Huawei Accelerates AI Chip Roadmap as It Challenges Nvidia in Chinese Market

Key Highlights

• Domestic AI chip demand far exceeds Huawei’s production capacity, restricting international expansion
• Testing underway for Ascend 950DT chip, with Chinese AI developers set to begin model training in 2025
• Next-generation Ascend 960DT processor accelerated to Q1 2027 release, nine months earlier than initially planned
• Peerium computing architecture introduced to enable integration of up to 1 million processors in unified systems
• Company asserts Ascend processors have captured greater market share than Nvidia within China

At its annual Huawei Connect conference held in Shanghai on Thursday, Huawei disclosed that production capacity for its AI processors falls significantly short of domestic market requirements, effectively eliminating any substantial international sales expansion plans. The company’s rotating chairman Eric Xu delivered these remarks during the event.

While Xu acknowledged that Huawei does serve select international markets with strong demand, he emphasized that shipment volumes remain minimal. The company’s immediate priority centers on satisfying China’s domestic requirements.

Next-Generation Ascend 960DT Timeline Moved Forward

The release date for Huawei’s Ascend 960DT processor has been accelerated to the first quarter of 2027, representing a nine-month advancement from the initial timeline. The subsequent Ascend 960PR model will arrive in Q3 2027, also one quarter ahead of the original plan.

The corporation intends to maintain an annual release cycle for new Ascend processor generations. Development roadmaps indicate the Ascend 970 and 980 are scheduled for 2028 and 2029 respectively.

During the conference, Huawei also presented the Atlas 960 SuperPoD computing cluster. This enhanced system builds upon the Atlas 950 SuperPod introduced just months earlier, delivering superior performance for both AI training workloads and inference tasks.

According to Xu, the Ascend 950DT processor has demonstrated strong performance during testing phases. He anticipates that Chinese artificial intelligence model developers will commence training operations on systems powered by this chip throughout the coming year. Huawei’s computing infrastructure has already supported the training of more than 40 AI models, while the Ascend chip software ecosystem maintains over 5,200 active developers each month.

Strategic Push Toward Complete Chip Independence

Xu spoke candidly about the motivations driving Huawei’s aggressive chip development strategy. Since 2019, US export regulations have curtailed China’s ability to procure Nvidia’s cutting-edge processors. Huawei itself has operated under US trade sanctions since May of that same year.

“We cannot accept a destiny where we cannot control our fate being determined by others in terms of willingness to sell chips to China or not,” Xu stated.

Beyond developing more powerful individual chips, Huawei is pursuing strategies to circumvent performance limitations by networking vast numbers of processors together. Current large-scale AI model training typically requires clusters containing approximately 100,000 chips.

The company introduced Peerium, a novel computing architecture engineered to ultimately support interconnection of up to 1 million processors functioning as a cohesive system. Its UnifiedBus technology facilitates integration of processors, memory, storage infrastructure and networking components within these massive systems.

According to Huawei, the Ascend 960 supernode configuration supports connectivity for up to 4,096 processors. These supernodes can subsequently be combined to construct even larger computational clusters.

Over 1,000 systems utilizing the previous-generation Ascend 910C processors have been deployed already. Systems based on Ascend 950 chips have transitioned into commercial production environments.

Nvidia continues to maintain a substantial competitive advantage through its CUDA software platform, which has become the industry standard for developers building and deploying AI applications. Nvidia declined to provide comment when contacted.

These announcements from Huawei arrived less than two weeks prior to a scheduled meeting between US President Donald Trump and Chinese President Xi Jinping in Washington on September 24.