The Pulse
Infinigence AI Raises RMB 2.3 Billion for Chip-Agnostic Infrastructure
China’s Infinigence AI has completed a RMB 2.3 billion Series C, bringing its cumulative funding to RMB 4.3 billion. The company says its platform connects computing capacity across 16 chip types, while its announcement does not identify in
AI.info Team · Written with AI ·
Infinigence AI says its infrastructure can connect computing capacity across 16 chip types; its RMB 2.3 billion Series C announcement, published October 12, does not name the investors in this round. The funding brings the Chinese company’s cumulative financing to RMB 4.3 billion, according to company materials cited by TechNode. The gap between a broad historical investor roster and an undisclosed round lineup matters: the company’s claims describe a technical platform, while the financing announcement offers little detail about the deal’s backers.
RMB 2.3 billion, with no round-by-round investor list
Infinigence, known in China as Wuwen Xinqiong, was founded in 2023 and develops software and infrastructure for running AI models across different chip platforms. The company’s materials say its infrastructure spans domestic and international processors and supports mixed-chip training, rather than tying model workloads to a single hardware family. The new Series C is its largest disclosed financing round to date, based on the figures in the announcement.
A separate company announcement carried by Sina lists investors from state-linked institutions, industrial companies and financial firms that have backed Infinigence over time. Those names include Lenovo, Tencent, Baidu, HongShan, Qiming Venture Partners and Shunwei Capital, alongside several Chinese investment organizations. The announcement does not say which of them, if any, participated in the new Series C, so the historical roster should not be read as the round’s investor list.
Capacity claims span 16 chip types
Infinigence says its platform has access to 37,000 petaflops of computing capacity across 16 mainstream chip types. Its materials describe a thousand-card mixed-chip training platform designed to scale to 10,000 cards, and report a peak computing utilization rate of 97.6% for certain mixed-chip training workloads. The company also says it has tested large-scale cross-region training with China Telecom, linking computing clusters between Hami in Xinjiang and Shenzhen in Guangdong across more than 4,000 kilometers.
Those figures are company-reported performance and deployment claims, not independent benchmarks published with the funding announcement. The distinction is important for a business whose pitch rests on coordinating hardware from multiple suppliers: the platform’s value depends not only on how much capacity it can reach, but on how reliably customers can put that capacity to work.
Inference efficiency is the business test
The company’s second set of figures concerns inference, the process of generating model responses after training. Infinigence says its cross-cluster heterogeneous prefill-decode separation architecture reduced time to first token by 51.5% and cost per token by 37.5%. It also reports that daily token calls on its Agentic MaaS model-service platform were 40 times higher in July than in December 2025.
These claims put the company’s financing in the context of its operating case: pooling diverse chips is intended to make model services faster or less expensive, while software and scheduling connect the hardware to customer workloads. Infinigence says it provides inference services and enterprise AI-agent infrastructure for sectors including gaming, health care, legal services and energy. The funding announcement does not disclose a valuation, the round’s detailed investor allocation or a breakdown of revenue from those services.
For now, the clearest public numbers are the RMB 2.3 billion raised and RMB 4.3 billion in cumulative financing. The company’s remaining case rests on reported platform capacity and efficiency metrics, with no round-specific investor list attached to the announcement.