Approaching.AI Completes Over 1.3 Billion Yuan Series B Funding; Average Daily AI Token Output Grows Over 100-Fold in Half a Year

Recently, Leiphone learned that Approaching.AI, a high-quality AI Token production service provider, has completed a Series B funding round of over 1.3 billion RMB; the company has raised a cumulative total of 2.4 billion RMB so far this year. Multiple core operating metrics have achieved more than tenfold growth within half a year, earning sustained high recognition from both industry and capital.

This round was jointly led by Zhejiang Caixin Innovation, Xiaoshan Capital, ICBC Investment, Zhongguancun Science City, Xinglian Capital, and Qianjiang Century City. YunChuang Capital, Hunan Gaoxin Zongheng, Jiantou Huake, Shenwan Innovation Investment, CITIC Juxin, Zhuhai Technology Industry Group, SenseTime, Wangsu Science & Technology, Gansu Shuzhi, Changjinkong, Yuanzhi Future, Qingzheng Xingzhu, Xiaoka Capital, and others participated, while existing shareholder Shangshi Capital continued to increase its investment.

English text is being prepared.

The funds raised in this round will be focused on expanding high-quality AI Token production capacity, upgrading and iterating the ATaaS (Approaching.AI Token as a Service) platform, R&D of underlying inference technologies and new products, and improving commercial cooperation and delivery systems.

Investor Says · Chen Dong, Chairman of Zhejiang Caixin Innovation

"As the AI industry moves toward large-scale application, inference efficiency, cost, and system optimization capabilities are becoming an important foundation for commercialization. Approaching.AI has sustained technological accumulation and engineering delivery capabilities in the field of large-model inference. We look forward to walking alongside the company for the long term and jointly promoting the construction and large-scale application of efficient, inclusive AI infrastructure."

Investor Says · Zhu Lifeng, Chairman of Xiaoshan Capital Group

"Approaching.AI is a benchmark project in Xiaoshan's AI Infra layout. As early as when the industry was focused on training, the company was already deeply cultivating the inference track with foresight. Relying on the Tsinghua-affiliated team's more than ten years of technological accumulation and its self-developed ATaaS platform, it efficiently converts computing power into high-quality AI Tokens that can be produced at scale, driving domestic computing power into the core production chain. The two sides are working together to advance the construction of a ten-thousand-card-level AI Token factory and accelerate the formation of an industrial closed loop of 'chip-model linkage.' Xiaoshan Capital will provide deep empowerment with long-term capital and patient capital, supporting the company in rapidly replicating its high-quality AI Token factory capabilities and injecting lasting momentum into the high-quality development of the AI industry."

Investor Says · Zhang Mingchen, Founding Managing Partner of Xinglian Capital

"Since becoming a shareholder in the angel round in 2024, Xinglian Capital has witnessed Approaching.AI continuously creating results far beyond expectations across technology, products, services, strategy, and commercialization; in less than three years, it has rapidly risen to become a dazzling first-class enterprise in the AI inference field. This is also the fundamental reason why Xinglian Capital has continued to make large additional investments in every round. We are optimistic about Approaching.AI's broad growth space as it moves from technological and commercial validation to large-scale construction and sustained operations, and we look forward to accompanying the team as it continues to lead the development of the AI inference industry."

From 'Extensive Supply' to 'Quality Service'

| High-Quality AI Tokens Become a New Industry Trend

In the early stage of the Token market's development, the supply side of the industry mostly 'emphasized Token volume and neglected service quality,' rapidly expanding supply scale through extensive volume expansion. At that time, large models were still in the pilot validation stage and had not yet been deeply embedded in enterprises' core production chains, so business had greater tolerance for performance, stability, and model quality, and this supply model was accepted by the market. Since its founding, however, Approaching.AI has adhered to anchoring production-grade inference standards, continued to invest in original innovation in inference systems, consolidated its engineering foundation, and taken the lead in defining the concept of high-quality AI Token services in the industry.

The 'high-quality AI Token' defined by Approaching.AI: anchored in production-grade scenarios, providing AI Token services with high intelligence, low latency, high speed, strong stability, and both large-scale delivery capability and controllable overall cost.

As the industry's focus shifts to Agent-driven business closed-loop implementation, large models are gradually carrying production-grade core processes, and the AI inference industry is ushering in a paradigm shift: enterprises are imposing strict production-grade rigid requirements on the intelligence level, response latency, output speed, context completeness, and service stability of AI Tokens. This shift coincides with the high-quality AI Token service concept that Approaching.AI has always adhered to and advocated.

With the industry inflection point emerging, the Token market is rapidly forming a clear quality stratification: the extensive supply of 'usable is enough' can no longer match the demands of production scenarios, while high-quality AI Token supply that balances multiple dimensions of metrics is scarce and in surging demand, becoming a rigid market need. As the earliest layout player and deep cultivator in high-quality AI Token services, Approaching.AI, with its long-accumulated underlying inference system technologies and engineering capabilities, continues to advance the large-scale delivery of high-quality AI Token services. Approaching.AI has now become a leader in the field of high-quality AI Token services.

Anchored in "fewer models, deeper optimization, high quality"

｜Accelerating the large-scale delivery of high-quality AI Tokens

The sustained, large-scale delivery of high-quality AI Tokens ultimately depends on two core propositions: whether per-unit compute efficiency can continue to improve, and whether large-scale supply capacity can grow steadily.

With this goal in mind, and to fulfill the production standards for high-quality AI Tokens, Approaching adheres to the technical route of "fewer models, deeper optimization, high quality," focusing on mainstream large models with clear market demand and broad application, and continuously advancing the engineering implementation of a number of globally pioneering technologies such as "heterogeneous inference," "trading storage for compute," and "virtual-real isomorphism," concentrating on the high-quality AI Token needs of enterprise production scenarios.

"Fewer models, deeper optimization" addresses the question of "how many high-quality AI Tokens the same batch of compute can produce," while stable production loads and cross-cluster operations address the question of "how much time this compute is actually in an effective working state." "Fewer models" by itself does not inherently bring higher utilization; only when per-unit compute output and effective utilization rate improve simultaneously can the unit economics of the AI Token business become viable.

The business logic behind this route is very clear: enterprise customers ultimately pay for business outcomes, not for the number of compatible models.

At present, Approaching Technology has already formed a replicable model for building and operating high-quality AI Token factories: there are projects already in production with daily Token capacity in the trillions, as well as strategic cooperation and under-construction layouts aimed at regional industrial ecosystems and leading customers. More importantly, the overall business has already crossed the cost line. Real performance has verified the ATaaS platform's large-scale delivery and continuous operation capabilities.

Technological innovation restructures the domestic compute system

｜Cost-effectiveness surpasses the world's most advanced compute

Enabling domestic compute to support the production of high-quality AI Tokens has always been Approaching Technology's goal.

Faced with performance differences between domestic and foreign chips, Approaching Technology has explored a new path to breakthrough compute efficiency through system-level technological innovation, having taken the lead in proposing a series of innovative technical solutions such as "domestic Prefill-Decode (PD) heterogeneous collaboration," "high-performance heterogeneous KV Cache conversion," and "heterogeneous compute fusion," restructuring the domestic compute computation model and inference system, which have already been officially put into production in high-standard production scenarios.

In a real production-grade environment, it has achieved domestic chips carrying trillion-level leading large models and large-scale production of high-quality AI Tokens, while overall cost-effectiveness surpasses the world's most advanced compute.

Approaching's pioneering heterogeneous collaborative inference technology, while ensuring model output quality, response speed, and service stability, achieves higher effective Token output with lower hardware investment, steadily serving the real business needs of leading customers.

This breakthrough has dual pioneering value: at the commercial level, it builds a differentiated core technology moat, opens up the growth ceiling for Approaching's large-scale co-building of high-quality AI Token factories, and supports the continued leap in performance and capital value; at the industrial strategic level, it helps break away from technical competition that merely compares hardware parameters, redefines the value yardstick of domestic compute, enables domestic chip assets scattered across the country and of different architectures to release full inference capacity, and provides a deployable core engineering solution for the country to build an autonomous, controllable, abundantly supplied, and commercially economical AI inference foundation, further consolidating the security of China's large model industry chain and the foundation of compute autonomy.

Expanding output through technology

｜Achieving a tenfold leap in performance in half a year

Continuous independent technological innovation is the core driving force behind the exponential growth of Approaching Technology's high-quality AI Token output.

Over the past six months, with the continuous introduction and engineering refinement of multiple pioneering technologies, the company's average AI Token production efficiency per unit of compute has climbed month by month, continuing to lead the industry; average daily output of high-quality AI Tokens has grown more than a hundredfold, climbing to the trillion-plus level, and business revenue for the year has reached nearly ten times that of the entire previous year.

In terms of the industrial landscape, Approaching Technology is driving the upgrade of AI infrastructure from traditional compute resource supply to large-scale production and operation of high-quality AI Tokens. Relying on the ATaaS platform and its Token factory construction and operation capabilities, the company has, within half a year, worked with industry partners to advance more than ten AI Token factory projects with daily output in the trillions, promoting the large-scale application of domestic compute, and joining hands with local state-owned capital, industry leaders, and ecosystem partners to build a new AI Token supply network connecting compute infrastructure, large models, and industry applications.

Regional infrastructure layout

The company focuses on Henan, Hangzhou, and Hainan, combining regional resources and industrial needs to advance the construction of a high-quality AI Token production and service network. In Henan, it is advancing joint operations with local state-owned capital, collaboratively carrying out Token production, sales, and customer consumption; the East China regional headquarters has landed in Hangzhou Qianjiang Century City, with plans to build a ten-thousand-card-level Token factory to serve enterprise-level inference needs in the Yangtze River Delta; in Hainan, it is exploring a cooperation model in which local state-owned capital builds and Approaching provides the platform and operational services. Through construction and operation models tailored to local conditions, it drives the conversion of compute investment into sustainable Token supply capacity, driving the development of local operational capabilities and the industrial ecosystem.

Industrial ecosystem and co-building and co-operation

The company has established ecosystem partnerships with SenseTime, Paratera, Hongxin Electronics, Wangsu Science & Technology, Moore Threads, MetaX, Tianshu Zhixin, and other enterprises to advance the construction and operation of high-quality AI Token factories, explore high-quality AI Token production technology solutions based on domestic chips, and jointly build an autonomous and controllable domestic AI compute ecosystem. Together with SenseTime, it is advancing the construction of a Token factory with a daily average of ten trillion tokens, and the co-built production line already has the capacity to produce and stably absorb a daily average of one trillion tokens; jointly with Paratera, it has built a Token factory with a daily average of one trillion tokens, connecting the entire chain from compute infrastructure, inference optimization, and standardized Token supply to industrial applications; with Hongxin Electronics, it has completed technical adaptation and pilot implementation at its two Token factories, Suihong Huachuang Wuxi and Xiamen; and together with Wangsu Science & Technology, leveraging its global distributed edge resources and network capabilities, it is carrying out collaborative inference optimization and intelligent scheduling to improve the production efficiency and delivery stability of high-quality AI Tokens.

The continuously released dividends of technological efficiency are driving the business and capital into a high-speed growth cycle in tandem. The financing pace is in sync with performance growth, with multiple rounds of closings landing in an orderly manner during the year, maintaining an average pace of one round every two months, and the scale of a single round of fundraising continuing to rise quarter over quarter; the cumulative total funds raised during the year reached 2.4 billion RMB, with continued heavy positions taken by state-owned capital, industrial capital, and market-oriented institutions. Following the completion of this series of financings, the shareholder structure has become further diversified and mature, bringing together local state-owned capital, financial capital, leading industrial groups, and first-tier market-oriented institutions, providing solid support for subsequent regional co-building of compute, collaboration across the domestic heterogeneous ecosystem, and large-scale industry delivery.

Top-tier scientific research and open-source foundation

| Building system-level long-term engineering barriers

Behind the commercialization progress is Approaching.AI's sustained accumulation in scientific research, open-source engineering, and enterprise-level production and delivery.

From cutting-edge technological exploration to rigorous engineering implementation, what runs throughout is a core team combining academic depth and industry experience: the company's management team consists of Founder and CEO Dr. Ai Zhiyuan, President and CFO Dr. Wu Wenjie, CTO Chen Xianglin, and others, who jointly advance technology R&D, capital coordination, industrial cooperation, and project delivery; at the technical research level, Academician Zheng Weimin serves as Chief Scientific Advisor, and Professor Wu Yongwei serves as Chief Scientist, providing academic and technical support for the company's underlying technology evolution; Associate Professor Zhang Mingxing serves as a director, guiding technical strategy and key R&D work and promoting the connection between frontier research and engineering practice.

Beyond academic research, the core team of Approaching.AI has long participated in building the open-source ecosystem for large-model inference. The team led the initiation of KTransformers and participated in the initiation and co-building of Mooncake, providing an engineering validation environment for related research oriented to different models, hardware, and load conditions, and also driving the team to continue exploring directions such as "full-system heterogeneous collaboration," "trading storage for compute," and "virtual-real isomorphism."

Starting from academic research and open-source engineering, the relevant technical capabilities have been further consolidated into the ATaaS platform and continuously tested under real business loads in AI Token factories. Upholding the philosophy of "more efficient compute, more inclusive intelligence," Approaching.AI uses real demand to drive production investment and long-term delivery to test technical capabilities. Through the ATaaS platform and AI Token factories, it is outputting a standardized production system for high-quality AI Tokens, reshaping the AI inference supply paradigm, and empowering production-grade intelligent upgrading across industries.

Ai Zhiyuan, CEO of Approaching.AI, said: In the race to transform compute into high-quality productivity, the truly scarce barrier is not scale, but unit energy efficiency and commercial delivery capability. Over the past six months, Approaching.AI has proven the enormous commercial potential of the "fewer models, deeper optimization, higher quality" route with business growth of more than tenfold. The continued capital injection is an affirmation of Approaching.AI's business delivery results. Looking ahead, Approaching.AI will rely on this round of funding to accelerate capacity expansion, deepen the implementation of domestic PD heterogeneous solutions, support customers' large-scale applications with more stable, more efficient, and more cost-effective high-quality AI Token services, promote the fuller entry of domestic heterogeneous resources into real production scenarios, and work with industry partners to jointly build a diversified, stable, and sustainable AI inference supply system.