Alibaba unveils Zhenwu V900 AI chip and opens up SAIL software stack

The race for AI chips is now as much about software as silicon. That was the message from Alibaba at its Yunqi Conference, reported by Chinese media outlet 量子位 (QbitAI). On 22 September, group CEO Wu Yongming introduced the new Zhenwu V900 chip. He called it "China's most powerful AI chip" and said it performs three times better than the M890.

Software, the submerged part of the iceberg

A processor is only as good as what you can run on it. The next day, 23 September, semiconductor subsidiary T-Head announced new progress in opening up its T-Head SAIL software stack. The company expanded framework adaptations, acceleration libraries, the toolchain, and communication libraries.

The company's analogy is clear: the AI chip is the engine, while the software stack acts as the gearbox, transmission, and driving system. No matter how powerful the engine, the result depends on how well the software migrates, runs, and optimizes models.

650 customers and industrial use cases

According to the figures presented, Zhenwu AI chips already serve more than 650 customers across more than twenty sectors. Companies such as Ant Group, Xiaohongshu, and carmaker Xpeng use T-Head SAIL. Xiaohongshu has even built an agent from SAIL's open code to handle model migration and operator optimization, speeding up the deployment of its generative recommendation models.

This opening traces back to the WAIC show in July, where T-Head announced that SAIL was moving to open source. The release included development kits, drivers, performance analysis tools, and technical documentation. Two months later, the open projects include framework adaptation for the PyTorch model, a source code migration tool, an operator development tool inspired by Triton, and computation acceleration projects.

Cutting the cost of migration

A manager in T-Head's software ecosystem summarized the top customer concern: "The strongest demand, from a business perspective, is the cost of migration." Many companies have run their software for years. Switching chips means reassessing that legacy without interrupting operations. Migration tools aim to preserve that asset.

The second issue is performance after migration. If efficiency drops to 30 or 40% of the original, the change becomes hard to accept. Hence the publication of acceleration projects, which allow business teams to understand the implementation and optimize it themselves. The third issue is the speed of adapting new models, with a target of availability on launch day.

On resources, T-Head says it offers 39 quantized models on the ModelScope platform. These cover the Qwen, DeepSeek, and Kimi series, with more than 348,000 cumulative downloads by the end of September 2026. Among the users, Xpeng migrated the training of its intelligent driving models from a GPU platform to a Zhenwu cloud cluster. Ant Group adapted inference for its main models on the Zhenwu 810E and M890 chips.

Chip manufacturing by major groups is shifting from a delivery logic to one of open co-construction.