Zhonghao Xinying presents low-cost TPU chip at Hong Kong Global Unicorn Summit
Chinese TPU chip maker Zhonghao Xinying showcased its new 'Xuyu' chip at the Global Unicorn Summit in Hong Kong, claiming inference costs as low as 35-50% of overseas GPUs.
The summit, themed "Born to Lead, Built to Create," gathered leaders from technology, capital and industry to discuss frontier fields including AI, integrated circuits, biomedicine, new energy, new materials and robotics. Hong Kong Chief Executive John Lee, Secretary for Innovation, Technology and Industry Sun Dong, and InvestHK Director-General of Investment Promotion Liu Kai-xuan attended the opening ceremony and delivered remarks.
During the roundtable "Intelligent Unbounded: AI Reshaping the Industrial Foundation and Global Innovation Paradigm," Huang Xu said that the rise of intelligent agents has driven exponential growth in computing demand. He noted that the real gap lies not in total token consumption but in how many executable, verifiable and value-creating intelligence results can be produced per unit of computing power. The AI chip competition, he said, is shifting from "who has the strongest computing power" to "who can build a dedicated computing system that better suits large model and agent companies' R&D and business, and who can provide the lowest token cost."
This is the direction Zhonghao Xinying has been pursuing. On June 30, the company released its next-generation TPU architecture AI chip "Xuyu," featuring a mixed-precision floating-point computing power of 896 TFLOPS, three times the overall performance of its predecessor "Chana." Its 8-bit inference computing power reaches 1792 TOPS, with dedicated optimization for token generation, context caching and batch concurrent inference. The chip has completed in-depth operator optimization with leading domestic large models, and the cost per million tokens for inference is only 35-50% of that of overseas GPUs.
As "Xuyu" enters mass production and thousand-card and ten-thousand-card clusters scale up, Zhonghao Xinying said it will continue to promote the large-scale deployment of high-performance, low-cost domestic computing power across industries, working with industrial partners to build a more efficient and open global future for domestic computing infrastructure.