Alibaba used its Apsara conference on Sept. 22 to show how far it wants to push beyond being a cloud host and become a vertically integrated AI supplier. In an Alibaba Cloud announcement, the company said Qwen 4 is in training, mapped later Qwen 4.5 and Qwen 5 systems at 5 trillion to 10 trillion parameters, introduced the Zhenwu V900 accelerator, and set a goal for more than 20 gigawatts of global data-center capacity by 2032. It also unveiled AgentCore for enterprise agents and a Qwen-based agent platform for smartphones.
That matters because the most practical AI competition in China is no longer only about who has the biggest model or the fastest chip. It is about who can still assemble enough compute, power, software and operational trust to keep training models and putting them into business use when access to some foreign AI hardware is restricted. Reuters reported Alibaba shares rose 5.1% on the day, a sign that markets heard the same message.
The immediate answer for customers and investors is mixed. The roadmap is strategically significant because it could make Alibaba a more resilient domestic AI platform and pull more workloads into its cloud. But it does not erase the hard limits underneath AI: foundry capability, manufacturing yield, electricity, networking, cooling, software reliability and customer confidence that agents can be deployed safely.
From chip scarcity to stack control
The appeal of a full stack is coordination. A chip designed for Alibaba’s own supernodes, tied to its own model training and inference software, can be tuned for the workloads Alibaba expects to run most. That can improve utilization and reduce dependence on imported accelerators, especially for domestic customers weighing long-term access risk. AP reported Alibaba presented the V900 as China’s most powerful AI chip, which is as much a statement about self-reliance as about speed.
Alibaba says the V900 has 216 GB of GPU memory, 1,200 GB/s of inter-chip bandwidth, and support for FP8 and FP4 workloads. It says the chip delivers three times the performance of the Zhenwu M890 launched in May 2026, with mass production and commercial release targeted for the first quarter of 2027. Alibaba also says Zhenwu chips already serve more than 650 customers. On the model side, its current flagship Qwen3.8-Max is reported by Reuters and AP at 2.4 trillion parameters, while Qwen 4 is still training and later Qwen 4.5 and Qwen 5 systems are projected to reach 5 trillion to 10 trillion parameters.
Those numbers mostly show ambition and integration, not finished advantage. Alibaba says Qwen3.8-Max completed 33 automated training cycles over a month and that, in a separate chip-design experiment, it ran for more than 60 hours, made more than 10,000 electronic-design-automation tool calls, and reduced chip area by 42% without hurting performance. The point is less that parameter counts or internal experiments settle the race, and more that Alibaba is trying to automate more of the engineering around chips and models at once.
The bottleneck may be moving, not disappearing
The harder question is where the constraint lands next. Building a domestic stack can reduce dependence on foreign accelerators, but it can also expose every other weak link more clearly. Alibaba’s own roadmap is a reminder that frontier AI is now a physical-infrastructure business: more memory, more high-speed networking, more data-center construction, more cooling, more trained operators, and much more power.
That is why the 20-gigawatt target matters as much as the model headlines. If Alibaba can add enough capacity and bring AI supernodes online at commercial scale, it can compete through system economics: more available compute, potentially tighter integration, and a clearer path from training to deployment. But the public record still leaves large blanks. There is no independent benchmark set showing how V900 compares with Nvidia, Huawei or other accelerators on common workloads, no public workload definition for the three-times-performance claim, and no disclosed price, yield, power-consumption or detailed availability schedule. A supernode configuration of up to 500,000 cards is a meaningful design statement, not proof that customers can yet rent a cluster at that scale.
Foundry capability is the other unresolved gate. As the AP reported, Counterpoint analysts argued that chip-design progress does not by itself close the U.S.-China gap if China’s semiconductor manufacturing remains weaker, and that Chinese frontier training has still often relied on Nvidia hardware. Alibaba has also said global AI data-center supply shortages are limiting its ability to scale. In other words, vertical integration can change who controls the system, but it cannot repeal the physics or the supply chain.
Where enterprise buyers will judge it
For business customers, the agent layer is where this roadmap stops being industrial policy and starts becoming a purchase decision. AgentCore is Alibaba’s attempt to give enterprises a managed way to build, run, monitor and secure AI agents. The smartphone agent platform aims at a different surface, but the logic is similar: if Alibaba can combine its own models, cloud infrastructure and tooling into a reliable runtime, it has a better chance of owning real workloads rather than just selling raw compute.
That opportunity comes with a tradeoff. A more optimized stack can also be a less portable one. Enterprises care about actual availability, price per useful token, latency, data residency, security controls, support and how safely agents connect to business systems. They also care about whether they can move workloads later if costs rise or performance disappoints. Independent security and reliability evaluations for AgentCore, the Agent Security Center and the smartphone agent platform are not yet public, so trust will have to be earned in production.
That makes Alibaba’s Sept. 22 roadmap important, but not yet decisive. It materially changes the conversation by shifting attention from access to one foreign chip supplier toward control of a whole AI system. For Chinese customers facing export-control uncertainty, that alone has value. The durable advantage, though, will be measured less by trillion-parameter headlines than by whether V900 ships on time, whether capacity gets financed and powered, whether software stays compatible, and whether enterprises trust the agents enough to run meaningful work on them. If Alibaba clears those hurdles, vertical integration becomes a competitive weapon. If it does not, the bottleneck has merely moved a few layers down the stack.
By
By
By
By
By
By
