Infinigence AI, a company building AI-native infrastructure, said on October 12 it had closed a 2.3 billion yuan Series C. Together with earlier rounds, that brings total funding to 4.3 billion yuan in the three years since it was founded, a record for China's AI-native infrastructure sector both for a single round and cumulatively. The Tsinghua-affiliated company does not make chips or models. It schedules and operates computing power.

Notably, the round's investor list includes state capital such as the Social Security Zhongguancun Fund and CDB Kechuang; corporate investors including Lenovo, Tencent and Baidu; and financial backers such as Sequoia China, Qiming Venture Partners and Fortune Capital. All three types of capital rarely show up at once in a company only three years old.

Four Rounds in Three Years, at a Quickening Pace

Infinigence AI was founded in May 2023 by Wang Yu, a tenured professor in Tsinghua University's Department of Electronic Engineering, together with his students Xia Lixue and Dai Guohao. Xia serves as co-founder and CEO. According to Securities Times, Wang had earlier co-founded the AI chip company DeePhi Tech with students in 2016; Xilinx acquired it in 2018.

Securities Times traces a tight funding cadence. In September 2023, just four months after founding, the company closed an angel round backed by Baidu, Zhipu, Matrix Partners China and ZhenFund. In September 2024 came a Series A of nearly 500 million yuan, co-led by the Social Security Zhongguancun Fund, Qiming Venture Partners and Hongtai Capital, taking cumulative funding to almost 1 billion yuan within 16 months. A Series A+ in November 2025 added nearly 500 million yuan more. In May 2026 the company raised over 700 million yuan, then a domestic single-round record for an AI-native infrastructure company. Now comes the 2.3 billion yuan Series C.

The financing pace mirrors the business. According to Securities Times, the company completed a joint-stock restructuring in March this year and renamed itself Shanghai Infinigence Intelligent Technology Co., Ltd.

The Technical Anchor Is Heterogeneity, Not Raw Scale

Unlike most computing power companies, Infinigence AI does not lead with size. By its own account, its computing power aggregation and distribution hub has deployed and reaches 37,000 petaflops of capacity across 16 mainstream chip types.

Securities Times describes the positioning as unusual: no chips, no models, purely a scheduler and operator of computing power. The company owns no large-scale data centers itself. Instead, a software layer stitches together heterogeneous capacity spread across dozens of third-party facilities.

Two technical metrics make that asset-light model work. The first is heterogeneous mixed training. The company previously launched what it called the world's first 1,000-chip-scale heterogeneous mixed-training platform, built to scale to 10,000 chips. According to Securities Times, it has run large-model mixed training combining four domestic chips, Huawei Ascend, Iluvatar CoreX, MetaX and Moore Threads, each paired with two imported chips from AMD and NVIDIA, with computing power utilization reaching as high as 97.6%. The second is cross-regional scheduling. Working with China Telecom, the company completed China's first large-scale cross-regional mixed-training validation spanning more than 4,000 kilometers, with the task chain running from Hami in Xinjiang to an inference cluster in Shenzhen, Guangdong, supporting stable training of a 30-billion-parameter model.

Why this matters: China's computing power market currently runs on multiple coexisting chip architectures, with idle installed capacity, difficult cross-regional scheduling and high costs for heterogeneous compatibility. Pulling capacity from different vendors in different regions into a single schedulable pool is itself a scarce capability.

Token Factory Cost Cuts Track the Inference Cycle

According to Ifeng Tech, Infinigence AI released a cross-cluster heterogeneous PD-disaggregation architecture (PDD) in July 2026, cutting first-token latency by 51.5% and per-token cost by 37.5%. The company says its Token Factory has driven inference costs down tenfold over the past year, and that as of July this year, daily token calls on its Agentic MaaS platform were 40 times higher than in December of last year.

Those numbers land at the moment inference demand becomes the main event. Securities Times cites Gartner forecasts that global inference spending will reach $23.3 billion in 2026, surpassing training spending of $19 billion for the first time and accounting for 55% of total AI infrastructure services spending. A CAICT report projects that inference will make up 70.5% of total AI computing load in 2026. Data from the National Data Administration shows China's daily token calls rose from the order of hundreds of billions in early 2024 to 100 trillion by the end of 2025, passing 140 trillion in March this year.

Securities Times reports that the industry's competitive logic is shifting from piling on scale to managing it finely, supporting more business value with less computing power. Comparable deals are heating up too: SiliconFlow has raised nearly 2.9 billion yuan so far this year, while US-based Fireworks AI is targeting a $30 billion valuation in its new round.

What the Investor Mix Reveals

The round's three investor groups are worth separating out. State-backed investors include Hangzhou High-Tech Jintou, the Social Security Zhongguancun Fund (managed by Legend Capital), Zhuhai Technology Industry Group, Xiamen Guoxing Investment, Xuhui Capital, Futeng Capital, China Insurance Investment and CDB Kechuang. Corporate investors include Dabanc Holding, Lenovo, Ubiquant Ventures, Tongge Capital, Tencent and Baidu. Financial investors include Sequoia China, Hongtai Capital, Huiyuan Capital, Qiming Venture Partners, Fortune Capital, Shunwei Capital and Lisi Capital.

According to Securities Times, it is rare in today's cautious primary market for hard tech to see a three-year-old company draw national social security money, local state capital and leading internet players at the same time.

The customer list backs up the logic. Securities Times reports that Infinigence AI's clients include Kimi, Zhipu, Lenovo and Li Auto. Ifeng Tech reports the company has signed partnerships with leading large-model firms including Zhipu and Moonshot AI, with industry solutions spanning entertainment and gaming, healthcare, legal services, and energy and power. An infrastructure provider that builds no models yet serves model companies, and takes money from both model developers and end-device makers, is itself a validation of the middle layer's value.