Author | Li Shuiqing, Editor | Mo Ying
ZhiDongXi learned on October 12 that Infinigence AI (Infini-AI), a leading domestic AI-native infrastructure company, has recently completed its Series C funding round, raising CNY 2.3 billion and bringing its cumulative funding to CNY 4.3 billion. Both the single-round and cumulative funding amounts have set a new industry high for domestic AI Infra (AI Infrastructure). According to foreign media reports, the company is preparing for an IPO. Since its establishment in 2023, this startup, which originated from the Department of Electronic Engineering at Tsinghua University, has stepped into the spotlight of the capital market. Behind it are both state-owned institutions and industrial investors and market-oriented financial investment institutions betting on the AI infrastructure track.
Looking overseas, the valuations of AI infrastructure companies are also climbing rapidly. The latest valuation of Fireworks AI has reached USD 17.5 billion, Baseten has reached USD 13 billion, and Together AI has reached USD 8.3 billion. NVIDIA's investments in these three companies further highlight the computing power industry chain's emphasis on AI infrastructure.
Currently, the global AI industry competition is extending from computing power and models to Token production efficiency and application productivity. How to transform diverse computing power into stable, efficient, and low-cost Token services is becoming a significant business. Behind the CNY 2.3 billion funding, what exactly do capitalists see in domestic Token factories? Is Infini-AI's competitiveness just riding the market trend, or has it already built a sufficiently deep technological barrier? Let us explore this in detail.
01.Starting from the Department of Electronic Engineering at Tsinghua University Tackling the Hard Nut of Heterogeneous Computing Power
Many people know that numerous companies in China's AI large model field originated from Tsinghua University, with leading enterprises such as Zhipu, Yuezhi Anmian, and Shengshu Keji being representatives. The founding of Infini-AI is also deeply rooted in the academic lineage of the Department of Electronic Engineering at Tsinghua University. The company's initiator, Wang Yu, is a tenured professor in the Department of Electronic Engineering at Tsinghua University, the former head of the department, and an IEEE Fellow; the co-founder and CEO, Xia Lixue, received both bachelor's and doctoral degrees from the Department of Electronic Engineering at Tsinghua University. He has long been engaged in research on AI system design methodologies and previously worked on the development and management of user growth products at Alibaba Cloud, with experience spanning from underlying system research to industrial product practice.
Another co-founder and chief scientist, Dai Guohao, also graduated from the Department of Electronic Engineering at Tsinghua University. He is currently an associate professor at Shanghai Jiao Tong University and has won multiple Best Paper Awards at international academic conferences; the CTO, Li Boxun, holds bachelor's and master's degrees from the Department of Electronic Engineering at Tsinghua University, and formerly served as the Algorithm Director at Megvii Technology, deeply engaged in AI algorithms and hardware-software co-optimization for a long time. The common accumulation of this team originates from the "Energy-Efficient Computing Group (NICS-EFC) of the Nano-Integrated Circuits and Systems Laboratory" led by Professor Wang Yu at Tsinghua University.
Since 2008, the research group has been continuously rooted at the forefront of hardware-software co-optimization and AI computing power system research, accumulating over a decade of technical experience in related fields. This long-term scientific research accumulation has also laid the foundation for the emergence of Infini-AI. At the beginning of the AI Infra industry, the development logic was very simple: stacking more computing power to support larger-scale models.
As domestic chips continue to evolve, the supply of computing power is gradually becoming diversified, and new problems arise: How can chips of different brands and architectures work collaboratively? How can scattered computing power resources be transformed into stable, efficient, and sustainably usable model training and inference capabilities? This is precisely the direction in which the Infini-AI team has been deeply engaged for a long time.
Public data shows that in 2024, Infini-AI completed the world's first "4+2" heterogeneous chip mixed training for a 70-billion-parameter large model, utilizing four domestic chips from Huawei Ascend, Tianshu Zhixin, MetaX, and Moore Threads, along with two imported chips from AMD and NVIDIA, achieving a computing power utilization rate of 97.6%. Currently, it can complete new chip adaptation within 2 working days, reducing the adaptation cost of domestic chips by over 70%; in the training of trillion-parameter MoE models, its throughput is 41% to 75% higher than that of mainstream open-source frameworks, and in inference, it has achieved a 10-fold decrease in cost within a year.
Behind these metrics is the team's technical accumulation, consistently delving into diverse heterogeneity and hardware-software co-innovation from the laboratory to the industrial frontline. In the long run, localization and the diversification and fragmentation of computing power supply will make heterogeneous collaboration a continuous industrial demand. Infini-AI has built its technological system in advance around "M models × N chips," which perfectly covers this long-term demand. This first-mover advantage has given it a favorable position in the domestic heterogeneous layout.
02.Front Store, Back Factory, One Center: Touching Neither Models Nor Chips A Neutral Platform Positioned for the Token Demand Explosion
As Agent applications accelerate, the tasks that AI infrastructure needs to handle become more complex. At the 2026 World Artificial Intelligence Conference, Infini-AI officially proposed the "Front Store, Back Factory, One Center" strategic layout, addressing these challenges in three parts.
1. Computing Power Distribution Center: 37,000P Heterogeneous Computing Power, Running Through Production-Level Mixed Training
The "One Center" refers to the "Computing Power Distribution Center," corresponding to the Agentic Infra autonomous infrastructure platform, which is responsible for aggregating, managing, and scheduling diverse heterogeneous computing power. In response to the current situation of scattered and fragmented heterogeneous computing power, Infini-AI has built a cross-domain training system for large models targeting heterogeneous clusters. This system has passed the production-level tests of three major types of cross-domain computing power aggregation training practices.
The first is ultra-long-distance aggregation. The company, in collaboration with China Telecom, completed the country's first cross-domain mixed training for large models over a distance of more than 4,000 kilometers, achieving a 165% improvement in joint training performance between clusters in Hami, Xinjiang, and Shenzhen, Guangdong.
The second is multi-data center aggregation. It connects 4 GPU clusters of different generations and models to jointly train a large model with nearly 10 billion parameters, achieving a 41% improvement in the collaborative performance of the four clusters.
The third is hybrid cloud computing power aggregation. It achieves secure interoperability and mixed training between local private clusters and public cloud computing power, improving overall performance by 68%.
Currently, this cross-domain training system has been deployed nationwide, reaching over 37,000P of computing power and covering 16 mainstream chips.
2. Token Factory: 40-Fold Growth in Daily Average Call Volume, 10-Fold Decrease in Inference Cost
The "Token Factory" corresponds to the Agentic MaaS large model service platform, which is responsible for transforming model capabilities into callable Token services. In July 2026, Xia Lixue disclosed at WAIC the optimization results of the new-generation inference system pioneered by Infini-AI, the cross-cluster heterogeneous PD separation architecture PDD. Based on the long-term large-scale practical accumulation of Agentic MaaS online business, this architecture achieves a 51.5% reduction in first Token latency while reducing the cost per Token by 37.5%. Over the past year, Infini-AI's Token Factory has driven a 10-fold decrease in inference cost through a series of original technological innovations. The company believes that with the large-scale implementation of the cross-cluster heterogeneous PD separation architecture, there is still room for a 10-fold decrease in inference cost. The growth flywheel is also fully operational. Official data shows that as of July this year, the platform's daily Token call volume has achieved a 40-fold explosive growth compared to December last year.
3. AI Productivity Store: From 3D Games to Law Firm Partners, Tokens Turn into Real Money Here
The "AI Productivity Store" corresponds to the Agentic Infra industry solutions. For the same scenario, different model scales and different hardware configurations may correspond to completely different Token costs. Infini-AI takes agent scenario adaptive technology as the core technological path to improve the conversion efficiency of AI productivity, outputs Agentic Infra solutions to the industry, and lands Tokens into deliverable industrial business outcomes.
In the cultural entertainment and gaming scenarios, Infini-AI joined hands with VAST's Tripo 3D large model to jointly create the AGENTIC 3D GAME GEN solution; in the legal terminal scenario, the company created "Lvdun Xinren," an exclusive AI partner for law firms, and cooperated with multiple law firms and ecological partners; in the energy and power scenario, the company is exploring an intelligent prediction and collaborative scheduling solution for the energy consumption of intelligent computing centers based on agent technology with Nanwang Nengyuan.
The three parts of "Front Store, Back Factory, One Center" respectively cover underlying computing power, intermediate model services, and upper-level industrial applications, forming a business chain from resource scheduling to Token production, and then to productivity conversion.
Infini-AI has another noteworthy feature: it adopts an asset-light operation model. Infini-AI does not own physical data centers itself, but reaches and manages the computing power in third-party facilities distributed across the country through technical services, while adapting to different AI chips such as Huawei Ascend and Moore Threads, exchanging a lighter balance sheet for faster expansion speed.
At the same time, the value of an independent third-party platform is emerging. Infini-AI does not make models, nor does it make chips. It remains neutral. As an independent platform connecting models, chips, and customer needs, it organizes scattered model capabilities and heterogeneous computing power, continuously optimizing the conversion efficiency from underlying resources to AI productivity.
Looking at the current industry landscape, the Infra of large tech companies mainly serves internal business, the Infra of chip manufacturers is bound to the hardware ecosystem, and the self-built Infra of model companies is limited by competitive relationships. Meanwhile, neutral independent cloud vendors are growing into core hubs of the Token economy. Currently, Infini-AI's customers and partners cover top model enterprises, research institutions, and industrial customers. In terms of model services, the company cooperates with enterprises such as Kimi and Zhipu, and the Agentic MaaS platform has launched over a hundred large models, supporting customers to call them on demand.
03.Double First in Funding: All-Star Investors, Valuation Still Has Room
The cumulative funding of CNY 4.3 billion has set a new high for China's AI Infra funding at once, which is a concentrated vote by capital on Infini-AI's development path. From the publicly disclosed investors, the company has formed a diversified shareholder lineup behind it. Since its establishment, its investor lineup can be described as luxurious, covering state-owned institutions, industrial investors, market-oriented financial investment institutions, and other categories.
Among them, state-owned institutions include Hangzhou Gaoxin Jintou, Social Security Zhongguancun Fund (managed by Legend Capital), Zhuhai Technology Industry Group, Xiamen Guoxing Investment, Xuhui Capital, Futeng Capital, CPIC Investment, and CDB Ventures; industrial investors include Dabanc Holding, Lenovo, Jiukun Chuangtou, Tongge Chuangtou, Tencent, Baidu, etc.; financial investment institutions include well-known institutions such as Sequoia China, HongTai Fund, Huiyuan Capital, Qiming Venture Partners, Dachen Caizhi, Shunwei Capital, and Lisi Ziben. The diversification of investors has made it a target of industrial consensus.
The national team looks at long-term compound interest, local state-owned capital looks at digital economy infrastructure, industrial capital looks at ecological synergy, and top VCs look at track explosion. Funds with different attributes are betting at the same time, pointing to the same judgment: AI infrastructure is transforming from a technology project into an industrial foundation.
Compared with overseas, the valuation space for domestic AI Infra manufacturers is still obvious. From public information, it is known that Fireworks AI completed a USD 250 million Series C funding in October 2025, with a valuation reaching USD 4 billion; after completing a USD 1.505 billion Series D funding in July 2026, the valuation rose to USD 17.5 billion. Baseten has taken two rounds of funding this year, and its valuation has also risen sharply to USD 13 billion in a short time.
Together AI, on the other hand, completed another USD 800 million funding in July 2026, reaching a valuation of USD 8.3 billion. It is worth noting that NVIDIA has invested in all three of the above companies. For the chip industry chain, AI infrastructure can help more models run on a larger scale and further release the application value of underlying hardware. The similarity between Infini-AI and them is that they both position themselves in advance in the inference market brought by the explosion of agents, and both serve model manufacturers and AI-native enterprises.
The differences lie in two points:
First, Fireworks, Baseten, and Together AI mainly run on the NVIDIA ecosystem, and the optimization targets are relatively concentrated; Infini-AI faces the reality of the coexistence of multiple chips such as NVIDIA, AMD, Ascend, Biren, and Moore Threads, and must solve the heterogeneous collaboration problem from the bottom up. This is a rigid demand domestically, but not a core proposition overseas.
Second, overseas Neo Clouds generally build or lease GPUs themselves, expand through debt, and rely on the price difference of computing power leasing for profits; Infini-AI does not own data centers, focuses on the software orchestration layer, earns money from efficiency improvement, has lighter assets, faster expansion, and lower dependence on a single hardware system.
These two points together point to a market that has not yet been fully priced by the overseas valuation system. The current cumulative funding of CNY 4.3 billion, compared with the valuations of Fireworks at USD 17.5 billion and Baseten at USD 13 billion, still has obvious room for appreciation.
The successful funding of Infini-AI may release an industrial signal: domestic AI Infra players are beginning to knock on the door of the main table of global Neo Cloud. By outputting high-quality, stable, and sustainable Token services, China has the ability to contribute a unique Chinese solution to global AI development.
04.Conclusion: On the Eve of Demand Explosion Domestic Players Positioned at the Hub of the Token Economy
The explosion on the demand side is no longer a long-term judgment. Agent applications are driving exponential growth in Token consumption, with daily average call volume growing by over 1,000 times in two years. Whoever can produce Tokens at a lower cost and higher efficiency will master the basic production factors in the AI era. Infini-AI has given the answer for domestic players with "Front Store, Back Factory, One Center."
The Computing Power Distribution Center aggregates heterogeneous computing power to expand the scale of available resources; the Token Factory reduces inference costs by 10 times in a year to improve Token production efficiency; the AI Productivity Store enables Tokens to be efficiently converted into productivity in scenarios such as gaming, law, and energy. It connects models, chips, and customers in a neutral posture, accelerates expansion with an asset-light Neo Cloud model. This determination to go all-in on Infra is representative on a global scale.
With a cumulative funding of CNY 4.3 billion, ranking first in both single-round and cumulative funding, numerous top model manufacturers and computing power suppliers have already voted with their cooperation. As multiple rounds of funding complete closing and IPO preparation is launched, the company will also enter a new stage of development. The definer of China's AI Infra is moving from the world's factory to the world's Token factory, taking a seat at the main table of global Neo Cloud.