China's most closely watched AI infrastructure startup has raised again. Infinigence AI has completed a RMB 2.3 billion Series C that has been fully delivered, according to multiple Chinese media reports citing people familiar with the deal, making it both the largest single round and the largest cumulative raise - RMB 4.3 billion, roughly $600 million - in the country's AI-native infrastructure sector. The company was initiated by Tsinghua University professor Wang Yu and co-founded by his former doctoral student Xia Lixue, who serves as CEO.
The investor list spans the full Chinese funding stack. State-backed names include Hangzhou Gaoxin Jintou, the Shezhongguancun fund managed by Legend Capital, Zhuhai Science and Technology Industry Group, Xiamen Guoxing Investment, Xuhui Capital, Futeng Capital, China Insurance Investment and SDIC Science and Technology Innovation. Industrial investors include Lenovo, Tencent, Baidu, Dabanc Holding, Jiukun Capital and Tongge Venture Capital. Financial backers include Sequoia China, Hongtai Fund, Huiyuan Capital, Qiming Venture Partners, Fortune Capital, Shunwei Capital and Lichi Capital.
What the company sells is the layer between chips and models. Infinigence describes an "Agentic Infra" strategy built on a "front shop, back factory, one center" layout: a compute distribution center that pools heterogeneous hardware, a token factory that serves models as a platform, and an industry-solutions arm that moves tokens into real applications. The platform currently reaches roughly 37,000 petaflops of compute across 16 mainstream chip types and supports training, inference and reinforcement-learning workloads at the 100,000-card scale.
The differentiator is the heterogeneous part. Infinigence built what it describes as the world's first thousand-card mixed-training platform spanning different chip architectures, scalable to 10,000 cards, and has demonstrated mixed training that combines four domestic chips - Huawei Ascend, Tianshu Zhixin, MetaX and Moore Threads - with AMD and Nvidia hardware, at reported utilization of up to 97.6%. Together with China Telecom it also completed China's first large-scale cross-domain mixed-training run over more than 4,000 kilometers, from Hami in Xinjiang to Shenzhen, training a 30-billion-parameter model.
The demand-side numbers are company-reported and unaudited. Single-day token volume on its MaaS platform has grown about 40 times since December last year, and the company says its token factory has pushed inference costs down 10 times over the past year, with another tenfold reduction still expected. Those figures come from Infinigence's own disclosures rather than an independent audit.
The IPO runway is the subtext. Around a week before the round was disclosed, Bloomberg reported that Infinigence had secretly filed for a Hong Kong listing, with a potential debut as early as the first half of 2027. Founded in 2023, the company would reach a public market in about four years - fast by Chinese standards, where the average STAR Market listing takes roughly 14 years. If it lists, it would be Wang Yu's second outcome in AI entrepreneurship, after his earlier deep-learning startup DeePhi Tech was acquired by Xilinx.
Why the money is flowing here: Gartner projects global inference infrastructure spending will overtake training spending in 2026 as agents and enterprise deployment take over, and Goldman Sachs has forecast that token consumption could grow to 24 times current levels by 2030. Infinigence is betting that the scarce asset in that world is not the model but the layer that turns whatever silicon is available into usable tokens - which is precisely the bet its new investors are underwriting.
Comments (0)
Log in to join the discussion
Log InNo comments yet