Infinigence AI raises about $344M to pool mixed AI chips

Infinigence AI has completed an approximately $344 million Series C for software that makes mixed fleets of AI chips work as a shared pool. Model developers and enterprises use its infrastructure to train and run models across processors from different vendors, rather than rebuilding every workload around a single hardware stack. Sina Technology reported the company’s October 12 announcement, while Dealroom supplied the dollar equivalent for the original financing amount.

The company sits above the hardware layer: it doesn’t make chips and, according to 36Kr, doesn’t own the physical data centers whose capacity it manages. Its platform aggregates third-party computing resources, schedules jobs across chip architectures and turns deployed models into callable services. That puts its business in orchestration and model-serving infrastructure instead of a large GPU estate of its own.

Sina reported that Infinigence’s platform spans 16 mainstream chip types. The company says it has mixed four Chinese processors—including Huawei Ascend and Moore Threads hardware—with AMD and Nvidia chips for large-model training, reaching computing utilization as high as 97.6%. The intended distinction is compatibility across capacity that would otherwise remain in separate vendor stacks.

On the inference side, Infinigence calls its service layer a “token factory,” but the work is straightforward: it runs models and returns their outputs through callable services. The company says a cross-cluster architecture reduced time to the first token by 51.5% and cost per token by 37.5%, while daily token calls in July 2026 were 40 times their December 2025 level. Those are company-reported operating results, not independently audited figures.

For enterprise deployments, its agent layer chooses models and computing resources based on task complexity, latency and token cost, then coordinates the work into an industry application. Sina says Infinigence collaborates with Zhipu AI and Moonshot AI, but described them as partners rather than disclosed paying customers.

The harder test is whether that coordination holds up when the hardware is geographically scattered. Infinigence and China Telecom validated cross-region mixed training over more than 4,000 kilometers, according to Sina, linking computing in Hami and Shenzhen.