Alibaba is turning its AI chip work into a broader infrastructure story. The company has unveiled the Zhenwu M890, a new accelerator from its T-Head chip unit, and framed it around one of the industry’s most expensive bets: AI agents that need to keep context, coordinate tools, and run for long periods.
According to Reuters, Alibaba says the M890 is three times faster than its previous Zhenwu 810E chip. The company also outlined a multi-year roadmap, including a V900 chip expected in 2027 and a J900 successor planned for 2028.
The bet is not just faster silicon
The important detail is the workload Alibaba is targeting. The M890 is pitched for agentic systems: software that can carry out multi-step tasks with less human supervision. Those systems put pressure on memory, interconnects, and scheduling because they often need long context windows, tool calls, and coordination across models.
That makes Alibaba’s announcement more than a simple chip refresh. It is an attempt to line up the full stack: chips, rack systems, cloud services, and models. Reuters reported that Alibaba introduced a Panjiu AL128 rack with 128 M890 accelerators, available to Chinese enterprise customers through Alibaba Cloud’s Bailian platform.
Alibaba also announced Qwen 3.7-Max, a flagship model aimed at advanced coding and long-running agent tasks. The company claims it can operate for up to 35 hours without performance degradation. That is a vendor claim, not an independent benchmark, but it shows where Alibaba wants the market to look: persistent AI work, not one-off chatbot responses.
China’s Nvidia problem is becoming a product strategy
The timing matters. U.S. export restrictions have limited Chinese companies’ access to the most advanced AI processors. That has pushed local cloud providers and model builders to treat domestic chips as a strategic requirement rather than a side project.
CNBC reported in April that Alibaba and China Telecom launched an AI data center using 10,000 Alibaba-developed Zhenwu chips, with plans to scale to 100,000. That context makes the M890 look like the next step in a deployment plan, not a lab demo.
Alibaba has also pledged more than 380 billion yuan, about $53 billion, for cloud and AI infrastructure over three years. The money is going into a market where AI capacity is increasingly political, local, and supply-constrained.
What enterprises should watch
For Chinese customers, the appeal is obvious: more local supply, fewer export-control surprises, and tighter integration with Alibaba Cloud. For global AI infrastructure watchers, the question is whether domestic Chinese accelerators can close enough of the performance and software gap to support real production workloads at scale.
Raw chip speed is only one part of that answer. Nvidia’s advantage is also CUDA, developer familiarity, networking, system integration, and a large ecosystem of optimized tools. Alibaba’s route is different: build a vertically integrated stack inside its cloud, then make the hardware less visible to the customer.
That can work if the platform is good enough. It also makes switching costs higher. If models, chips, and cloud services are tuned together, customers may get better performance inside Alibaba’s environment but less portability outside it.
The practical read
The M890 does not mean China has solved its AI chip bottleneck. It does suggest that the workaround is maturing. Alibaba is no longer talking only about chips as substitutions for restricted hardware. It is packaging them around the next wave of AI workloads and giving enterprises a reason to care beyond national supply chains.
The real test will be boring and important: uptime, developer tooling, price per useful task, and whether customers can run agent workloads reliably at scale. If Alibaba can deliver there, the domestic AI chip push becomes less about catching Nvidia on spec sheets and more about owning enough of the stack to compete on economics.
