The archive · AI & Models · Strategic decision · 2023–2025
Infinigence AI bets software can solve computing power shortage
In 2023, Wang Yu founded Infinigence AI to unify fragmented chips via software; by November 2025, it raised nearly RMB 500 million and managed 25000P.
Infinigence AI (无问芯穹)
What the business is
Builds AI computing power infrastructure: creates an M×N middleware layer between models and chips, providing a heterogeneous computing power cloud platform, intelligent computing all-in-one machines, and inference acceleration tools, and sells computing power services to large model companies.
Starting capital：In September 2023, the first round was invested in by Sequoia China (红杉中国), GSR Ventures (金沙江), Xuhui Capital (徐汇资本), and others; Sequoia held about 10.7% and was the largest external institutional shareholder. In November, Tencent (腾讯), Baidu (百度), Matrix Partners China (经纬), ZhenFund (真格), Qiming Venture Partners (启明创投), and Zhipu AI (智谱AI) followed.
How it started
In May 2023, Wang Yu (汪玉), chair of the Department of Electronic Engineering at Tsinghua University (清华), started a new venture after Deephi (深鉴科技) was acquired by Xilinx (赛灵思), co-founding Infinigence AI (无问芯穹) with his students Xia Lixue (夏立雪), Zeng Shulin (曾书霖), and Dai Guohao (戴国浩). At the time, domestic large models were booming, supply of Nvidia (英伟达) high-end cards was restricted, and domestic chip ecosystems operated separately; the team bet that software could assemble these computing resources.
What happened
In September 2023, Sequoia (红杉), GSR Ventures (金沙江), and others invested; in November, Tencent (腾讯), Baidu (百度), Zhipu AI (智谱AI), Matrix Partners China (经纬), ZhenFund (真格), and Qiming Venture Partners (启明创投) followed. In July 2024, at the World Artificial Intelligence Conference (世界人工智能大会), the company released the industry's first thousand-card-scale heterogeneous chip mixed-training platform, with computing power utilization up to 97.6%, enabling cross-training among six chip types: Huawei Ascend (华为昇腾), Iluvatar CoreX (天数智芯), MetaX (沐曦), Moore Threads (摩尔线程), AMD, and Nvidia (英伟达). Zhipu (智谱), Moonshot AI (月之暗面), Shengshu Technology (生数科技), and others have already stably used heterogeneous computing power on its cloud platform.
How it ended up
Still expanding: in November 2025, it completed a nearly RMB 500 million Series A+ round, led by Zhuhai Technology Group (珠海科技集团) and Futeng Capital (孚腾资本), with follow-on investment from existing shareholders including Huiyuan Capital (惠远资本), Shangqi Capital (尚颀资本), Honghui Fund (弘晖基金), Lenovo Capital (联想创投), and Legend Capital (君联资本). It has already managed more than 25000P of computing power nationwide, covering 53 core data centers in 26 cities; customers include Baichuan Intelligence (百川智能), Kimi, Lenovo (联想), Li Auto (理想汽车), Shengshu Technology (生数科技), and Zhipu (智谱).
Background
Infinigence AI was founded in May 2023 by Tsinghua's Wang Yu and students Xia Lixue, Zeng Shulin, and Dai Guohao. It bet on computing power middleware, not large models, to assemble fragmented domestic chips into unified computing power.
Sequoia, GSR, and Xuhui invested first in September 2023; Tencent, Baidu, and others followed in November. In July 2024, it released a thousand-card heterogeneous training platform with 97.6% utilization, enabling cross-training across six chip types. Zhipu, Moonshot, and others used its cloud.
In November 2025, it completed a nearly RMB 500 million Series A+ led by Zhuhai Technology Group and Futeng Capital. It manages over 25000P across 53 data centers, serving Baichuan, Kimi, Lenovo, Li Auto, and Zhipu, and bets on agent infrastructure.
What has to be true
- The real bottleneck is fragmented domestic chip ecosystems, not total supply; software middleware fills this gap.
- It avoids building models or hardware, earning by making computing power usable, avoiding direct competition with Nvidia and model companies.
- The team, validated by Tsinghua and Deephi, made Sequoia, Tencent, and Baidu bet at pre-revenue stage.
- Once the M×N middleware becomes standard, switching costs are high, and first movers capture the computing power pool.
- The Series A+ brought in Zhuhai state capital that brings resources; the bet includes local governments' willingness to pay for computing power.
What can be applied
When resources are scarce, the winner standardizes fragmented resources into services; assemble domestic chips into unified computing power and sell it.
Aftermath
As of December 2025, Infinigence AI remains in expansion: the Series A+ funds go to software-hardware collaboration, AI cloud and terminal solutions, and agent infrastructure. It launched Infra Agents and Kernel Mind. CEO Xia Lixue sees Agentic AI as a strategic opportunity. No revenue or IPO disclosed.
Sources
spotted an error? The archive wants to know.
Your turn
You just read one. Describe what you are building, and see who is betting on the same thing.
Free account · 3 free questions · no card