Alibaba used its 2026 Apsara Conference to lay out a broader full-stack AI strategy spanning proprietary accelerators, the next Qwen model family, cloud infrastructure and agent tooling. Rather than treating models and hardware as separate product lines, the company is positioning them as one integrated stack.
The hardware centerpiece is T-Head’s Zhenwu V900. Alibaba says the accelerator delivers three times the performance of the earlier Zhenwu M890, with 216 GB of memory and up to 1,200 GB/s of inter-chip bandwidth. Mass production and commercial availability are scheduled for the first quarter of 2027.
What happened
Alibaba also confirmed that Qwen 4 is currently in training. Its roadmap extends to Qwen 4.5 and Qwen 5, with future models projected to scale into the 5-to-10-trillion-parameter range. That figure is a development target rather than a released specification, and parameter count alone does not establish real-world model quality.
The company described a set of automated self-improvement experiments around Qwen3.8-Max. Alibaba says a month of autonomous iteration lifted the model’s Artificial Analysis score from 40 to 45. In a separate chip-design experiment, the model reportedly ran for more than 60 hours and made over 10,000 EDA tool calls while optimizing production-grade bus modules. Those measurements come from Alibaba and have not been presented as independent benchmark results.
The infrastructure roadmap goes beyond the V900 itself. Alibaba unveiled a supernode server combining the new accelerator with in-house networking, storage and control components, alongside future Yitian CPUs intended for agentic AI workloads. The company says its Zhenwu chips are already serving more than 650 customers across industries.
Alibaba Cloud is also reorganizing its software stack around agents. AgentCore is meant to provide a managed foundation for building and operating agents, Agent Security Center adds governance and threat controls, and Agent Context is designed to supply enterprise data and long-term memory to AI applications.
The largest infrastructure commitment is a target to operate more than 20 GW of global data-center capacity by 2032. Associated Press independently confirmed the main chip and model announcements and reported that Alibaba uses Zhenwu hardware in its data centers for both internal workloads and cloud customers.
Why it matters
The strategic significance is vertical integration. A cloud provider that controls chips, servers, networking, storage, models and agent runtimes can optimize the stack end to end and reduce dependence on outside accelerator suppliers.
That matters particularly in China, where access to some leading U.S. AI accelerators remains constrained. A domestic chip does not automatically match the efficiency or software maturity of global market leaders, but it gives Alibaba more control over capacity planning and product economics.
The near-term test will be execution rather than headline specifications. Independent performance data, commercial availability of the V900 and actual deployment costs will determine how competitive the stack is. For now, the announcement shows Alibaba moving aggressively toward a self-contained AI platform from silicon through agent software.
