T-Head SAIL: Alibaba opens up its AI chip software stack

Alibaba's semiconductor unit T-Head (平头哥) is building out an open software stack for its own AI chips. QbitAI (量子位) reported on 23 and 24 September 2026 on the progress of the SAIL program and on the launch of the new Zhenwu V900 chip (真武 V900).

The chip: Zhenwu V900

Zhenwu V900 was presented on 22 September 2026 at the Apsara Conference in Hangzhou. Alibaba CEO Eddie Wu (吴泳铭) called it the most powerful AI chip in China today and said it delivers three times the performance of its predecessor, the M890. He also said the architecture had absorbed more than ten years of engineering work. Those performance claims come from the presentation itself, and no outside party has confirmed them. The second part of the announcement rests on firmer ground.

SAIL as a counterpart to CUDA

T-Head says it has released SDK, drivers, profiling and debug tools and documentation step by step since the opening began at WAIC in July 2026. The projects that are now open include PyTorch-for-sail, the migration tool sailify, Triton-for-sail, DeepGEMM-for-sail and FlashAttention-for-sail. The report also mentions adaptations for TensorFlow and JAX, an in-house inference engine, the communication library PCCL and a DeepEP-for-sail.

SAIL connects PyTorch to the Zhenwu hardware. It does the same job as CUDA in Nvidia's ecosystem: it gives existing models a layer to run on without deep reworking. Lu Shenghua (陆生华), senior director for the software ecosystem at T-Head, named migration costs as the biggest hurdle.

Who uses the chips

According to the company, the Zhenwu chips serve more than 650 customers in over 20 industries. The report says Ant, the platform Xiaohongshu and the carmaker XPeng already rely on SAIL. Xiaohongshu has reportedly built an agent that ports models for generative recommendation systems to the new environment. XPeng says it moved GPU-based models for training driver assistance systems into a Zhenwu cloud cluster. On the model platform ModelScope, T-Head published 39 quantized models from vendors such as Qwen, DeepSeek and Kimi. By September 2026, those models had drawn more than 348,000 downloads in total.

The assessment

Makers of AI software now have a second target system to think about, one that could eventually squeeze licensed and paid alternatives. The real question is whether toolchains and libraries can keep pace with framework development. That is where the work remains unfinished. Migrating existing code is still the biggest cost, and no software package makes it disappear with a single command.

The chip alone does not decide the outcome. The stack beneath it does. That is the lesson SAIL addresses.