By Da Cheung
Chinese artificial intelligence company Intellifusion (688343.SS) has secured a 420 million yuan ($58 million) contract to build a landmark AI infrastructure project in the southern city of Zhanjiang. According to NFNews, the project will establish a 1,000-card AI computing cluster using the company’s proprietary AI chips, the first such cluster to utilize domestically developed accelerator cards.
The cluster is designed to host domestic AI models, including those of DeepSeek, to provide lower-cost AI capabilities for government and industrial applications. Zhanjiang is the hometown of DeepSeek’s founder Liang Wenfeng, and the city’s government has actively deployed the company’s models for local administrative tasks.
From surveillance to AI inference chips
Intellifusion, whose formal name is Shenzhen Yuntian Lifei Technology, initially built its business on visual technology hardware and software. Its earliest flagship product, Shenmu, was designed to help police surveillance cameras analyze data in real-time. The company expanded its product line into broader urban management and public transportation systems before pivoting into the highly competitive AI inference chip sector.
Today, the company is attempting to capitalize on a critical shift in the AI hardware market. AI computing generally consists of two phases: training and inference. Training is the highly resource-intensive process of feeding massive datasets into an AI model so it can learn, while inference is the application of that trained model to process new data and answer user prompts in real time.
The push for domestic autonomy
Driven by US export controls, China has been aggressively pushing for domestic autonomy in both hardware and software, a policy that has spawned various combinations of domestic chips and local software models like DeepSeek. While DeepSeek initially relied heavily on processors from Nvidia for its training phase, domestic chips — including Intellifusion’s inference neural processing units — are increasingly being used to support its inference operations.
According to NetEase Tech, tech giants such as Huawei and full-stack AI companies like Cambricon have launched dedicated AI chips to challenge Nvidia’s dominance. Cambricon has secured significant orders from DeepSeek, emerging as a heavyweight domestic supplier.
Focusing on edge computing
Unlike competitors focused on comprehensive cloud-based solutions, Intellifusion has carved out a niche in edge computing. While cloud computing processes data in centralized, remote data centers, edge computing processes data locally on or near the physical device generating the data, which reduces latency. Because of this niche focus, Intellifusion’s brand awareness and market capitalization still lag significantly behind full-stack companies like Cambricon, even after a recent surge of more than 250% in its share price.
The company, which was founded in 2014, remains unprofitable. Its unaudited 2025 earnings preview forecast a net loss of 402.44 million yuan ($55.8 million), though that was a 30.5% improvement on 2024.
In terms of hardware capability, the first phase of the Zhanjiang cluster will deploy Intellifusion’s X6000 accelerator cards. The X6000’s computing power is roughly in the same tier as Nvidia’s A10 chip, which sits just below the threshold of U.S. export bans but remains difficult to procure officially in China. While the X6000’s performance trails Nvidia’s flagship A100 and H100 processors by a significant margin, it is specifically optimized for the “cost-per-token” metrics that local governments prioritize.
Ambitious plans in a growing market
Chinese manufacturers, including Intellifusion, have significant ambitions for future development. The company says the planned second and third phases of the Zhanjiang computing cluster will be equipped with its next-generation chips.
CEO Chen Ning told the 21st Century Business Herald in a recent interview that the AI industry is at a turning point where the focus is shifting from model training to large-scale, cost-effective inference applications. He predicted that in 2026, the global market for inference chips will match or even exceed that of training chips, providing a unique window for Chinese firms to overtake their competitors by offering extreme cost-efficiency. His goal is to reduce the cost of processing one million AI tokens by one million times by the end of China’s 15th Five-Year Plan in 2030.
Sources:
