Alibaba's T-Head Open-Sources More of Its Zhenwu AI Chip Software Stack
The AMW Read
T-Head's open-sourcing of its CUDA-like Zhenwu software stack, backed by named production customers Ant Group, XPeng, and Xiaohongshu, marks a shift from closed chip delivery to ecosystem-building against Nvidia's software moat.
Alibaba's T-Head Open-Sources More of Its Zhenwu AI Chip Software Stack
On September 22 at Alibaba's Yunqi conference, CEO Eddie Wu unveiled the Zhenwu V900 chip, calling it China's most powerful AI accelerator at roughly 3x the prior M890. A day later, chip unit T-Head (平头哥) expanded open-source access to T-Head SAIL, its CUDA-like stack linking PyTorch to Zhenwu hardware, adding the sailify migration tool, Triton-for-sail, and accelerators DeepGEMM-for-sail and FlashAttention-for-sail. T-Head says Zhenwu serves 650-plus customers across 20-plus industries, with Ant Group, Xiaohongshu, and XPeng Motors in production; Xiaohongshu built its own migration agent on the open code. T-Head lists 39 quantized Qwen, DeepSeek, and Kimi models on ModelScope with 348,000-plus downloads to date.
The release targets migration cost, not raw throughput — the real barrier to switching chips. T-Head's software lead says customers' first question is how much existing code and tuning experience carries over, similar to the lock-in behind Nvidia's CUDA dominance. Open-sourcing migration and kernel tools lets customers tune their own operators instead of routing requests through Alibaba, while adaptations get pushed upstream into PyTorch, vLLM, and Triton so support tracks mainstream releases rather than forking. This sits inside Alibaba's stated AI models-chips-cloud commitment, following the roughly RMB 380 billion (~$53B) three-year infrastructure plan disclosed in 2025.
For teams evaluating non-Nvidia capacity in China, switching cost is lowest for PyTorch-based workloads, since sailify targets porting existing models, not rewrites. Judge readiness by named customers, not spec claims: Ant's inference work on Zhenwu 810E/M890 and Xiaohongshu's migration agent are the evidence. Per the AI Market Watch index, name-matched T-Head coverage nearly doubled to 7 items in the past 90 days versus 4 prior, consistent with an accelerating cadence.