The US government has given a new clarification that export restrictions on advanced AI chips apply not only to companies operating in China but also to Chinese-owned firms and subsidiaries located overseas. The move closes a potential loophole that may have allowed Chinese companies to access restricted AI hardware through operations in countries outside mainland China.

Why the US Made This Move

The clarification follows concerns in Washington that Chinese firms could obtain advanced AI processors through overseas subsidiaries despite existing export controls. Some policymakers argued that uncertainty created after the cancellation of the Biden-era AI Diffusion Framework may have opened a pathway for such purchases. The new guidance makes it clear that Chinese-headquartered companies remain subject to licensing requirements regardless of where they operate.

Pressure Grows on China’s AI Industry

The tighter enforcement is expected to make access to advanced US AI hardware more difficult for Chinese companies. NVIDIA’s most powerful Blackwell AI chips remain restricted, while purchases of other advanced processors continue to face regulatory scrutiny. As a result, Chinese technology firms are being pushed to rely more heavily on domestic alternatives.

China Accelerates Its Chip Strategy

In response to years of US restrictions, China has been rapidly expanding its domestic AI semiconductor industry. Companies including Huawei, Moore Threads, Biren Technology, Cambricon, and Alibaba have increased investments in AI processors designed for both training and deploying large AI models. Rather than focusing solely on replacing Nvidia products, many firms are now building complete AI hardware and software ecosystems.

Domestic Alternatives Gain Importance

Industry analysts expect Chinese AI companies to increasingly adopt locally developed accelerators as access to foreign chips becomes more limited. Huawei’s Ascend series and Alibaba’s custom AI processors are among the products attracting attention as China works to reduce dependence on US technology.

What to Expect Next

The latest US action is unlikely to slow China’s AI ambitions. Instead, it may further accelerate investment in domestic chips, software platforms, and AI infrastructure. The competition between the US and China is increasingly shifting from individual products to entire technology ecosystems. In the coming years, the key question may no longer be whether China can replace Nvidia, but how quickly it can build a self-sufficient AI industry capable of competing on a global scale.

Read More:

• China vs US in Robotics: Nvidia CEO Reveals the Real Power Shift
• Huawei aims to launch lithography equivalent to TSMC’s 1.4nm by 2031 in a major defiance of U.S. sanctions
• OnePlus Ace 6 Ultra vs OnePlus Ace 6: Are the Upgrades Really Worth the Extra Money?

The post US Expands AI Chip Restrictions Worldwide: What It Means for China appeared first on Gizmochina.

Highlights:

• Huawei launches Atlas 350, focused on AI inference, not training
• Claims up to 2.8× performance boost over Nvidia’s H20 chip
• Powered by Ascend 950PR, part of China’s push for AI independence

New Focus: AI Inference Over Training

Huawei’s Atlas 350 reflects a clear strategic shift toward AI inference, the stage where trained models are deployed in real-world applications. Instead of focusing on training, this accelerator is designed to efficiently run AI systems such as recommendation engines, large language models, and multimodal tools. This aligns with a broader industry transition, where inference is increasingly becoming the main challenge as AI adoption scales across industries.

Performance Designed for Real-World Use

At the heart of the Atlas 350 is Huawei’s Ascend 950PR chip, capable of delivering around 1.56 petaflops of FP4 computing power. Huawei claims this enables up to 2.8 times better performance than Nvidia’s H20 in certain inference workloads. The use of FP4 precision allows faster processing and lower memory usage, making it particularly effective for large-scale deployment where efficiency and speed are more critical than high precision.

Memory and Architectural Advancements

A notable development is Huawei’s move toward in-house high-bandwidth memory, with reported configurations of up to 128GB and 1.6 TB/s bandwidth. This reduces reliance on external suppliers and strengthens China’s semiconductor capabilities. Additionally, the chip is optimized for the “prefill” stage of inference, improving how quickly input data is processed, which is essential for real-time AI applications and emerging agent-based systems.

Building a Complete AI Ecosystem

Huawei’s approach extends beyond a single accelerator. The company is developing an integrated AI ecosystem that includes hardware, storage, and large-scale computing systems. This vertically integrated strategy mirrors established players but is tailored to domestic infrastructure, enabling tighter control over performance and deployment.

Impact on the US–China Tech War

The Atlas 350 is significant for the ongoing US–China technology competition. With restrictions limiting access to advanced US chips, Huawei’s development of competitive domestic alternatives represents a step toward technological self-reliance. For China, this strengthens its ability to build independent AI infrastructure, while for US companies such as Nvidia, it introduces growing competition in a key market.

Implications for Consumers and the Market

For everyday users, this could mean faster and more affordable AI services, especially in China. It may also lead to more local AI apps and platforms. Over time, the tech world could split into separate systems, with different regions building their own AI ecosystems.

Read More:

• China vs US in Robotics: Nvidia CEO Reveals the Real Power Shift
• Xiaomi Robot Vacuum 6 Max vs Roborock S9 MaxV Ultra: One Cleans Harder, One Cleans Smarter
• BMW Starts Using Humanoid Robots in Factory to Build Electric Cars

(via)

The post US – China Tech War: Huawei Takes On Nvidia with Atlas 350 AI Accelerator appeared first on Gizmochina.