Approaching.AI Completes a Pre-A Round Worth Hundreds of Millions of RMB, Strengthening High-Quality AI Token Production Infrastructure
Daily ATaaS call volume is approaching 1 trillion
Approaching.AI, an AI Token production service company, has announced the completion of a Pre-A financing round worth hundreds of millions of RMB. This round was co-led by Starlink Capital and Hua Kong Technology, with participation from Honghui Capital, Tianhao Energy, Shangshi Capital, Tianjin Renai Hongsheng, Hangzhou Fucheng, and others. Existing shareholder GL Ventures also made additional investment.
Following this financing, Approaching.AI will further increase its investment in “Approaching AI Token as a Service (ATaaS),” its high-performance AI Token production service platform, with a particular focus on securing computing resources and building its underlying inference system. This will enable the continuous delivery of model output capabilities that combine low latency, high throughput, stable structured output, reliable function calling, and predictable service quality, further enhancing the company’s large-scale supply capacity for high-quality Tokens in enterprise production environments.
Approaching.AI: A Leading Domestic Provider of High-Quality Tokens, with ATaaS Daily Call Volume Nearing 1 Trillion
As large models continue moving into enterprise production environments, the evaluation criteria for AI inference infrastructure are also changing. Enterprises now care not only about the scale of compute resources, the number of models, or the breadth of APIs, but also about whether each call can reliably, efficiently, and predictably complete the intended task. At this stage, the core competitiveness of inference services is shifting from “providing models” to “producing high-quality Tokens.” Time to first token, tokens per second, stability of structured output, reliability of function calling, and service predictability under high-concurrency workloads have become key indicators for enterprises selecting AI infrastructure.
Approaching.AI believes that Tokens are no longer just the basic unit of input and output for large models, but a critical production factor connecting model performance, system performance, service stability, and cost efficiency. Based on this view, the company introduced the industrial concept of Token as a Service (TaaS) and built ATaaS, a high-performance AI Token production service platform. Unlike traditional MaaS, which focuses on model invocation and management, ATaaS is more focused on delivering inference efficiency in enterprise production scenarios, helping enterprises gain scalable and operable high-quality Token production capabilities.
In its model strategy, Approaching.AI has consistently followed a “fewer models, deeper optimization” approach. Rather than blindly supporting hundreds of models, the company concentrates on a smaller number of high-productivity models and continuously improves output quality, inference efficiency, TTFT stability, and TPS performance based on real enterprise use cases. For enterprise customers, having more models does not automatically translate into productivity; what truly matters is whether each call can reliably support business outcomes.
On the systems side, Approaching.AI transforms underlying compute resources into sustainable, high-quality AI Token production capacity through capabilities such as heterogeneous compute scheduling, cross-cluster cache sharing, inference path isolation, elastic scale-up and scale-down, and quality monitoring. With end-to-end systems engineering capabilities, the company can deliver more stable TTFT, high-speed output of 30–50 TPS, and reliable service guarantees to enterprises while keeping costs under control.
Today, through the ATaaS platform, Approaching.AI provides services to multiple enterprise customers, including Zhipu GLM and Moonshot Kimi, with daily Token processing volume reaching nearly 1 trillion. After long-term validation in highly complex, high-concurrency scenarios, the company has established its core capabilities for large-scale inference delivery.
A Unique Team Structure, With Business Execution and Technical Foundations Driving ATaaS Growth
TaaS is not merely an application-layer product, but a systemic capability spanning the entire AI inference pipeline. As a result, beyond understanding enterprise customer needs, industry resources, capital flows, and the pace of commercialization, it also requires long-term accumulation in foundational architecture areas such as compute, storage, scheduling, caching, and inference systems. Approaching.AI’s core team combines business execution with deep technical research, laying the foundation for ATaaS to move from an advanced concept to large-scale deployment with major customers within just two years of founding.
On the business and operations side, Approaching.AI has developed organizational capabilities that integrate technology productization with business capitalization. Founder and CEO Aizhi Yuan holds a PhD in Computer Science from Tsinghua University and combines systems research expertise with commercialization experience at major tech companies. He proposed the industrial logic of TaaS and has driven ATaaS from a technical platform into an enterprise production service. President Dr. Wujing Wu holds a doctorate in finance and the CFA credential, has management experience in major industry players and investment institutions, and has led investment and M&A projects for several representative companies. She currently oversees the company’s strategy, internal controls, and global operations. Chairman Ren Xuyang was an early founding member of Baidu and has led the launch of iQiyi, Yidian Zixun, Haizhi, News Break, and others. He supports the company’s growth through industry insight, organizational development, capital alignment, and ecosystem resource integration.
On the technical and research side, Approaching.AI is backed by more than 20 years of technical accumulation from Tsinghua University’s Institute of High Performance Computing, and it has completed the equity participation process based on the contribution valuation of Tsinghua-related technological成果. These成果 were developed over many years by research teams led by Academician Zheng Weimin, Professor Wu Yongwei, and Associate Professor Zhang Mingxing, covering key areas such as high-performance computing, parallel and distributed systems, storage systems, intelligent computing resource systems, and large-model inference infrastructure. The injection of these成果 marks the transition of collaboration between the company and Tsinghua’s research teams in AI infrastructure from concept to substantive execution.
Among them, Academician Zheng Weimin, Approaching.AI’s Chief Science Advisor, laid the academic foundation for Tsinghua’s high-performance computing field. Chief Scientist Professor Wu Yongwei has spent many years researching distributed systems and storage systems and has received multiple national-level science and technology awards. Associate Professor Zhang Mingxing focuses on large-model inference architecture, and open-source projects he leads, such as KTransformers and Mooncake, are widely used across the industry. With the contribution-based participation of Tsinghua’s core technological成果 and the ongoing support of top research teams, Approaching.AI has built a strong barrier to entry in systems engineering and research implementation for AI inference infrastructure.
This technical accumulation is also validated by the open-source ecosystem. KTransformers, open-sourced under the leadership of Approaching.AI and the Tsinghua team, is the world’s first edge heterogeneous inference framework. Its GitHub Stars have surpassed 17k, and it has become the first-recommended inference engine for top large models such as GLM, Kimi, Minimax, and Qwen. In distributed inference, Approaching.AI is co-developing the open-source project Mooncake together with Tsinghua University, Moonshot Kimi, 9#AISoft, Alibaba Cloud, Ant Group, and other industry-academia-research institutions. Yang Ke, a technical expert at Approaching.AI and PhD from Tsinghua University, is also deeply involved as a core contributor in many key technical implementations and architectural efforts. In addition, Approaching.AI actively contributes to the global inference community, including SGLang, vLLM, and NVIDIA Dynamo, continuously promoting the development of an open AI inference infrastructure ecosystem.
With the business team accurately capturing industry needs, customer scenarios, and capital flows, and the technical team building on years of accumulation in high-performance computing, distributed systems, and large-model inference infrastructure, Approaching.AI has the full-stack capability to cover everything from foundational system R&D to large-scale enterprise delivery. As ATaaS continues to evolve, this hybrid team structure will further support the company’s ability to mass-produce and supply high-quality Tokens.
Huakong Fund Chairman Zhang Yang said:
As domestic large-model capabilities rapidly improve and application demand expands across the board, the huge demand for Tokens is reshaping the compute resource industry chain. Huakong Fund believes that the AI Infra industry, which can provide high-quality Tokens at scale and with stability, will become a critical foundational infrastructure supporting the strong development of the AI industry, with broad market potential and extremely high investment value. The Approaching.AI team is rooted in Tsinghua University’s Institute of High Performance Computing, possesses deep research foundations and solid technical strength, and has succeeded in breaking through the silos of underlying compute resources, earning high recognition from both upstream and downstream ecosystems. Huakong Fund will leverage its rich AI industry ecosystem resources to fully support Approaching.AI’s future fundraising and listing process, and will accompany the company as it grows into a world-leading next-generation intelligent computing “Token factory.”
Starlink Capital partner Li Wenjue said:
Approaching.AI has demonstrated exceptional technical depth and engineering capability in the AI infrastructure space, especially in Token production efficiency, where it leads at a world-class level. In this investment, we highly value the system-level breakthroughs achieved by the company’s ATaaS platform and its ability to rapidly translate cutting-edge academic成果 into large-scale commercial deployment. As AI Agents become more widespread and Token demand surges, companies that can efficiently convert compute resources into intelligent output will become the core of industrial competition. Starlink Capital believes that whether next-generation AI infrastructure can coordinate and optimize control plane, scheduling, runtime, memory management, and model architecture will be key to seizing leadership in the next wave. Approaching.AI combines top-tier Tsinghua technical roots with rich commercialization experience, and investors have high expectations for its continued growth and market leadership in AI Infra. We strongly support the company in building the industry benchmark for efficient AI Token production.
Source: Approaching.AI
This article is reprinted with permission from 量子位, and the views expressed belong solely to the original author.