Alibaba unveils its most powerful AI chip and opens up the software stack

On 22 September, at the Yunqi Conference in Hangzhou, Alibaba chief executive Wu Yongming presented the new 真武 V900 accelerator. The company calls it the most powerful Chinese AI chip, claiming a threefold performance edge over the previous generation M890. A day later, subsidiary T-Head published the open source work on the accompanying software stack.

What was opened

The package, called T-Head SAIL, covers compute framework adaptations, acceleration libraries, the toolchain, and communication libraries. The release lists PyTorch-for-sail, sailify, Triton-for-sail, DeepGEMM-for-sail, and FlashAttention-for-sail, with PCCL and DeepEP-for-sail announced as upcoming. Opening the tools, not just the silicon, serves a specific purpose: it allows customers to adopt the chip without rewriting existing models.

Alibaba states that more than 650 customers across over 20 industries have adopted the technology, including Ant, Xiaohongshu, and carmaker XPeng. Xiaohongshu developed agents capable of managing workload migration and operator optimization. The ModelScope catalogue offers 39 quantized models, including Qwen, DeepSeek, and Kimi, with over 348,000 downloads in September 2026. This confirms that the competitive advantage lies as much in the developer community as in raw performance.

Compute remains scarce

A supply problem looms in the background. A joint report by Inspur and IDC, presented at AICC 2026, estimates that the coverage rate for global AI compute demand will drop from 79% in 2024 to an expected 71% in 2027, before rising to 77% in 2030. The absolute gap is valued at 380.9 billion dollars in 2030, nearly ten times the 2024 value. Meanwhile, global token consumption is forecast to grow at a compound annual rate of 4,822.6% through the end of the decade, reaching approximately 4,000 trillion inference operations per year.

At the same conference, Inspur presented the 元脑 SD200 Ultra system, which scales the configuration from 64 to 128 chips, with expansion memory and storage up to 8 and 64 terabytes, alongside the HC2000 platform. This marks a shift in competition from the single processor to the entire system: those who integrate silicon, software, and distribution scale reduce the risk of being left behind.