T-Head Semiconductor

What T-Head is

T-Head Semiconductor is the chip-design arm of Alibaba Group. It builds processors for cloud computing and AI workloads, uses them in Alibaba's own services, and sells them to outside customers.^[1][2]

Origins

Alibaba set up T-Head in 2018, announcing the new unit at that year's Apsara Conference.^[1] It was assembled from two pieces: Zhongtian Micro, a chip company Alibaba had acquired, and the chip team of DAMO Academy, the group's research institute. The unit sits under full ownership of the group.^[1]

First chips: an inference NPU and a server CPU

The unit's first widely noted product was Hanguang 800, an NPU for AI inference, released in September 2019.^[1] Inference is the stage at which a trained model answers requests, and an NPU is a processor built to speed up neural-network arithmetic, so this chip was aimed at running models rather than training them.

In 2021 T-Head released Yitian 710, a server CPU built on a 5 nm process.^[1] A server CPU is the general-purpose processor that carries most ordinary computing in a data centre; the process node describes the manufacturing generation, and smaller numbers generally mean more transistors packed into the same area.

The Zhenwu PPU

T-Head began internal development of its PPU AI processor in 2020, before the latest wave of growth in AI models.^[1] The chip is sold as Zhenwu 810E. According to the company, it uses a self-developed parallel computing architecture and a die-to-die interconnect — the links that let separate pieces of silicon inside a package exchange data at high speed — together with a full self-developed software stack.^[1] The software side matters because a processor is only usable at scale if developers have compilers and libraries that turn model code into work the chip can run.

Where it is used

Alibaba uses Zhenwu PPUs at scale for training and inference of its Qwen models.^[1] The chips are deployed in several clusters of more than ten thousand cards each on Alibaba Cloud, and customers include State Grid, the Chinese Academy of Sciences, XPeng and Sina Weibo.^[1][2]

T-Head said that by February 2026 it had delivered 470,000 Zhenwu PPUs in total, more than 60% of them to external customers, serving over 400 enterprises.^[2] It also said that, as of April 2026, more than 300,000 Zhenwu PPUs were in use on Alibaba Cloud by more than 30 automakers and autonomous-driving solution providers.^[2]

Position in the industry

IDC data put T-Head second among Chinese AI chip vendors in 2025, with 265,000 units shipped.^[2] Taken as a sequence, the products trace a path from an accelerator for inference and a general-purpose server CPU toward a processor Alibaba uses for both training and inference of its own model family, paired with a software stack the company says it built itself.^[1] Combined with the disclosed delivery figures, that makes T-Head one of the larger suppliers of AI chips in China by volume.^[2]