Alibaba Pushes AI Independence With M890 Chip and Qwen 3.7-Max

Alibaba is moving deeper into full-stack AI with a new accelerator and model built to work together. At the Alibaba Cloud Summit, the company introduced the Zhenwu M890, a training-and-inference AI chip from its T-Head semiconductor unit, alongside Qwen 3.7-Max, a model designed for long-running agentic workloads.

The launch is not only about performance. It is also about independence. Alibaba, like other major Chinese cloud and AI vendors, is trying to reduce its reliance on Nvidia GPUs by building more of the hardware, model, and cloud layer inside its own ecosystem.

Alibaba Wants More Control Of The AI Stack

The Zhenwu M890 is aimed at AI agents, especially workloads that require large memory, long context windows, and multiple models communicating with one another. Alibaba says the chip is built for both training and inference, making it part of a broader strategy to optimize AI systems from silicon to software.

Qwen 3.7-Max is the model meant to show why that matters. Alibaba says it can run on the M890 for up to 35 hours, perform extended reasoning, handle more than 1,000 tool calls, and work with a 1-million-token context window. The company is pitching it for demanding software tasks such as multi-file code editing, refactoring, and prototyping.

That combination matters because AI agents are increasingly bottlenecked by memory, context, tool use, and cost. A model that can reason for longer periods is only useful if the underlying infrastructure can support the workload efficiently.

Nvidia Dependence Is The Strategic Pressure

The release comes as Chinese AI companies continue looking for ways to operate with less exposure to U.S. chip policy and Nvidia supply constraints. Although exports of some advanced processors remain available, local vendors such as Alibaba, Baidu, and Huawei are pushing domestic alternatives more aggressively.

For Alibaba, the M890 is both a cost strategy and a differentiation strategy. If Alibaba Cloud can run more workloads on its own chips, it can tune performance for its own models, reduce dependence on outside suppliers, and give enterprise customers a more integrated platform.

That mirrors the direction taken by other hyperscalers. Google has TPUs, AWS has Trainium and Inferentia, and Baidu has also pursued custom AI silicon. The logic is simple: as AI workloads become core cloud infrastructure, owning the chip layer gives cloud providers more control over cost, availability, and product direction.

The Hard Part Is Manufacturing Scale

Alibaba still faces a difficult road. China’s chip supply chain remains weaker than the global ecosystem used by Nvidia and other leading AI hardware suppliers. Even if Alibaba can design useful accelerators, producing them efficiently and at scale is a separate challenge.

That limits how far the savings can go. A company can reduce reliance on outside GPUs, but if its own chips are less efficient, constrained by manufacturing, or harder to deploy broadly, the economic advantage may be smaller than the strategy suggests.

Still, the M890 and Qwen 3.7-Max show where Alibaba wants to compete. The company is not just releasing another model. It is trying to build the hardware-model pairing that makes agentic AI cheaper, longer-running, and more controllable inside its own cloud.