Infinigence AI closed a RMB 2.3 billion Series C round, according to reports published on October 12, 2026 by Pedaily, The Paper and Securities Times, taking its total funding since founding to RMB 4.3 billion — the largest single round and the largest cumulative total in China's AI-native infrastructure sector. Founded in 2023 as a spin-out driven by Tsinghua University's Department of Electronic Engineering, the company builds neither chips nor data centers. Its business is the scarcest one of the moment: turning scattered heterogeneous compute into token capacity that can be drawn on demand.
A middleware layer that sells no chips — why it is worth this much
Infinigence AI works in the middle layer: it does not design GPUs upstream and does not build foundation models downstream. It focuses on heterogeneous compute scheduling, hardware-software co-optimization and utilization across large clusters. Its M×N paradigm aims to let M kinds of models run efficiently on N kinds of chips. Its Agentic Infra platform now reaches 37,000 petaflops of deployed compute across 16 mainstream chip types and supports training, inference and reinforcement-learning workloads at the 100,000-accelerator scale. Earlier, together with China Telecom, it completed China's first long-distance cross-domain hybrid training validation spanning more than 4,000 kilometers, with compute utilization peaking at 97.6%.
What investors are really buying is token capacity
In 2026 the center of gravity in large-model competition is shifting from training to inference. Infinigence AI describes its layout in three parts: a compute distribution hub, a token factory and industry solutions. As of July 2026, daily token calls on its MaaS platform had grown 40-fold compared with December 2025; over the past year its token factory cut inference costs roughly tenfold through optimizations such as a cross-cluster heterogeneous disaggregated architecture. In plain terms, investors are not buying the capacity of any single chip. They are buying the efficiency saved when electricity and silicon are converted into tokens — and in a year when agent applications multiply call volumes, that efficiency gets priced again with every call.
What this money changes, and what it does not
For model companies and enterprise customers, a neutral scheduling layer means capacity does not have to be bet on a single chip or a single cloud, and bargaining power over migration and expansion stays in their own hands; this site's earlier report on Anthropic's reported multi-year cloud deal with IREN shows the same shift, with compute supply moving from buying hardware to contracting capacity. The boundary is equally clear: no scheduling layer can outrun the physical supply of high-end chips or cluster network conditions. RMB 4.3 billion buys expansion speed, not a higher physical ceiling. The real test ahead is whether 100,000-accelerator scheduling holds its claimed utilization under real customer workloads.
