--- title: "Infinigence raises US$324 million to stitch China's fragmented AI compute together" date: 2026-10-25 category: Capital & Markets site: NeuroAI canonical: https://neuroai.site/a/na-capital-infinigence-series-c language: en --- # Infinigence raises US$324 million to stitch China's fragmented AI compute together > Shanghai Infinigence (无问芯穹) closed a 2.3 billion yuan Series C in October 2026, lifting cumulative funding to 4.3 billion yuan and setting a record for China's AI-native infrastructure sector. Most of China's AI chips do not speak the same language. A single training job can span silicon from a dozen vendors, each with its own software stack and quirks. One Shanghai startup just convinced investors that the layer tying all of it together is worth backing with serious money. ## Key takeaways - Infinigence (无问芯穹) completed a 2.3 billion yuan (≈ US$324 million) Series C in October 2026. - Cumulative funding since its 2023 founding reached 4.3 billion yuan (≈ US$606 million), described as a record for China's independent AI-native infrastructure (AI Infra) companies. - The company runs a platform that aggregates about 37,000P of compute across 16 chip types and says its inference cost has fallen about 10x. - Backers span state-linked funds, internet giants, and financial investors, including Tencent, Baidu, Lenovo, Sequoia China, and local government funds such as Xuhui Capital; partners include Zhipu (智谱) and Moonshot (月之暗面). - A Bloomberg report, unconfirmed by the company, says Infinigence has made a confidential Hong Kong IPO filing targeting early 2027. ## The raise, in numbers According to disclosures reported by Chinese science and financial media on October 12, 2026, Infinigence closed a 2.3 billion yuan Series C. Combined with prior rounds, total funding since the 2023 founding reached 4.3 billion yuan. Both the single round and the cumulative sum are described as records for China's independent AI infrastructure companies — the kind of milestone that signals a sector, not just a company, has arrived. ## What Infinigence actually does Founded in 2023 as a spin-out from Tsinghua University and Shanghai Jiao Tong University, Infinigence was initiated by Tsinghua electrical-engineering professor Wang Yu (汪玉) and is led by CEO Xia Lixue (夏立雪). It builds what it calls "Agentic Infra" — the plumbing between raw chips and usable AI. Its compute-aggregation center schedules heterogeneous hardware. The company says the platform now reaches about 37,000P of compute and covers 16 mainstream chip types, including domestic accelerators from Huawei Ascend (昇腾), Iluvatar (天数智芯), MetaX (沐曦) and Moore Threads (摩尔线程) mixed with NVIDIA and AMD parts in the same training run. The "token factory" side is its Model-as-a-Service layer: Infinigence says inference cost on its platform fell about 10x over the past year, and daily token calls grew roughly 40x between December 2025 and July 2026. Those are company-stated metrics, useful as direction rather than audited fact. ## Why investors care about the middle layer China's compute supply is unusually fragmented: multiple chip vendors, multiple generations, multiple regions, often running incompatible software. That fragmentation is a headache for anyone training or serving large models at scale, and it has created a distinct local category — the neutral, third-party AI infrastructure (AI Infra) layer that sits above the chips and below the applications. Infinigence positions itself as that neutral base, and it already works with model labs such as Zhipu (智谱) and Moonshot (月之暗面). The bet is that as AI moves from one-off model training to continuous, high-volume inference, the bottleneck shifts from "who has the biggest cluster" to "who turns compute into cheap, reliable tokens most efficiently." The pattern is not unique to China: overseas "NeoCloud" players such as Fireworks AI, Baseten, and Together AI have also pulled in large funding rounds, a sign that global capital now treats inference infrastructure as a core growth market rather than a back-office cost. According to Li An, chief scientist at BrainNet (脑机网), the value of AI Infra is shifting from how much compute you own to how cheaply and reliably you convert it into tokens, and that pivot is exactly what makes the middle layer investable. The logic is structural: a fragmented hardware base needs a unifying software layer far more than a unified one does. ## The IPO rumor, handled carefully Bloomberg, citing people familiar with the matter, reported that Infinigence has made a confidential filing to list in Hong Kong, with a pre-IPO valuation around 14.3 billion yuan and a possible debut in early 2027. The company has not confirmed this. I treat it as reported-but-unverified and keep it out of any hard claim in this article. ## Honest limitations The funding figures — 2.3 billion yuan Series C and 4.3 billion yuan cumulative — are corroborated across Chinese state-science outlets (Science and Technology Daily), financial papers (China Securities Journal), and a district government release, so they are reasonably solid. The 37,000P, 16-chip, 10x-cost, and 40x-token figures are company-stated and echoed by those same outlets; they indicate trajectory, not audited performance, and I have not independently verified them. The IPO rumor is a single, unconfirmed Bloomberg report. I have not verified the full cap table or the precise lead investor of this round, because primary disclosure was not public; I therefore avoid naming a lead investor rather than guess. ## What readers can do now - Watch whether Infinigence's "token cost" claims appear in independent benchmarks; that is the real test of an Infra company, not the funding headline. - Treat the Hong Kong IPO report as a signal to watch, not a confirmed event — wait for the official filing before drawing conclusions. - If you build on Chinese AI stacks, the fragmentation problem is real; evaluate neutral aggregation layers on compatibility breadth and verified cost, not marketing language. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-capital-infinigence-series-c Free to quote with attribution and a link to the original.