Alibaba Unveils Full-Stack AI Roadmap: Zhenwu V900 Chip and 10-Trillion-Parameter Qwen Aim at the 'Machine Intelligence' Era

Alibaba used its biggest technology event of the year, the Apsara Conference 2026, to make a full-scale push into AI infrastructure. Held in Hangzhou from September 22 to 24 under the theme "Intelligence goes beyond," the conference saw CEO Eddie Wu (Wu Yongming) present a roadmap for a "full-stack AI" strategy: designing and operating every layer of the stack in house, from semiconductor chips to cloud infrastructure, foundation models, and AI agents. At the center of it all are a proprietary AI chip, an ultra-large model, and the compute infrastructure to support both.

In-House 'Zhenwu V900' — 3x the Predecessor, 500,000 Cards per Cluster

The biggest draw was a new AI training-and-inference chip, the Zhenwu V900, developed by Alibaba's semiconductor unit T-Head. Alibaba said the chip delivers roughly three times the performance of its predecessor, the Zhenwu M890, and can connect up to 500,000 cards within a single cluster. The company described V900 as "the most powerful AI chip in China today." Mass production and commercial availability are slated for the first quarter of 2027.

With U.S. export controls on high-end AI chips to China still in place, the move to secure large-scale compute with a homegrown chip is significant. Alibaba's intent to reduce reliance on Nvidia and meet AI demand with a domestic stack shows up right at the silicon layer.

Qwen: From 2.4 Trillion to 10 Trillion Parameters

At the model layer, Alibaba disclosed plans to scale up the Qwen lineup. Its current flagship, Qwen 3.8 Max, runs at about 2.4 trillion parameters; the company said it is preparing next-generation models in the 5–10 trillion-parameter range. Future generations — Qwen 4, 4.5 and 5 — are expected to reach that scale to handle more complex, longer-horizon tasks.

Alibaba also said it is making meaningful progress on recursive self-improvement (RSI), in which a model improves its own capabilities, and is ultimately aiming toward artificial superintelligence (ASI). That, however, reads more as a statement of direction; concrete verification will only be possible once the models actually ship.

A 'Machine Intelligence' Declaration and 20GW of Compute

In his keynote, Eddie Wu framed AI's next phase as the "machine intelligence" era. "Machine power already drives 99.9% of the world's physical work," he said, adding that "machine thinking will expand even faster, ultimately shouldering 99.9% of all thinking." At the same time, he stressed that this is only the beginning: "The truly groundbreaking products of the machine intelligence era have not yet arrived."

Underpinning that vision is a massive compute build-out. Alibaba Cloud said it aims to grow the global data-center capacity it operates to more than 20GW by 2032 — an extension of the company's previously signaled ramp in AI and cloud investment, and the physical foundation for a strategy that binds chips, models and infrastructure together.

Market Reaction and What's Left to Prove

Alibaba's shares rose 5.1% on the day of the announcement, hitting a roughly one-month high as the in-house chip, ultra-large model and a clear compute roadmap stoked investor expectations. Yet the market's gaze is already shifting from "expectation" to "results." Cathy Chan, an analyst at CCB International, praised Alibaba's "disciplined execution across its full-stack AI ecosystem," while noting that investor priorities have moved from "last year's enthusiasm for AI catalysts to the need for tangible returns and infrastructure efficiency."

What It Means

The announcements show that U.S.–China AI competition is expanding beyond model performance into a vertical-integration contest that runs "from chips to agents." In an environment where access to Nvidia chips is constrained, if Alibaba can carve a path with its own stack, the self-sufficiency of China's broader AI ecosystem could rise a notch. The key is not the declaration but the verification: V900's actual production yields, the performance of a 5–10 trillion-parameter Qwen, and the ability to convert 20GW of infrastructure into revenue will be the storylines to watch over the next 12–18 months.