Alibaba has unveiled an AI processor designed specifically for long-running agents, pairing the hardware with a multi-year chip roadmap and a new Qwen model. The combination shows that the company is no longer treating US export restrictions as a temporary procurement obstacle. It is building an integrated domestic platform.
The Zhenwu M890, developed by Alibaba’s T-Head semiconductor unit, is claimed to deliver three times the performance of its predecessor. More significant is its intended workload: agents that retain long context, coordinate models and execute multi-step tasks with limited intervention.
Those demands place pressure on memory bandwidth and communication rather than simply rewarding the same architecture used for isolated inference requests.
Agent workloads reshape accelerator priorities
A conventional inference chip may process a prompt and return an answer. An enterprise agent can remain active for hours, call tools, exchange context with other models and maintain state through a changing task.
Alibaba says the M890 is built for those extended patterns. The chip will be offered through its domestic Bailian model platform inside the Panjiu AL128, a rack system combining 128 M890 accelerators.
T-Head already has a material deployment base. It reports shipping more than 560,000 Zhenwu units, with over 400 external customers across 20 industries, including automotive and financial services.
That footprint provides operational data that a laboratory programme lacks. Customers expose the chip to real software, utilisation and reliability demands, allowing the roadmap to respond to production rather than benchmark behaviour alone.
A three-generation roadmap signals long-term commitment
Alibaba plans to follow the M890 with the V900 in the third quarter of 2027, targeting another roughly threefold gain, and the J900 in the third quarter of 2028.
The cadence resembles the predictable upgrade cycles that helped Nvidia build customer confidence. Enterprises designing large systems need to know that a platform will improve and remain supported beyond one product generation.
Huawei has outlined a similar path for its Ascend accelerators. Both companies are responding to the structural risk of depending on foreign components whose availability can change through export policy.
Alibaba committed more than 380 billion yuan, approximately $53 billion, to cloud and AI infrastructure over three years. Developing successive chips, servers, software and capacity turns that spending into a domestic capability rather than a one-off substitute.
The result could outlast the restrictions that accelerated it. Once engineers, customers and applications optimise around a local architecture, switching back to an external supplier becomes a strategic choice rather than the default.
Chip, model and cloud form a closed feedback loop
Alibaba announced Qwen 3.7-Max alongside the processor. The model is presented as suited to coding and long-running agent tasks, with the ability to operate continuously for up to 35 hours without degradation.
Launching both for the same workload creates a platform play. T-Head controls the silicon, Qwen provides the model, Bailian delivers cloud access and Panjiu supplies the server architecture.
Co-design can improve utilisation because each layer is tuned with knowledge of the others. It also gives Alibaba direct feedback from enterprise deployments and reduces reliance on external roadmaps.
Customers should still evaluate the costs of a vertically integrated stack. Deep optimisation may limit portability, while claimed performance must be tested on actual agent workflows, including tool calls, memory use and failure recovery.
The M890 is not simply evidence that China can replace restricted Nvidia products. It represents a view that agentic computing will require a different class of infrastructure and that control over every layer will become a commercial advantage.
More than half a million prior shipments give the strategy credibility, while announced successors demonstrate intent. If Alibaba delivers the cadence and developer ecosystem, export controls may have done more than encourage an alternative supplier. They may have helped create a durable platform competitor designed around the next generation of AI workloads from the beginning.
