Alibaba introduced its new artificial intelligence accelerator, Alibaba Zhenwu V900, which can scale up to 500 thousand chips. Here are the details of the new processor.
Alibaba introduced its new generation artificial intelligence accelerator, Zhenwu V900. This new processor, developed by the company's semiconductor unit T-Head, offers three times higher performance than the previous generation Zhenwu M890.
Alibaba Zhenwu V900 features stand out with its 216 GB memory capacity and huge cluster support that can scale up to 500 thousand chips. This hardware was developed specifically for both training artificial intelligence models and inference processes.
Technical Specifications and Scalability
While the new accelerator has 216 GB GPU memory, the inter-chip connection bandwidth reaches 1,200 GB/s. The system supports different data sensitivities including FP8 and FP4.
The previous generation Zhenwu M890 was limited to 144 GB memory and 800 GB/s inter-chip link bandwidth. While 128-chip super nodes are used in M890-based systems, it is stated that more than 1,000 accelerators can be run under a single system on the V900 side.
On a larger scale, Alibaba's new super node infrastructure can handle up to 500 thousand V900 cards. This new super node server includes not only the processor.
System; It brings together Alibaba's ICN Switch network solution, Panmai SmartNIC network card and Zhenyue SSD controller in the same infrastructure. Compute, storage and networking components are optimized to work together.
This configuration provides efficiency in the training and inference processes of very large artificial intelligence models. Alibaba describes the Zhenwu V900 as the most powerful artificial intelligence chip in China.
However, the FLOPS value, which shows the absolute computational performance of the processor, has not yet been announced. Basic technical details such as production technology, manufacturing foundry, power consumption and direct memory bandwidth were also not shared.
Market Competition and Future Roadmap
One of the domestic options against the V900 will be the Huawei Ascend 960PR. 192 GB memory capacity, 2.4 TB/s memory bandwidth and 2.2 TB/s scale-up connection were announced for Ascend 960PR, which is planned to arrive in the third quarter of 2027.
A direct comparison of the computing power of the two accelerators cannot be made at this stage, as not all equivalent performance data for the Zhenwu V900 has been shared yet. Mass production and commercial availability of the V900 will begin in the first quarter of 2027.
In Alibaba's previous road map, the third quarter of 2027 was indicated for V900. Thus, the schedule for the new accelerator was brought forward by a few months.
As of May 2026, T-Head had shipped more than 560 thousand Zhenwu chips to more than 400 external customers. The current number of customers has grown to over 650 in different sectors including automotive, finance, large language models, robotics artificial intelligence, energy and manufacturing.
Alibaba's hardware plan is progressing with the new generation Qwen models. Qwen 4 is currently in the training phase.
In the subsequent Qwen 4.5 and Qwen 5 families, the model scale is planned to be increased to 5 to 10 trillion parameters. The current Qwen3.8-Max model has 2.4 trillion parameters.
In the automatic development work carried out on Qwen3.8-Max, the model completed 33 iterative cycles over the course of a month. At the end of the process, which included data validation, experiment execution, error detection and training optimization, the Artificial Analysis score of the model increased from 40 to 45.
In a separate chip design experiment, the model ran for more than 60 hours and made more than 10 thousand EDA tool calls. In this process, the designed chip area was reduced by 42 percent without loss of performance.
Alibaba Cloud's worldwide data center capacity is planned to be increased to more than 20 GW by 2032. The company also announced that global supply constraints in artificial intelligence infrastructure are limiting its current growth rate.
The deployment of commercial-scale Alibaba AI Supernode systems will begin in the third quarter of 2026. Zhenwu V900 and the new Qwen roadmap were announced at a time when the US continues to impose Chinese export controls on advanced artificial intelligence chips.
Alibaba's announcement comes just ahead of the meeting between US President Donald Trump and Chinese President Xi Jinping on September 24, 2026, as Chinese technology companies work on domestic alternative solutions to Nvidia accelerators. While artificial intelligence was one of the topics discussed in the talks between the two countries, a new artificial intelligence security notification mechanism also came to the agenda during the preparatory talks in New York.
