Alibaba's Qwen3.8-Max and ZTE's Nebula Share First Place in SuperCLUE Embodied Brain Ranking

Alibaba officially announced the Zhenwu V900, its latest processor designed for training and inference tasks of artificial intelligence models. According to the company, this accelerator demonstrates three times the performance of the Zhenwu M890, its previous version, and features 216 GB of memory.

This launch took place during the Apsara 2026 conference, held in Hangzhou, China. Eddie Wu, CEO of Alibaba, stated that the V900 is currently the 'most powerful AI chip in China.' However, the company did not provide sufficient data to validate this claim or to conduct a performance comparison with its competitors' chips.

The Zhenwu V900 was designed to meet two primary demands of AI: model training and inference, which is the phase where a trained model executes the necessary calculations to produce responses. In addition to 216 GB of memory, the model offers a bandwidth of 1,200 GB/s in inter-chip communication; for comparison, the M890 has 144 GB of memory and 800 GB/s inter-chip communication.

The T-Head semiconductor division reported that more than a thousand units can operate together as a single system. On an even larger scale, Alibaba stated that the accelerator has the potential to integrate clusters containing up to 500,000 boards.

Mass production and commercial release of the device are scheduled for the first quarter of 2027. Despite the promises of great capacity, crucial information is still missing for a complete evaluation of the V900. Alibaba also omitted details on FLOPS performance, manufacturing method, the company responsible for component production, or its energy consumption, leaving the claim of being the most powerful chip in China as a statement from the manufacturer itself.

Additionally, this new accelerator integrates Alibaba's plans to develop larger AI models. The company plans to create a future Qwen with a scale of 5 to 10 trillion parameters, while the current Qwen3.8-Max has 2.4 trillion.

This model expansion aligns with the Chinese giant's projects aimed at increasing its computing infrastructure. Alibaba aims to exceed 20 gigawatts of capacity in its data centers by 2032. However, Wu mentioned that shortages in the global supply chain for AI data centers are restricting the speed of this expansion.

XFEON, a company based in Chengdu, unveiled a product matrix integrating chips and algorithms at an event held on September 22, 2026. This matrix encompasses edge silicon, embodied data tools, Token workstations, and an embodied brain, according to reports from EEFocus and other local sources referencing www.xfeon.com.

The company's corporate materials use the official English designation XFEON; the presented product lines include Xinghe S1, AGLobe, and Xinghui Token workstations. While Chinese press uses additional explanations, some English funding reports mention Xingfan Intelligence, but this review prefers to use the official XFEON designation.

At the silicon level, Xinghe S1 is positioned as the core computation engine for physical AI at the edge. It is a specialized multi-core design with a proprietary AI acceleration core, supporting native mixed precision INT2–INT8, delivering approximately 105 TOPS equivalent to INT8 on the die, typical power consumption of around 25W, and a claimed efficiency exceeding 4 TOPS/W. Company materials also assert that through co-design of software and hardware, approximately 50% higher inference throughput, about 85% smaller model memory footprint, and roughly 90% less memory usage for continuous learning are achieved compared to previous baseline models.

Around S1, XFEON describes SKU series such as Xinghe R (for robotics, up to 420 TOPS), Xinghe K (for orbital satellite computing), and Xinghe E (for edge/industrial applications).

Regarding the algorithmic component, AGLobe is presented as an embodied brain built around conceptual learning rather than exclusively on vision-language-action pathways or world models. It aims to extract transferable concepts from multiple or even single demonstrations across time, space, quantity, and causality. Computing products include the Xinghui D/N series Token workstations, designed for private deployment, training, and high-performance inference, featuring mixed-precision packages from BF16 to INT2. It is claimed that an eight-card node can host a model with 671 billion parameters.

Supporting components mentioned in reviews include the XGAIA data toolkit, DexCore simulation devices, and the XBoost acceleration layer for the edge. The XFEON matrix, presented in September, represents a vertical stack of physical AI, extending from the Xinghe S1 edge silicon through AGLobe and Xinghui Token systems, and is not merely a chip release or a stock market story.