US and China remains at a critical juncture where technology innovation, security concerns, and geopolitical realities intersect. The ongoing negotiations, regulatory adjustments, and strategic corporate responses will shape the supply chain and AI capabilities available within China’s burgeoning AI market.
For Nvidia and AMD, maintaining a foothold in China will depend on navigating stringent export controls, alleviating security apprehensions, and justifying the value proposition amidst competitive pressures from rising domestic chipmakers like Huawei. Conversely, China’s drive for technological self-sufficiency and secure AI infrastructure underscores the urgency of advancing homegrown semiconductor technologies.
This dynamic tension will continue to create both challenges and opportunities, accelerating advancements in AI hardware while redefining global competition and cooperation in the critical domain of artificial intelligence.