Tsingmicro readies space AI chip as Nvidia reshapes data center systems

Chinese startup Tsingmicro and Oriental Starlink plan to launch a space-based AI agent built on reconfigurable chips into a dawn-dusk orbit in the second half of 2026. The system is designed to process satellite data in orbit, addressing the heat, power and radiation limits that make conventional GPUs difficult to use in space.

Tsingmicro’s architecture uses software-defined hardware that can reorganize computing units for tasks such as remote sensing inference, signal processing and data preprocessing. The companies outlined a roadmap beginning with on-orbit verification, followed by 5 to 10 launches and network iterations within three years, with Oriental Starlink planning 1,000 space AI agents and total computing power of 900,000 POPS.

The ground business gives Tsingmicro a base for that push. Its TX8 series shipments have exceeded 30 million units, computing card orders have surpassed 40,000, and deployed plus under-construction computing power exceeds 5,000 PFLOPS. In June 2026, it completed pre-listing tutoring for China’s ChiNext board.

Nvidia is pursuing a parallel shift with Vera, a custom data center CPU intended to improve GPU utilization and system-level token throughput rather than compete only on core count. Vera is built into systems such as Vera Rubin NVL72, reflecting Nvidia’s effort to control the CPU, GPU, interconnect, network card, DPU and software stack in the AI factory.