Is the US afraid of Huawei? Reports point to the Ascend 910C accelerator’s performance to explain the surprising reversal of Nvidia's H200 AI GPU ban on China

• The White House allows Nvidia H200 exports to China with a 25% fee
• US officials evaluated strategies ranging from full export bans to market flooding
• President Trump says H200 exports support American jobs and manufacturing efforts

The White House has authorized the export of Nvidia H200 AI accelerators to China, attaching a 25% fee per shipment.

According to reports, the decision was influenced by Huawei’s rapid development of its Ascend 910C chips, particularly the CloudMatrix 384 system, which integrates 384 of these accelerators.

Internal sources suggest that the US move aims to maintain American dominance in the global tech ecosystem while keeping the country’s proprietary Blackwell and Rubin architectures restricted.

Huawei’s CloudMatrix 384 performance

Huawei’s CloudMatrix 384 has been described as a “nuclear-level product” capable of delivering 300 petaflops of dense BF16 compute.

It surpasses Nvidia’s GB200 NVL72 system on certain performance metrics, highlighting its raw computational power.

The system also provides 3.6 times more aggregate memory and over twice the memory bandwidth compared to Nvidia’s platform.

However, these gains come at the cost of nearly four times the power consumption, raising efficiency concerns.

These accelerators have been deployed in Huawei data centers, where abundant electricity reduces the importance of energy efficiency.

The company plans to scale Ascend 910C production to hundreds of thousands of units next year, with projections suggesting millions could be manufactured by 2026.

Despite China’s development of its own AI instruction set through CANN, Nvidia GPUs remain the preferred choice for many AI developers, including companies such as Deepseek.

Huawei has made its CANN software for Ascend GPUs open source, offering multi-layer programming interfaces for AI applications.

The move aims to challenge CUDA’s nearly two-decade dominance, encouraging a domestic ecosystem that reduces reliance on American hardware.

Early adoption remains uncertain, as CANN’s ecosystem is still immature compared to the long-established CUDA platform.

With the progress of Huawei, the US allegedly reviewed multiple scenarios, ranging from full export bans to attempts to overwhelm Huawei by flooding the market.

The final decision represents a middle ground, balancing national security, global AI competitiveness, and economic interests.

President Trump emphasized that authorized exports will support American jobs and manufacturing while retaining leverage over advanced AI technology.

Nvidia’s CEO, Jensen Huang, acknowledged uncertainty over whether Chinese customers would fully purchase H200 systems, noting a $5.5 billion revenue shortfall in AI chips earlier in 2025.

The reversal, therefore, appears to be driven by Huawei’s Ascend 910C performance trajectory, which poses a potential threat to US leadership in AI hardware.

While exporting H200 chips allows the US to maintain influence over AI software ecosystems.

It also reflects the recognition of China’s growing capabilities in high-performance accelerators.

Via Tom's Hardware

Follow TechRadar on Google News and add us as a preferred source to get our expert news, reviews, and opinion in your feeds. Make sure to click the Follow button!

And of course you can also follow TechRadar on TikTok for news, reviews, unboxings in video form, and get regular updates from us on WhatsApp too.