Product1 publisher2 min readPublished
Alibaba unveils new chip tripling performance, sets 20-gigawatt cloud target
Alibaba says the Zhenwu V900 trebles the performance of its previous accelerator and can run in clusters of up to 500,000 units. The supply question behind those clusters went unanswered at the launch.
The Product Desk · Product desk
What happened
- Alibaba outlined a Zhenwu V900 accelerator from its T-Head division on Sept 22, which chief executive Eddie Wu said can be combined in clusters of up to 500,000 units for frontier-model training.
- Wu's target of 20 gigawatts of Alibaba Cloud data centre capacity by 2032 could drive more than US$160 billion of external revenue for the division, on Citigroup's estimate.
- The company has committed more than US$53 billion over three years to AI and raised about US$10.2 billion from a follow-on share offering in August.
- Alibaba's profit fell by more than 75 per cent in the last quarter after spending of nearly US$10 billion, while annualised revenue from AI products is expected to reach US$10 billion this quarter.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- constraint Without a named foundry or a volume, a buyer cannot plan V900 capacity the way it plans a purchase order, and the supply ceiling sits with Chinese fabs that export controls have already narrowed.
- decision Every capacity figure Alibaba attached to the chip is its own, so a team wanting this silicon is making a cloud tenancy decision and inheriting Alibaba's build schedule with it.
- cost Qwen's low prices are funded out of a capex programme whose margin cost has already landed, and Alibaba's own payback window is three years, so a tenant signing today is pricing against the investment phase.
- exposure Anyone standardising on Alibaba's models or cloud takes on a live dispute: Washington has accused the company of illicitly accessing US models, and Beijing has promised retaliation if American measures follow.
Building half a million of these chips is the open question. According to the Straits Times, Wu stopped short of addressing whether Alibaba can produce enough silicon in the short term to meet surging domestic demand, with Nvidia's most advanced parts restricted in China [11]. US export controls largely block Chinese firms from using TSMC, and it is not known whether Alibaba is working with domestic chipmakers such as the capacity-constrained SMIC [12].
The capacity number belongs to Alibaba's own build. Wu's 20 gigawatt target is for Alibaba Cloud globally, and he showcased a full stack of proprietary chips, cloud services and AI models [7][26]. SoftBank envisions a 10 gigawatt-scale computing centre in Ohio alone [9]. Alibaba's entire 2032 global goal is twice the size of that one American site [10].
The spending line is more than US$53 billion over three years [4], an average of about US$17.7 billion a year [25]. Alibaba wants to quintuple annual cloud and AI revenue to US$100 billion within five years [21], implying a base of roughly US$20 billion today [22]. Wu's team has said the company should recoup its overall investment within three years [20].
Alibaba has the world's most downloaded AI model and has kept Qwen pricing low to bring users in [17]. A tenant signing now is buying at investment-phase prices. Wu also said Alibaba plans models with between five and 10 trillion parameters for longer and more complex tasks [6].
"Last year, we asserted that the more capable AI becomes, the more powerful humanity will be," Wu said. "Today, I remain steadfast in that conviction" [14]. Alibaba's shares rose more than 5 per cent in Hong Kong after the event, and Tencent's more than 7 per cent [13]. Anthropic chief executive Dario Amodei has urged a slower pace to better handle the risks [15].
For a team deciding where to train next year, the useful split is between a design claim and a delivery commitment. A quote that names the V900 comes back either as an instance type with an hourly rate and a term length, or as a 2032 date.
What to watch
- Whether Alibaba names a foundry or a wafer volume for the Zhenwu V900.
- Whether the V900 appears as an Alibaba Cloud instance type with a published hourly price and term.
- Whether the AI dialogue Scott Bessent and He Lifeng agreed to set up changes the restrictions on Nvidia's top chips in China.