Jiliu Technology has completed nearly 100 million yuan in Series A+ financing, co-led by Shanghai State Investment Futeng Capital and Guofang Innovation.

36Kr has learned that AI infrastructure provider "Jiliu Technology" recently completed a nearly 100 million yuan Series A+ round of financing, co-led by Shanghai Guotou Futeng Capital and Guofang Innovation, with Karamay Urban Development Fund and Zhangjiang Suixin Fund participating in the investment. Existing shareholder Zhuoyuan Asia made additional investments for the fourth consecutive round. The funds raised in this round will be mainly used for core technology R & D, market expansion, and team building.

Jiliu Technology was founded in February 2023 and originated from the Cybersecurity Laboratory of Tsinghua University. Its founder, Hu Xiaohe, is a Ph.D. and postdoctoral fellow from Tsinghua University, with over a decade of research experience in the fields of high - performance networks and distributed systems. The core team members come from top universities such as Tsinghua and Peking University, as well as leading companies like Alibaba, Baidu, and ZTE.

Jiliu Technology positions itself as a full - stack autonomous AI computing power builder. The team focuses on solving the distributed computing and communication problems of large - scale computing power clusters. As one of the few AI infrastructure providers in China with experience in deploying several ten - thousand - GPU clusters, the company has served multiple users, including Zhipu AI, SenseTime, operators, data centers, and local state - owned enterprises.

"Different from the 'real estate developers' who focus on computing power platform operation and the 'decorators' who focus on system optimization, we focus on building a high - performance, open - source, and autonomous intelligent computing system covering both computing and communication software and hardware, similar to 'building a house'," Xie Wenqi, co - founder and CFO of Jiliu Technology, told 36Kr.

With the explosion of large models and AIGC applications, computing power has become a new type of infrastructure after water, electricity, and the Internet. Subsequently, there are problems such as network communication bottlenecks, high failure rates, and high costs caused by the software - hardware binding of suppliers. NVIDIA, relying on its CUDA ecosystem, not only sells GPUs but also bundles its dedicated network devices, forming a closed and expensive solution.

"Our core advantage lies in cost - effectiveness and openness, and the essence is to achieve 'decoupling'," Hu Xiaohe, founder of Jiliu Technology, once pointed out. Through Jiliu Technology's self - developed RoCE networking solution, customers' costs can be reduced by 40%, and the delivery cycle can be shortened from several months to several weeks. In the second half of 2023, Jiliu Technology successfully deployed the earliest third - party thousand - GPU H800 RoCE cluster in China, strongly supporting the iteration of the base models of domestic large - model companies.

Currently, Jiliu Technology has developed its self - developed high - performance open - source intelligent computing system, Galaxy HPAC (High - Performance AI Computing). The system mainly includes three core products: the high - performance converged AI network (Mercury - X), the artificial intelligence platform (Venus - AICloud), and the computing power construction and maintenance platform (Venus - AIDOC). In addition, the company's self - developed MS6426 all - domestic - chip 25.6T high - speed open - source AI switch has passed the 72 - hour long - term stability test of the intelligent computing cluster and has been deployed on a large scale.

Xie Wenqi introduced that after more than two years of development, the company has cumulatively optimized and delivered 23 clusters, with over 66,000 GPUs, more than 4,000 switches, and over 320,000 optical modules. At the same time, it has successfully deployed several long - distance training and inference clusters (50 km, 100 km, 1500 km).

Regarding the question of whether there is an oversupply of computing power, Xie Wenqi told 36Kr that computing power is not only for pre - training. Currently, large - model companies both at home and abroad are still conducting pre - training, and the computing power demand for inference and post - training is growing rapidly. But in terms of demand proportion, he mentioned that the ratio of computing power demand for training and inference was about 7:3 in 2023, 5:5 in 2024, and may reach 3:7 in 2025. While based in core regions such as Beijing and Shanghai, Jiliu Technology also actively responds to national strategies and participates in the implementation of computing power projects in Xinjiang, Ningxia, and other places.

Currently, Jiliu Technology has invested in the simulation and future architecture design of a one - hundred - thousand - GPU cluster. Hu Xiaohe admitted that although China is still in the exploration stage of how to build and use a one - hundred - thousand - GPU cluster, technology must take the lead. The company is collaborating with more industry partners to promote the domestic substitution of hardware and the opening and maturity of the AI infrastructure ecosystem.