Alibaba Now Builds the Chip, the Model and the Cloud

At its annual Apsara (Yunqi) conference in Hangzhou this week, Alibaba made several announcements that each would have been a headline on its own. It unveiled a new accelerator. It confirmed a trillion-parameter model roadmap. It set a data-center target measured in tens of gigawatts. And it named three new countries for Alibaba Cloud.

CEO Eddie Wu described the strategy as "three pillars: models, chips and cloud infrastructure." That is the right way to read the week. Alibaba is not trying to beat Nvidia on a spec sheet. It is trying to become the one Chinese company that doesn't have to depend on anyone else for any layer of the stack.

The chip: Zhenwu V900

The Zhenwu V900 comes from T-Head, Alibaba's semiconductor unit. The company claims three times the compute of the previous Zhenwu M890. Wu called it "the most powerful AI chip in China today."

Reported specs include 216GB of memory and 1,200GB/s of inter-chip bandwidth. The chip supports native precision from FP32 down to FP4, with improved FP8 and FP4 instructions. It is designed to scale into clusters of up to 500,000 chips. Mass production begins in the first quarter of 2027.

The benchmark claim needs context. Reporting describes the M890 as roughly twice as powerful as Nvidia's H20. The H20 is the export-compliant part Nvidia cut down specifically for China, not a frontier training chip. Three times a chip that was about twice an H20 is a large step. But it is a step measured against a ceiling Washington set, and Alibaba has not published an independent comparison against Nvidia's current flagship parts.

The volume numbers matter more than the peak performance. According to reporting from the conference, the Zhenwu family has shipped 560,000 units to more than 650 external customers in automotive, finance, large-model and embodied-AI work. That is no longer a lab project. T-Head is a merchant chip business with a real installed base, and it now has a flagship that its owner's own models are meant to train on.

The model: Qwen at 5 to 10 trillion parameters

Alibaba confirmed that Qwen 4 is in training. It also laid out development paths for Qwen 4.5 and Qwen 5, which are aimed at the 5-to-10-trillion-parameter range. Wu said Alibaba is "planning to train an AI model with five to 10 trillion parameters," and cloud executive Li Feifei described the company as building the ability to train at that scale. For reference, the current flagship, Qwen 3.8-Max, is reported at 2.4 trillion parameters.

The company also pointed to Qwen's work on recursive self-improvement. Alibaba says Qwen 3.8-Max raised its Artificial Analysis score from 40 to 45 across 33 automated improvement cycles. In a chip-design test, it says the model reduced a design's area by 42% after more than 60 hours and 10,000-plus EDA tool calls. These are Alibaba's own figures. Even so, a model helping to design the silicon it will later train on is the stack-integration story in miniature.

The link to the V900 is the point. A 500,000-chip cluster is a training-scale claim, not an inference one. Alibaba is telling the market that its next frontier models are planned around hardware it controls.

The cloud: 20GW by 2032

The infrastructure target is more than 20 gigawatts of global data-center capacity by 2032. Alibaba Cloud also plans its first regions in Turkey, Finland and the Netherlands within the next 12 months, alongside more capacity in Malaysia, Germany, the UAE, France and Hong Kong.

Two details keep this from being pure ambition. First, Alibaba did not raise its capex guidance. It kept the roughly 380 billion yuan (about $53 billion) three-year AI and cloud commitment it announced in 2025, and said some capacity may come through partnerships funded as operating expenses. So 20GW is a capacity target, not a promise to fund it all from Alibaba's own balance sheet. Second, the spending is already hitting results. Alibaba's recent net profit fell 75% on AI infrastructure investment even as AI cloud revenue grew sharply.

Citi, in an analyst note, estimated that the 20GW network could support more than $160 billion in external cloud revenue by fiscal 2033. Investors liked the framing: Hong Kong shares rose about 2% and U.S. shares about 3% premarket on the news.

Why it matters

Vertical integration is the only hedge left. U.S. export controls limit Chinese access to Nvidia's best accelerators and to TSMC's leading processes. Flux has already covered how the memory shortage pushed up prices for domestic accelerators. Under those constraints, the companies that own several layers can absorb shocks that single-layer players can't. Huawei has long played this game in hardware. Alibaba is now the most credible version that owns a top-tier open-weight model family and a hyperscale cloud as well.

The European regions are the quieter story. Finland and the Netherlands put Alibaba Cloud inside the EU's data-protection regime at a moment when Chinese AI providers face growing scrutiny over where data goes. The regions are a bet that European customers will buy Qwen-backed services if the data physically stays in Europe. They also give regulators a clear place to test that promise.

The dates are the risk. The V900 does not mass-produce until 2027. Qwen 4.5 and 5 are roadmap items, and 20GW is six years out. Manufacturing is the biggest open question: coverage of the launch does not say where the V900 will be fabricated or at which process node. Until independent benchmarks and production volumes show up, the V900 is a well-specified promise. That promise is still the most coherent answer any Chinese company has given to the question of how to build frontier AI without American hardware.