On February 28, Hangzhou hosted an AI development summit at the Hangzhou Civic Center, aiming to establish the city as China’s premier hub for artificial intelligence (AI). Reports from East Money and Xinhua indicated that 12 AI initiatives secured funding, totaling an investment of CNY25.5 billion (US$3.71 billion).
Among these, Sunrise’s “High-Performance GPU and Inference Chip R&D Project” was the sole endeavor focused specifically on AI inference GPUs. This Zhejiang-based company committed to the project in Hangzhou, aligning with the city’s broader computing infrastructure development.
Sunrise Co-CEO Wang Zhan emphasized that future industrial progress would be driven by computing power, rather than solely by model capabilities. He also noted that Hangzhou’s AI strategy provides a favorable environment for his company’s expansion efforts.
In a published statement, Sunrise affirmed that this agreement signifies a new phase in its growth within Hangzhou, contributing to the city’s AI computing infrastructure and overall innovation framework.
Sunrise was formerly SenseTime’s large-chip division, becoming an independent entity at the close of 2024. It specializes in developing high-performance GPUs and multimodal inference chips for commercial applications. While the company claims multiple generations of validated products, assertions like “ten-thousand-card-scale delivery” stem from corporate and local announcements, lacking independent third-party corroboration.
China’s AI chip deployment expands across multiple vendors
While Sunrise’s project exemplifies local government backing for domestic AI computing providers, it represents only one facet of China’s AI chip deployment. International and domestic media outlets report that China’s strategy for AI computing supply and substitution has evolved into a diverse, multi-player ecosystem.
Huawei Ascend Series: A unified platform for training and inference
Huawei’s Ascend series, which includes the 910C, is widely recognized by major media as a key domestic platform in China for extensive model training and inference. Reports suggest Huawei is preparing for substantial shipments to address the demand for alternatives to Nvidia, despite persistent limitations in manufacturing capacity and policy-related constraints.
US officials have indicated that domestic production might remain constrained, with one projection suggesting 2025 shipments could be capped at approximately 200,000 units, underscoring ongoing supply chain and technological challenges.
Multiple domestic AI accelerator providers
Beyond Huawei, China’s AI chip supply is also significantly bolstered by various AI accelerators and processors primarily focused on inference tasks.
Cambricon is frequently cited as a prominent domestic AI accelerator supplier. Its MLU series has appeared on procurement lists for both government and enterprise clients. The Financial Times reported that both Huawei and Cambricon chips were included in official domestic procurement initiatives designed to foster indigenous technology development.
Reuters previously reported that the China Unicom data center in Xining, Qinghai, implemented a range of domestic AI chips, featuring products from Alibaba’s T-Head, MetaX, and Biren. This project demonstrates the practical application of Chinese AI processors within large-scale infrastructure.
The Reuters article also mentioned that Moore Threads and Enflame Technology are anticipated to supply chips for the subsequent phases of the data center and associated cloud initiatives.
Chinese media outlets have reported that Enflame is pursuing a STAR Market IPO, indicating continued capital market engagement alongside its product development efforts.
Credit: AFP
Domestic CPU replacement efforts expand
At the central processing unit (CPU) level, vendors such as Loongson and Phytium are positioned as domestic alternatives for government, state-owned enterprise, and industrial server environments, though their role is distinct from directly replacing AI training accelerators.
China is progressively restricting the use of Intel and AMD chips in government-funded systems. This policy aims to promote domestically developed alternatives, thereby offering significant support to vendors like Loongson and Phytium.
China’s AI chip ecosystem enters a multi-vendor phase
Sunrise’s GPU and inference chip project in Hangzhou exemplifies a collaboration between local governments and emerging chip manufacturers. However, broader deployment patterns highlighted by major global media and specific project implementations reveal the following:
● Huawei Ascend stands as one of the few domestic computing platforms with confirmed shipments and practical deployments for large-model training.
● Chips from Alibaba T-Head, MetaX, Biren, and Cambricon have been either deployed or earmarked for major data center projects, including those at China Unicom.
● Moore Threads and Enflame are emerging as additional sources of supply.
● Chinese CPUs, notably Loongson and Phytium, are advancing in server and government applications through policy-driven substitution.
China’s AI chip ecosystem is transitioning from reliance on a single supplier to a more diversified, multi-vendor framework. Nevertheless, notable disparities persist in computing scale, manufacturing capacity, and global ecosystem integration when compared to Nvidia and established international supply chains.
Article edited by Joseph Tsai