Alibaba used its September 22 Apsara Conference appearance in Hangzhou to lay out a full-stack AI roadmap that runs from foundation and multimodal models through chips, cloud infrastructure, and enterprise agent services. The announcement, detailed by Alibaba Cloud, was a company statement about future products, capacity, and performance. It was not an independently audited account of deployments already operating at the stated scale.
The breadth of the plan matters because Alibaba presented the pieces as a connected stack. Qwen development sits at the model layer. T-Head’s processor work addresses computing hardware. Alibaba Cloud’s architecture covers training, inference, enterprise agent operations, security, real-time data, and long-term memory. The roadmap therefore describes a path for the company to coordinate software, silicon, and infrastructure rather than treating agents as a single application feature.
Alibaba’s announcement also extends a sequence of EastFrontier coverage around the company’s agent strategy. The earlier Alibaba Cloud agent control product layer story focused on the control function. The September 22 plan places AgentCore inside a broader Agent Native Cloud layer alongside the hardware and model systems that Alibaba says will support enterprise deployment.
Qwen’s 10-Trillion-Parameter Roadmap Raises the Compute Question
Alibaba said Qwen 4 was in training and projected a roadmap in which Qwen 4.5 and Qwen 5 could scale to between 5 trillion and 10 trillion parameters. Those figures describe Alibaba’s roadmap statements, not confirmed release specifications, and the company did not announce that any of the models had been released. The relevant point in the announcement is the intended direction: larger foundation and multimodal models are being planned alongside the infrastructure Alibaba says they will require.
A model roadmap at that scale has direct consequences for the rest of the stack described at the conference. Training and inference need processors, high-bandwidth connections, storage, and systems for handling model context. Alibaba did not present the Qwen figures as an independent benchmark or completed deployment. They remain future-facing company plans, which makes the accompanying hardware and data-center targets part of the same stated strategy rather than evidence that the full system is already in place.
The company’s cloud target is similarly prospective. Alibaba Group CEO Eddie Wu said Alibaba Cloud aims for its globally operated data-center capacity to exceed 20GW by 2032. That is a future capacity target, not current capacity and not proof of completed construction. It gives the roadmap a long time horizon and ties the model program to a planned expansion of the physical facilities used to run cloud services.
This infrastructure emphasis follows a pattern visible in Alibaba Cloud’s European data-center map, but the September 22 announcement is broader than geographic expansion. It links global capacity with model training, inference, and agent operations. The available evidence does not establish how the 20GW target will be funded, where the additional capacity will be built, or how much is already available, so those questions remain outside the announcement’s verified claims.
T-Head’s Zhenwu V900 Adds a Planned Silicon Layer
Alibaba’s chip-design unit T-Head announced the Zhenwu V900 AI processor. Alibaba characterized it as a threefold performance step up from the Zhenwu M890, alongside 216 GB of GPU memory and 1,200 GB/s of inter-chip bandwidth. These are Alibaba’s reported product metrics, not independently validated results. T-Head scheduled mass production and commercial release for the first quarter of 2027, making the processor a planned component of the roadmap rather than a currently available product under these source limits.
The V900 announcement also described the surrounding server design. Alibaba said an upgraded supernode integrating the Zhenwu V900, ICN Switch, Panmai SmartNIC, and Zhenyue SSD controller could support a cluster of up to 500,000 cards. That number is a stated platform capacity, not evidence that a cluster of that size is operating or commercially deployed. It shows how Alibaba framed the chip as part of a system that includes switching, networking, and storage control rather than as a standalone accelerator.
The design choice is consistent with the roadmap’s full-stack logic. A processor can supply compute, but a large model system also depends on memory, inter-chip communication, network ports, and data access. T-Head’s V900 figures and the supernode description were presented with cloud and model plans, positioning the hardware as infrastructure for Alibaba’s wider architecture.
Alibaba Cloud also reported product metrics for the surrounding infrastructure. Its Cloud Parallel File Storage was described as reaching the hundred-terabyte-per-second throughput range and the hundred-million-IOPS range, with a claimed 69% reduction in enterprise AI storage costs. HPN 8.0 Pro was reported at 100 petabits of bandwidth and more than 130,000 800G network ports in one cluster. None of those figures was independently benchmarked in the supplied materials.
The hardware story is about integration as much as speed. The V900, ICN Switch, Panmai SmartNIC, and Zhenyue SSD controller form a system around the processor. Storage and networking metrics sit beside it. Alibaba’s roadmap presents these parts as ways to handle data and model work across large infrastructure, while the available evidence supports no claim that they have displaced other hardware or established a market position.
Alibaba Cloud Defines Three Layers for Enterprise Agents
Alibaba Cloud described its agentic-cloud architecture in three layers. AI Native Cloud is intended for model training and inference. Agent Native Cloud is for enterprise agent deployment, operation, and security. Context Engine addresses real-time data and long-term memory. Together, the layers describe how Alibaba Cloud wants enterprises to move from a model endpoint to agents that can operate with data and managed controls.
AgentCore sits within the Agent Native Cloud layer. Alibaba Cloud introduced it as an enterprise platform to build, run, and manage AI agents across their lifecycle. The company also described an Agent Security Center for security and compliance management. AgentCore is one component of the broader architecture, not a synonym for the full-stack roadmap. That distinction matters because the September 22 announcement also covered Qwen, T-Head processors, data-center capacity, storage, and networking.
Alibaba Cloud reported that Agent Context could reduce token usage by up to 67%. This is a company-reported product metric, not an independent benchmark. In the architecture Alibaba described, context is a distinct infrastructure concern because agents may need real-time information and long-term memory while they operate. The reported metric indicates the company is presenting context management as an efficiency feature within the agent stack, not merely as a model capability.
The roadmap also connects to Alibaba’s recent physical-device push. EastFrontier’s report on how Alibaba’s Apsara Conference brought AI agents to more physical devices covered a related direction, while this announcement centers on the infrastructure underneath enterprise agents. The two themes should not be collapsed: the September 22 announcement supports a chips-to-cloud-to-agents roadmap, with AgentCore and Agent Security Center operating inside that design.
For customers and investors, the immediate takeaway is the scope of Alibaba’s stated plan, not proof that every target has been delivered. Qwen 4.5 and Qwen 5 remain projected milestones, the 20GW figure remains a 2032 capacity target, and the V900 release is scheduled for Q1 2027. The roadmap’s next test will be whether these layers appear as products and operating capacity on Alibaba’s timelines.