Alibaba unveils new AI chip and 20-gigawatt data center plan to challenge Nvidia

Alibaba Group used its annual Apsara Conference in Hangzhou on Tuesday to unveil the Zhenwu V900, a new artificial-intelligence accelerator the company describes as the most powerful chip built in China to date, alongside a plan to expand global data-center capacity to 20 gigawatts by 2032. The announcements, delivered by chief executive Eddie Wu in a keynote framed around what he called the "Machine Intelligence" era, position Alibaba as one of the most aggressive Chinese challengers to Nvidia's dominance in AI computing.

The chip comes from T-Head, Alibaba's semiconductor division, and was shown alongside a broader hardware stack, including a new interconnect switch, a smart network card and a solid-state drive controller, that the company said is designed to let hundreds of thousands of chips operate as a single system. Alibaba said the conference, which runs through Thursday, would center on what it calls a three-part strategy spanning AI models, AI chips and AI cloud infrastructure.

The numbers

According to specifications reported by TrendForce's supply-chain research desk, the Zhenwu V900 carries 216 gigabytes of memory and 1,200 gigabytes per second of chip-to-chip bandwidth, with native support for the low-precision FP8 and FP4 formats now standard in large-model training. Alibaba said the chip delivers roughly three times the performance of its predecessor, the Zhenwu M890, and that a new in-house interconnect chip called ICN Switch can link up to 500,000 V900 units into a single training cluster, alongside companion parts for storage and networking that make up what the company calls its Apsara Supernode stack. Mass production and commercial sales are targeted for the first quarter of 2027, meaning the chip is a roadmap commitment rather than something customers can buy today.

Wu also set a goal of 5 to 10 trillion parameters for Alibaba's next flagship Qwen model, several times the scale of systems the company currently runs in production, where the M890-based "AI Supernode" already supports inference for models exceeding 2 trillion parameters. He said Alibaba Cloud aims to operate more than 20 gigawatts of data-center capacity worldwide by 2032, up from a fraction of that today. That buildout sits on top of a spending commitment of more than $53 billion over three years that Alibaba disclosed previously, and which it partly funded with a $10.2 billion follow-on share sale in Hong Kong in August. Alibaba's Hong Kong-listed shares closed about 2% higher on Tuesday, and its U.S.-listed stock added more than 3% in premarket trading, according to a market wrap from Yahoo Finance. Alibaba also said its open-weight Qwen-27B model has become one of the most widely downloaded open models among outside developers, a data point it is using to argue that hardware and model gains are reinforcing each other.

How we got here

The launch lands almost exactly three years into a tightening regime of U.S. export controls that has barred Nvidia and other American chipmakers from selling their most advanced AI accelerators to Chinese customers. Washington has eased and re-tightened the rules repeatedly since, most recently sparring over a loophole that let some restricted chips reach China through third countries, a dispute that remains unresolved heading into this week's diplomacy between Washington and Beijing. With Nvidia's newest hardware effectively off-limits, Chinese cloud providers and AI labs have leaned on a mix of stockpiled Nvidia chips, the H20 model still cleared for Chinese sale, and homegrown silicon from Huawei, Cambricon and Alibaba's own T-Head unit, which was founded in 2018 and has previously released the Hanguang 800 inference chip and the Yitian 710 server processor.

Alibaba is a comparatively small player in that domestic hardware race today. Huawei's Ascend line and Cambricon's accelerators together account for the large majority of China's AI-chip shipments this year, with Alibaba trailing both, and Alibaba has instead built its reputation on cloud services and on the Qwen family of models. The Zhenwu V900 is Alibaba's attempt to close that hardware gap without giving up its model and cloud advantage, by building chips tuned specifically to run its own software efficiently rather than competing as a general-purpose merchant chip supplier.

"Thinking will become as abundant and accessible as power," Wu told the Apsara audience, comparing the current buildout of AI infrastructure to the early, halting years of electrification, when the first commercial power plant lit only a few hundred bulbs before the technology reshaped industry decades later.

Who is affected

The most direct competitive target is Nvidia, which built its dominance of AI computing on chips that are now largely barred from the Chinese market it once supplied heavily. A credible in-house alternative from a hyperscaler the size of Alibaba narrows the opening for Nvidia to re-enter later even if controls ease, since customers who have already redesigned their software around domestic chips have less reason to switch back. Alibaba Cloud's enterprise customers, which the company says now number more than 650 across sectors including autonomous driving, finance and manufacturing, stand to get access to the new hardware once it ships, and Chinese firms shut out of Nvidia's top-end chips gain another domestic option beyond Huawei and Cambricon. Developers building on Alibaba's Qwen models gain a company more explicitly committed to training far larger systems, while employees and outside investors in the T-Head unit are watching a possible spinoff: Alibaba has separately been preparing to carve out the chip division and explore an initial public offering for T-Head, a move that would let the market price the chip business on its own rather than as a line item inside Alibaba's cloud unit.

Reaction and what happens next

Alibaba's own account of the keynote, published by its corporate newsroom, cast the announcements as the start of a multi-year build-out rather than a finished product line. In a recap of Wu's keynote, the company said machine "thinking" currently amounts to less than 3% of total human thinking capacity but argued that share would eventually invert, with machines handling the overwhelming majority of cognitive work. Alibaba also used the event to introduce Qwen Intelligence, a mobile-device product meant to let phones carry out multistep reasoning tasks locally, and to lay out a roadmap for new Yitian server processors due in the third quarter of 2027.

The near-term test is execution: the Zhenwu V900 will not enter mass production until the first quarter of 2027, giving Nvidia, Huawei and Cambricon more than a year to extend their own roadmaps before Alibaba's chip reaches customers at scale. Alibaba has a mixed record translating chip announcements into shipped volume, and outside verification of the V900's claimed performance will not be possible until independent benchmarks appear closer to launch. The timing also puts the announcement in a wider geopolitical frame. It lands just ahead of an expected meeting between President Trump and Chinese leader Xi Jinping at which AI competition and chip export policy are anticipated to be discussed, and amid an unresolved dispute in Washington over how tightly to police the export of Nvidia's China-permitted H20 chip. How that meeting and the broader export-control fight play out could determine whether Chinese cloud providers keep pouring capital into homegrown silicon like the Zhenwu V900 or shift back toward Nvidia's hardware if access loosens.