Technology

China's TPU "Xuyu" Chip Arrives: 896 TFLOPS, 50% Less Power, No Foreign Dependencies

on July 07, 2026

Table of Contents Expand

China's TPU "Xuyu" Arrives
Key Specifications
The TPU Difference
The Token Cost Advantage
Real-World Deployment
Key Takeaways

China's TPU "Xuyu" Arrives

On June 30, 2026, Zhonghao Xinying — a Chinese AI chip company founded by former Google TPU engineers — unveiled its second-generation fully self-developed TPU AI chip, codenamed "Xuyu" (须臾). Alongside it, the company launched the Taize 2.0 AI high-performance computing platform .

The numbers are striking: 896 TFLOPS of mixed-precision floating-point performance — three times the power of the company's first-generation "Chana" chip. 1,792 TOPS for 8-bit inference. Power consumption held to 600W, which is 50% less than comparable chips .

This is not a prototype. Xuyu has already entered mass production, with批量 products delivered to customers. It is already deployed in multiple large-scale AI data centers across China .

Key Takeaway: Zhonghao Xinying's Xuyu TPU delivers 896 TFLOPS — 3x the performance of its predecessor — with 50% less power consumption. Fully self-developed IP, no foreign dependencies. Already in production and deployed in commercial data centers. The era of competitive domestic AI chips is here.

Key Specifications

Specification	Xuyu Performance
Mixed-Precision FP	896 TFLOPS (3x previous generation)
8-bit Inference	1,792 TOPS
Power Consumption	600W (50% less than comparable chips)
Single Node Performance	7.168 PFLOPS (8 chips)
Cluster Scale	Up to 2,048 chips direct interconnect per supernode
Architecture	Fully self-developed IP, instruction set, operator library, and system software
Power Efficiency vs GPU	80% of traditional GPU server power consumption for same workload
Cost Advantage	60% of overseas high-end compute infrastructure cost

Xuyu is built on a fully self-developed TPU architecture. Zhonghao Xinying owns the complete stack: chip IP cores, proprietary instruction sets, operator acceleration libraries, and system software — with no dependencies on foreign technology . This is a critical differentiator for industries with strict security and compliance requirements, such as government, finance, and power grid sectors .

The TPU Difference

Why TPU instead of GPU? Zhonghao Xinying's CEO Yang Gongyifan, a former Google TPU engineer, explains the strategic choice: "General-purpose GPUs have their value in the domestic market, but from a global industry perspective, it is difficult for new players to overtake on that path" .

NVIDIA and AMD have spent decades refining their GPU architectures and CUDA software ecosystems. "Later entrants cannot easily overtake them," Yang argues . TPU, by contrast, is a newer architecture — Google only began developing TPU in 2015 — and still has significant room for micro-architecture innovation .

TPU is purpose-built for AI workloads. It is not a general-purpose compute engine — it is optimized specifically for the tensor math that drives large language models . This yields several advantages:

Higher data reuse: TPU architectures keep data closer to the compute units, reducing the memory bandwidth bottleneck that plagues GPUs
Lower overhead: By focusing on AI-specific operations, TPUs eliminate the circuitry needed for general-purpose computing
Better scaling: The architecture is designed for high-performance, low-latency interconnect across thousands of chips

Yang notes that TPU architectures have "more uncharted territory to explore" compared to mature GPUs, making them a more fertile ground for innovation .

What Makes Xuyu Different

PD separation support: Xuyu natively supports Prefill-Decode separation architecture, allowing different chips to specialize in prompt processing vs. token generation
KV cache optimization: Re-architected data flow scheduling dramatically improves cache hit rates for long-context scenarios
2048-chip direct interconnect: Single supernode supports massive distributed training and inference

The Token Cost Advantage

Perhaps the most compelling metric is cost. In large model inference scenarios, Zhonghao Xinying's platform delivers per-million-token costs at just 35-50% of leading overseas GPU solutions .

This is not a theoretical advantage — it is the primary reason customers are choosing the platform. According to Yang, the main drivers are: "The huge cost difference is the primary reason customers choose us, followed by long-term cluster stability, chip interconnect reliability, and supply chain independence" .

Yang confirmed that Xuyu is already in mass production and has shipped批量 products to customers . The company uses a proactive inventory strategy: "We lock in orders and stock based on business forecasts, not waiting for customer orders before starting production. Our supply chain is stable and we can meet customer demand" .

Cost Context: The hardware cost of next-generation AI servers has risen 60-80% due to component price increases. But Xuyu's performance-per-dollar advantage allows Zhonghao Xinying to absorb these cost pressures through architectural innovation rather than passing them to customers .

Real-World Deployment

The company's first-generation "Chana" chip has already been deployed in multiple large-scale AI data centers, including projects operated by China Unicom Shenzhen, China Mobile Tianjin, Taiji Computer, and Jiangxi Shangrao . These deployments span finance, media, education, and healthcare sectors .

Xuyu has been optimized for major Chinese AI models, including the Qwen series, DeepSeek, GLM, and MiniMax . The software stack supports PyTorch, vLLM, SGLang, and distributed training frameworks DeepSpeed and Megatron-LM .

For AI agent workloads, Xuyu is compatible with the OpenClaw open-source framework, providing on-premise deployment options for enterprise automation use cases .

Analysis: Zhonghao Xinying's Xuyu is not a "lab project" — it is already generating revenue in commercial deployments. The combination of 3x performance, 50% power reduction, and 35-50% token cost savings gives it a clear value proposition for Chinese enterprises looking to reduce dependence on foreign GPU suppliers. The real test will be scaling software ecosystem and developer adoption — but the hardware foundation is now genuinely competitive.

Key Takeaways

#	Key Takeaway
1	896 TFLOPS, 3x previous generation — Xuyu delivers massive performance gains over its predecessor, with 1,792 TOPS for 8-bit inference
2	50% power reduction — 600W power consumption vs. comparable chips, crucial for green data center strategies
3	Fully self-developed — No foreign IP dependency across chip, instruction set, software stack. Meets security and compliance requirements for government and finance sectors
4	35-50% token cost savings — Per-million-token cost is significantly lower than leading GPU alternatives in real-world inference scenarios
5	Already in mass production — Not a prototype. Xuyu is shipping to customers with stable supply chain
6	Deployed in major data centers — China Unicom Shenzhen, China Mobile Tianjin, Taiji Computer, and others
7	Optimized for Chinese AI models — Qwen, DeepSeek, GLM, MiniMax, and more
8	OpenClaw compatible — Supports on-premise AI agent deployment for enterprise privacy and security
9	2048-chip cluster scale — Single supernode can support trillion-parameter model training