Alibaba has launched its most powerful in-house AI accelerator, the Zhenwu M890, alongside a new flagship large language model, deepening China's drive for homegrown alternatives to Nvidia hardware. The company says the chip delivers three times the performance of its predecessor and is built for the memory-intensive demands of agentic AI.
Inside the Zhenwu M890
The M890 is based on Alibaba's in-house PPU (Parallel Processing Unit) architecture with a Transformer core engine. It carries 144 GB of on-chip HBM3 memory, a 50% increase over the 96 GB in the older Zhenwu 810E, and offers 800 GB per second of inter-chip bandwidth. Alibaba lists precision formats ranging from FP32 down to FP4, aiming to cover both high-accuracy training and cheaper inference on the same hardware family.
A system, not just a chip
Alibaba is positioning the M890 as part of a wider cluster design rather than a standalone processor. Its ICN Switch 1.0 fabric is rated at 25.6 terabits per second across clusters of 64 accelerators, and a 128-card AI supernode links chips with latency measured in the hundreds of nanoseconds. That fabric-level pitch matters for large training jobs, where interconnect speed determines whether adding chips improves throughput or simply adds cost.
A new flagship model: Qwen3.7-Max
Alongside the silicon, Alibaba unveiled Qwen3.7-Max, the latest version of its flagship LLM. The model is engineered for advanced coding and long-running agent tasks and can reportedly operate continuously for up to 35 hours without performance degradation. The launch underscores how AI hardware and frontier models are increasingly designed in tandem, a trend echoed in large-scale physical AI deployments.
Scaling a domestic alternative
Alibaba's chip subsidiary T-Head says it has already shipped more than 560,000 Zhenwu units to over 400 external customers across 20 industries, evidence that its silicon program has commercial scale. The company plans to follow the M890 with the Zhenwu V900 in 2027 (216 GB of memory, 1,200 GB/s bandwidth) and the J900 in 2028. The push lands amid intensifying competition in AI chip manufacturing and rising policy attention to the sector, including moves by the European Parliament on AI and robotics oversight.
Caveats
Many of the headline figures are company-supplied. Alibaba has disclosed memory capacity, supported numeric formats and relative performance versus its prior chip, but not absolute FLOPS, HBM bandwidth, process node, power draw or independent MLPerf results. Buyers will need to test software compatibility and operating cost before treating the M890 as a proven substitute for constrained Nvidia supply.
Reporting based on coverage from CNBC, WinBuzzer and WCCFTech.