Approaching.AI Raises Over RMB1.3 Billion in Series B to Scale High-Quality AI Token Delivery

The Series B was co-led by Zhecai Innovation, Xiaoshan Capital, ICBC Investment, Zhongguancun Science City, Xinglian Capital and Qianjiang Century City. Yuncheng Capital, Hunan Gaoxin Zongheng, Jiantou Huake, Shenwan Innovation Investment, CITIC Juxin, Zhuhai Science and Technology Industry Group, SenseTime, Wangsu Science & Technology, Gansu Shuzhi, Changzhou Financial Holdings, Yuanzhi Future, Qingzheng Xingzhu and Xiaoka Capital followed, while existing shareholder Shangshi Capital increased its investment. The round follows an A round in July. Approaching.AI has completed several rounds in 2026 at an average pace of one every two months, with each round larger than the previous one, according to Leiphone.

The company said several core operating indicators grew more than tenfold in the past six months. Average AI token production efficiency per unit of compute rose month by month, daily output of high-quality AI tokens increased more than 100-fold to more than 1 trillion, and business revenue so far this year reached nearly ten times its full-year revenue last year. It said its overall business has crossed the cost line.

Approaching.AI defines high-quality AI tokens as services for production-grade scenarios that combine high intelligence, low latency, high speed, strong stability, large-scale delivery and controllable overall cost. The company argues that as AI agents move into core business processes, enterprises are demanding stricter performance and reliability from inference services, making coarse token supply insufficient. It follows a 'fewer models, deep optimization, high quality' route, focusing on widely used mainstream models and engineering technologies including heterogeneous inference, 'storage-for-compute' and 'virtual-real isomorphism.'

The company said it has built a replicable model for building and operating AI token factories, with projects already in production that have daily capacities of trillions of tokens and additional projects under construction for regional industrial ecosystems and major customers. In the past six months, it has worked with partners on more than 10 projects with daily trillion-token capacity. The company said it has also developed heterogeneous inference technologies for domestic chips, including prefilling-decoding (PD) disaggregation, high-performance heterogeneous KV cache conversion and heterogeneous compute fusion. It said these have been put into production and allow domestic chips to run trillion-level leading models while producing high-quality tokens at scale, with overall cost-performance exceeding that of the most advanced international compute.

Approaching.AI is building production and service networks in Henan, Hangzhou and Hainan. In Henan it is working with local state-owned capital on joint operation, token production, sales and customer offtake; in Hangzhou's Qianjiang Century City it plans a 10,000-card-level token factory to serve Yangtze River Delta enterprise inference demand; and in Hainan it is exploring a model in which local state-owned capital builds facilities and Approaching.AI provides platform and operating services.

The company has formed partnerships with SenseTime, Parallel Technology, Hongxin Electronics, Wangsu, Moore Threads, MetaX and Iluvatar CoreX, among others, to build and operate high-quality AI token factories and explore production technology based on domestic chips. It is working with SenseTime on a token factory with daily capacity on the order of 10 trillion tokens, with a jointly built line already capable of producing and absorbing trillions of tokens per day. It has also worked with Parallel Technology on a daily trillion-token factory, completed technology adaptation and pilot deployment at Hongxin Electronics' token factories in Wuxi and Xiamen, and is cooperating with Wangsu on inference optimization and intelligent scheduling using its distributed edge resources.

Approaching.AI's management team includes founder and CEO Ai Zhiyuan, president and CFO Wu Wenjie, and CTO Chen Xianglin. Zheng Weimin serves as chief scientific adviser, Wu Yongwei as chief scientist and Zhang Mingxing as a director. The core team initiated the open-source project KTransformers and participated in the founding and co-building of Mooncake, according to the company.

Ai said the scarce barrier in converting compute into high-quality productivity is not scale but per-unit energy efficiency and commercial delivery capability. He said the company will use the new funds to expand capacity, deepen domestic PD heterogeneous solutions and support customers' large-scale use of AI token services.

Editor's Summary

Approaching.AI has raised more than RMB1.3 billion in a Series B round and RMB2.4 billion in 2026, with backing from local government funds, financial institutions and industrial partners. The company says its high-quality token output and revenue have grown sharply and that its heterogeneous inference technology can improve the cost-performance of domestic chips. The funds are earmarked for capacity expansion, platform development and commercial delivery.