Chinese AI chip company Zhonghaoxinying is making significant strides, with plans to launch its second-generation 7nm Tensor Processing Unit (TPU) in the second quarter of 2026. The chip is currently in the post-silicon testing phase and aims to compete directly with top-tier international products like NVIDIA's H100 and Google's TPU v5p.
Founded by Yang Gongyifan, a former core developer of Google's TPU, the company successfully mass-produced its first-generation 12nm "Chana" TPU in 2023. Leveraging its fully self-developed IP cores and instruction set, that chip reportedly delivered 1.5x the computing power of NVIDIA's A100 while consuming 30% less energy under similar workloads. Based on this foundation, Zhonghaoxinying has established a rapid 1 to 1.5-year product iteration cycle.
The new 7nm chip is designed for demanding scenarios like autonomous driving model training and large-scale data center inference. It is expected to deliver a single-chip performance of 400-800 TFLOPS. Due to its architecture being custom-built for deep learning from the ground up—unlike GPUs adapted from graphics processing—the company claims its chips can achieve 3-5x the performance uplift at the same process cost compared to traditional GPUs, with cloud inference cost-performance potentially over 2x better.
CEO Yang Gongyifan emphasized the explosive growth in demand for inference-side computing power, which is expected to surpass training as the market's core growth driver. As large model algorithms mature, he believes custom architectures like TPUs are better suited for diverse scenarios. The company has built an integrated "chip-cluster-model" solution, expanding its customer base from research institutes to sectors like finance and healthcare, which now contribute over 30% of its revenue, making it one of the few profitable domestic AI chip firms.
ICgoodFind's Insight
Zhonghaoxinying's progress demonstrates the growing capability and ambition within China's specialized AI chip sector. Its focus on a rapid, competitive product cadence and a full-stack solution positions it as a notable player in the quest for diversified, domestic AI computing supply.
