Alibaba is pushing its AI ambitions to a new scale. At the Apsara Conference in Hangzhou on September 22, 2026, the Chinese tech group said future Qwen generations could reach 5–10 trillion parameters, unveiled the Zhenwu V900 AI chip, and set a target to lift its global data-center capacity past 20 GW by 2032.

Alibaba is building a whole “AI stack”

At Apsara, Alibaba did not just announce a single model or chip; it laid out an AI strategy spanning chips, cloud infrastructure, foundation models, AI agents and applications. CEO Eddie Wu said the company plans to train future models at 5–10 trillion parameters — a sizable jump from today's flagship, Qwen3.8-Max (about 2.4 trillion parameters).

One thing to be clear about: Alibaba has NOT released a 10-trillion-parameter model. Qwen 4 (the just-announced Max, Flash, Plus and 27B variants) is still in training; the 5–10 trillion figure is on the roadmap for future Qwen 4.5 and Qwen 5.

How big is 10 trillion parameters?

Using Qwen3.8-Max (2.4 trillion parameters) as a baseline, the 5–10 trillion Alibaba is aiming for is roughly two to four times as many parameters. Parameters are, simply put, the values a model learns during training and uses to produce results. But more parameters do not automatically mean better AI: data quality, model architecture, training, inference and compute all shape the final capability. What's more notable is that bigger models drive huge demand for chips, memory, networking and data centers — which is the second part of Alibaba's strategy.

Zhenwu V900: Alibaba's new AI chip

Alibaba's chip unit T-Head unveiled the Zhenwu V900, a processor for both AI training and inference. Per Alibaba's stated specs, the V900 has 216 GB of memory, 1,200 GB/s of inter-chip bandwidth, support for FP8 and FP4 formats, performance the company claims is three times the Zhenwu M890, and systems that can scale to clusters of about 500,000 cards. Alibaba plans mass production and commercial availability in Q1 2027; per the company, its Zhenwu chips already serve more than 650 customers. Note that the “three times” figure is Alibaba's own performance claim, not an independent benchmark.

Why does Alibaba want to build its own AI chips?

A trillion-parameter model needs enormous compute. Meanwhile, U.S. export restrictions have kept Nvidia from selling some of its most advanced AI chips to China, pushing Chinese companies to accelerate domestic hardware. Alibaba is therefore moving toward a relatively closed loop: Zhenwu chips → compute → Alibaba Cloud → Qwen models → AI agents → applications. Rather than competing only at the model layer, it wants to control more layers of the infrastructure needed to run AI.

Over 20 GW of data centers by 2032

The third piece is compute capacity. Alibaba targets total capacity across Alibaba Cloud-operated data centers worldwide to exceed 20 GW by 2032. Eddie Wu argued that AI compute demand is rising very fast, while the infrastructure supply chain is currently a limiting factor on how quickly it can expand. The figure shows the AI race shifting from “who has the better model?” to “who has enough chips, power, data centers and cloud to run AI at scale?”.

Alibaba also wants AI to improve AI

Another notable detail is Recursive Self-Improvement (RSI) based on experimental feedback. According to Alibaba, Qwen3.8-Max ran 33 automated improvement iterations over more than a month — covering pipeline design, data checks, testing and error diagnosis; in one chip-design trial, the system made more than 10,000 EDA tool calls over 60-plus hours. It's a direction worth watching: AI is increasingly taking part in research and engineering, not just generating content or answering questions.

Why this announcement matters

Three seemingly separate announcements are actually connected: 5–10 trillion-parameter models need more compute → compute needs AI chips → chips need large-scale data centers. Alibaba is betting on all three. The AP also noted the move amid a broader push by Chinese firms like Alibaba and Huawei to develop domestic AI chips as access to Nvidia's most advanced hardware is restricted. The bigger story isn't just “Alibaba is about to have a 10-trillion-parameter model,” but that it is trying to build a closed AI ecosystem from chip → cloud → model → agent.

30 seconds to get it

  • What did Alibaba announce? A future Qwen roadmap at 5–10 trillion parameters, the Zhenwu V900 AI chip, and a plan for over 20 GW of data centers by 2032.
  • Has the 10-trillion model launched? No. Qwen 4 is in training; the 5–10 trillion scale is on the Qwen 4.5/Qwen 5 roadmap.
  • How powerful is the Zhenwu V900? Per Alibaba, three times the M890, with 216 GB of memory, commercializing in Q1 2027.
  • The most important point? Alibaba isn't just building models — it's building a full-stack AI: chip, compute, cloud, model and agent.

Frequently asked questions

  • Has Alibaba released a 10-trillion-parameter model? — No. It's a roadmap target for Qwen 4.5 and Qwen 5; Qwen 4 is currently in training.
  • How big is Qwen3.8-Max? — About 2.4 trillion parameters; the 5–10 trillion goal is roughly two to four times that.
  • What are the Zhenwu V900's specs? — Per Alibaba: 216 GB memory, 1,200 GB/s inter-chip bandwidth, FP8/FP4 support, 3x the M890 (company claim), commercializing in Q1 2027.
  • Why is Alibaba building its own AI chips? — U.S. export limits restrict advanced Nvidia chips into China, pushing the company toward hardware self-reliance to run ever-larger models.