Will the NVIDIA Monopoly Myth Be Broken? Undercurrents in the AI Computing Chip Race

Introduction: NVIDIA—Designing a New “Export-Compliant AI Chip”

Recently, after the U.S. government banned NVIDIA’s H20 chip for the Chinese market, the American chip giant is accelerating the development of another AI chip designed to comply with U.S. export regulations. According to U.S. media, NVIDIA has informed three Chinese companies that it is redesigning its AI chips to continue supplying products to China without violating U.S. export controls. Industry analysts point out that this move may reshape the global AI chip supply chain landscape.

1. AI Computing Chips: Exploding Demand and Market Restructuring

In China, short-video platforms like Douyin and Kuaishou have become the “main battlefield” for short dramas. For instance, in March 2024, Douyin’s daily active users (DAUs) for short dramas grew by 120%, views increased by 350%, and paid user numbers jumped tenfold. The variety of domestic short dramas is immense—comedies like Aunt’s World 2, which used the theme of “new-style elderly care,” reached over 1.7 billion views, while Chinese New Year focused on family emotions, topping 1 billion views and moving audiences to both laughter and tears.

AI Chip vs. AI Computing Chip

What exactly is an AI computing chip? While often used interchangeably, “AI chips” and “AI computing chips” serve different roles.

AI chips broadly refer to processors that support AI functions—such as NPUs and FPGAs—used across cloud, edge, and terminal devices (e.g., AI processing in smartphones). AI computing chips are specialized for delivering massive computing power, primarily for large-model training and high-performance inference.

Exponential Growth in Computing Demand

As AI technology advances, demand is shifting from general AI chips to AI computing chips that emphasize “thinking” and “reasoning.” With large models like DeepSeek and ChatGPT driving training and deployment, global computing demand is soaring exponentially. Between 2012–2023, computing demand increased hundreds of thousands of times, far exceeding Moore’s Law, with model training demands continuing to climb each year.

(Source: Jazzyear)

According to IDC, global computing capacity is projected to grow from 1,397 EFLOPS in 2023 to 16 ZFLOPS by 2030, with a compound annual growth rate (CAGR) of 50%. In China, intelligent computing capacity was 725.3 EFLOPS in 2024, expected to reach 2,781.9 EFLOPS in 2028, with a CAGR of 46.2% from 2023–2028.

(Source: Central China Securities)

With this surge in demand, the AI server market is also expanding rapidly. Projections show the global AI server market will hit $158.7 billion in 2025 and $222.7 billion in 2028 (CAGR: 15.5%), with generative AI servers’ share rising from 29.6% in 2025 to 37.7% in 2028.

In China, IDC forecasts the AI server market will reach $19 billion in 2024, grow to $25.9 billion in 2025 (+36.2% YoY), and hit $55.2 billion in 2028 (CAGR: 30.6%).

(Source: Central China Securities)

2. From GPU Monopoly to ASIC Rise: Diversified Development of AI Computing Chips

Currently, GPUs dominate the AI computing chip sector, with NVIDIA holding a leading global position. However, U.S. restrictions have accelerated China’s independent innovation and development, boosting domestic chip alternatives and narrowing the performance gap.

In the first half of 2024, China’s accelerator chip market surpassed 900,000 units. GPUs accounted for 80%, while non-GPU chips held 20%. Domestic brands shipped nearly 200,000 units (20% share), with chips like Huawei’s Ascend 910B and Zhonghao Xinying’s “Shana” achieving performance breakthroughs, particularly in inference tasks.

(Source: Jazzyear)

Meanwhile, North American cloud service providers are also heavily investing in self-developed AI ASICs, including Google’s TPU, Amazon’s Trainium and Inferentia, Microsoft’s Maia 100, and Meta’s MTIA.

Compared to GPUs, ASICs—customized for specific algorithms and applications—offer strong computing power and higher energy efficiency for targeted tasks.

While GPUs outperform ASICs in absolute compute power and inter-chip connectivity, ASIC servers rely on Ethernet interconnects, which bring broader compatibility, open ecosystems, and lower costs.

(Source: Southwest Securities)

3. Collaborative Development: AI Processing Shifts from Cloud to Edge

AI computing chips can be categorized by application: cloud, edge, and terminal.

(Source: Central China Securities)

• Cloud AI chips handle massive datasets and computation, requiring the highest performance and density.
• Edge AI chips balance power and performance, sitting between cloud and terminal needs.
• Terminal AI chips prioritize low power consumption and efficiency, with relatively lower compute demands.

(Source: Central China Securities)

Currently, AI processing is gradually shifting from cloud toward edge. Enterprises are advancing cloud–edge collaboration, enabling faster upgrades in device-side AI capabilities and enhancing competitive differentiation.

(Source: SPDB International)

The shipment of AI-enabled devices capable of running large models locally is rising significantly, boosting demand for higher-performance chips. Looking ahead to 2025, AI smartphones are expected to reach 29% penetration, supported by chips optimized for on-device AI large models.

Conclusion

As enterprises drive toward cloud–edge integration, 2025 will bring new opportunities for collaborative AI ecosystems. At the same time, major cloud providers’ self-developed chips are emerging strongly, increasingly challenging NVIDIA’s dominance. Before NVIDIA’s next-generation GPUs hit the market, the AI computing chip sector is poised for a fundamental reshaping.