Chip War 2025: US vs China AI Technology

👋 If you are a new reader, my name is Danar Mustafa. I write about product management focusing on AI, tech, business and agile management. You can visit my website here or visit my Linkedin here.

In today’s article, I will talk about the current state in the Chip War between US and China in 2025.

US Nvidia Chips is market leaders

US is still the number one in AI chip manufacturing. Deepseek shook the AI world. However, it only had access to lower performance AI Chips called A100. In some cases, it used H800 chips. Here is the full Nvidia chips list with performance metrics.

1. NVIDIA Blackwell Architecture

Transistor Count: 208 billion transistors
Manufacturing Process: Custom-built 4NP TSMC process
Chip-to-Chip Link: 10 TB/second
Performance: Supports AI training and real-time inference for models scaling up to 10 trillion parameters
Key Features:
- Second-Generation Transformer Engine with micro-tensor scaling support
- Fifth-Generation NVLink with 1.8TB/s bidirectional throughput per GPU, allowing seamless communication among up to 576 GPUs
- RAS (Reliability, Availability, Serviceability) Engine for enhanced system uptime and diagnostics
- Secure AI capabilities for protecting models and data

2. NVIDIA Hopper Architecture

Transistor Count: Approximately 80 billion transistors (specific count may vary by model)
Manufacturing Process: TSMC’s N5 process technology
Performance: Designed for high-performance computing and AI workloads, supporting large-scale models.
Key Features:
- Advanced tensor cores optimized for deep learning tasks
- High bandwidth memory (HBM3) for faster data access
- Support for multi-instance GPU technology, allowing multiple networks to run simultaneously on a single chip

3. NVIDIA A100 Tensor Core GPU

Transistor Count: 54 billion transistors
Memory Configuration: Up to 80 GB of HBM2 memory
Performance:
- FP32 Performance: Up to 19.5 TFLOPS
- FP16 Performance: Up to 156 TFLOPS
- INT8 Performance: Up to 624 TOPS
Key Features:
- Multi-instance GPU technology enabling multiple workloads on a single GPU
- High bandwidth memory interface providing up to 1555 GB/s memory bandwidth

4. NVIDIA RTX Series (e.g., RTX A6000)

Transistor Count: Approximately 28 billion transistors (for the A6000)
Memory Configuration: Up to 48 GB GDDR6 memory
Performance:
- FP32 Performance: Up to around 38.7 TFLOPS
- Ray Tracing Cores and Tensor Cores for real-time ray tracing and AI-based tasks.
Key Features:
- Hardware support for real-time ray tracing
- AI-enhanced graphics capabilities through DLSS (Deep Learning Super Sampling)

China is 2 years behind global leaders

China’s position in AI chip technology is complex and varies across different aspects of the industry. In chip design, Chinese firms are about 2 years behind global leaders for logic chips and several more years behind for memory chips. For semiconductor manufacturing equipment, China could be five generations behind.

In terms of advanced chip manufacturing processes:

SMIC, China’s leading chipmaker, is 1-2 generations behind global leaders, having only begun large-scale production of 14nm chips in 2022.
Compared to TSMC, China is estimated to be:
- At least 10 years behind according to Taiwan’s National Science and Technology Council.
- Approximately 3 years behind according to some recent analyses.

The gap in AI chip capabilities is influenced by U.S. export controls on advanced chipmaking equipment and design tools. These restrictions aim to keep China 10-15 years behind in high-end chip production.

Despite these challenges, China has made significant progress in AI model development. Some Chinese AI models now rival top open-source models from the U.S., though they still lag about a year behind the best closed models. However, AI models need increasingly more computing power. The chip restrictions may significantly impact China’s ability to keep pace in AI development.

U.S. vs. Chinese AI Model Performance

How is China trying to minimize the gap?

China is actively working to bridge the gap in AI chip technology through several strategic initiatives:

State-driven efforts: The Chinese government has placed AI development high on its agenda, as evidenced by discussions during the recent two sessions. The State-owned Assets Supervision and Administration Commission (SASAC) is pushing centrally administered State-owned enterprises to integrate AI development into their overall planning.
Focus on vertical domains: Chinese companies are concentrating on developing specialized AI models for specific business domains, which could potentially surpass general-purpose models like GPT-4.0 in certain areas1.
Emphasis on applications: Chinese AI startups are focusing more on adapting technology to various industries and commercializing different applications, showing flexibility in creating personalized products and innovations.
Domestic chip production: Huawei and SMIC are leading China’s efforts to produce advanced AI chips domestically. SMIC has begun large-scale production of 14nm chips, though it still lags behind global leaders by one to two generations.
Government financial support: The Chinese government is providing financial aid to nurture high-potential firms, especially in regions that might otherwise be overlooked, to foster a broader base for innovation.
Academic-industry collaboration: Prestigious institutions like Tsinghua University are supporting a new generation of AI startups that are pushing the boundaries of AI innovation, particularly in generative AI.
Decentralized approach: Local governments in China have been allowed to tailor AI development strategies to their specific contexts, promoting a more diverse and adaptable innovation ecosystem.

Despite these efforts, China still faces challenges in key areas such as computing power, cloud servers, and AI chips, where there remains a significant gap with the United States. The impact of U.S. export controls on advanced chipmaking equipment and design tools also continues to pose obstacles for China’s AI chip development.

Huawei vs Nvidia Chips

Huawei’s AI chips are currently lagging behind Nvidia’s in performance and stability, but the exact gap is difficult to quantify precisely. While Huawei has made significant strides in developing its Ascend series of AI chips, they still face several challenges:

Performance: Huawei’s latest Ascend 910B chip is reported to be comparable to Nvidia’s A100 in raw computing power. However, multiple industry insiders have stated that Huawei’s chips still lag far behind Nvidia’s for the most advanced AI work.
Software ecosystem: Nvidia’s CUDA software platform, introduced in 2006, gives it a significant advantage. Huawei’s software ecosystem is less mature, leading to stability issues and slower inter-chip connectivity.
Product maturity: Huawei’s Cann (Compute Architecture for Neural Networks) was only introduced in 2018, while Nvidia’s ecosystem has been developing since the early GPU days.

Despite these challenges, Huawei is making progress:

The upcoming Ascend 910C chip is reported to match Nvidia’s H100 in performance.
Huawei plans to produce 1.4 million 910C chips by December 2025.
The company is leveraging its strong customer service to work closely with clients and address issues.

It’s important to note that the AI chip landscape is rapidly evolving, and the gap between Huawei and Nvidia may narrow in the future, especially with continued investment and support from the Chinese government.