China's Most Powerful AI Chip and a 20-Gigawatt Bet: Inside Alibaba's Full-Stack Apsara Announcement

At Alibaba Cloud’s annual Apsara Conference in Hangzhou on Tuesday, the company made its most aggressive infrastructure play yet: a new AI accelerator it calls China’s most powerful domestically developed chip, a roadmap toward frontier-scale models with up to 10 trillion parameters, and a commitment to operate more than 20 gigawatts of global data center capacity by 2032.

The announcements, delivered by Chairman Joe Tsai and CEO Eddie Wu Yongming, amount to a coordinated bet across every layer of the AI stack — silicon, cloud, foundation models, and on-device intelligence — at a moment when Chinese technology firms are racing to secure computing capacity and reduce reliance on foreign technology amid US export restrictions on advanced AI chips.

The Zhenwu V900: triple the performance, half a million cards per cluster

The centerpiece is the Zhenwu V900, the latest AI processor from Alibaba’s chip unit T-Head. The company says the V900 delivers three times the computing performance of its predecessor, the Zhenwu M890, which was released only in May. SCMP reports the chip features 216 gigabytes of high-bandwidth memory and inter-chip bandwidth of 1.2 terabytes per second — figures that put it in the same conversation as top-tier accelerators from Nvidia and Huawei when it comes to feeding large training clusters.

Just as significant as the chip itself is the scale it is designed for. Alibaba says the Zhenwu V900 can scale to clusters of up to 500,000 cards — a figure that signals the company’s intent to build training infrastructure at a scale previously associated only with the largest US hyperscalers. For context, Huawei last week unveiled AI infrastructure it says can scale to as many as one million processors, underscoring how quickly the cluster-size race is escalating in China.

The new chip is scheduled for mass production and commercial release in the first quarter of 2027. Alibaba’s existing Zhenwu silicon is already deployed with more than 650 customers across automotive, finance, energy, and manufacturing — a customer base that gives the V900 a ready-made market when it ships.

A 10-trillion-parameter moonshot

On the model side, Alibaba confirmed plans to train a new AI model with between 5 trillion and 10 trillion parameters, a range that would place it among the largest dense-model training runs ever attempted. The company also disclosed that its next-generation Qwen 4 model is currently in training, with Qwen 4.5 and Qwen 5 series already on the public roadmap.

The parameter target is not just a vanity number — it aligns with Wu’s stated thesis about the trajectory of machine intelligence. At the keynote, Wu argued that machine thinking today represents less than 3 percent of humanity’s total thinking capacity, and that the growth runway is enormous. “In the future, the total amount of human thinking will continue to grow, but machine thinking will expand even faster, ultimately shouldering 99.9% of all thinking,” he said.

Alibaba is also advancing what it calls Recursive Self-Improvement — research toward AI systems that can improve themselves — as it works toward artificial superintelligence (ASI), a goal Wu first articulated publicly at last year’s Apsara conference.

20 gigawatts by 2032

The infrastructure commitment anchors the whole strategy. Wu set a target for Alibaba Cloud to surpass 20 gigawatts of global data center capacity by 2032, on the expectation of what he called “exponentially” rising demand for AI computing.

“Alibaba will join forces with all our partners to commit fully to AI infrastructure development, forging an AI cloud designed for the machine intelligence era,” Wu said. “Our target is that by 2032, the global data center capacity operated by Alibaba Cloud will surpass 20GW, fueling the industry’s exponentially rising demand for AI.”

The scale is easier to grasp by comparison. Earlier this month, Nvidia outlined plans with Australian partners to support up to 2 gigawatts of AI infrastructure in that country by 2027. In July, Meta unveiled plans for its first Canadian data center — a 1-gigawatt facility in Alberta expected to cost around $9 billion and take up to three years to build. Alibaba’s 20GW target is an order of magnitude beyond both, and reflects earlier reporting that the company was weighing a three-year capital expenditure plan of up to 480 billion yuan (roughly $69 billion) for AI data center development.

Full-stack logic: from chips to agents

Chairman Joe Tsai framed the announcements as the maturation of AI itself. AI development, he said, has entered a new phase — moving beyond breakthroughs at the technological frontier toward creating value at scale in real-world applications, with deployment accelerating in autonomous driving, embodied intelligence, and world models.

To capture that value, Alibaba is continuing to invest in what Tsai described as a full-stack AI system: chips, cloud infrastructure, AI models, model services, and agent-based applications, connected end to end. “Alibaba is firmly investing in building full-stack AI … to let AI move from a technological breakthrough to value creation,” Tsai said in his opening remarks.

The company also launched Qwen Intelligence for mobile devices — an end-to-end solution that enables phones to reason through and execute complex tasks locally, extending the Qwen family from the cloud to the edge.

Wu, for his part, reached for a historical analogy: comparing today’s AI development to the early stages of electrification, he quipped that “AI coding is simply the light bulb of the machine intelligence era” — the first killer application of a general-purpose technology still finding its footing.

What it means

Three takeaways stand out. First, the gap between Chinese and Western AI infrastructure announcements is narrowing in ambition, if not yet in proven performance: a 20GW target and 500,000-card clusters are hyperscaler-scale numbers by any definition. Second, Alibaba is uniquely positioned among Chinese firms to monetize a full stack — it owns the chip design (T-Head), the cloud (Alibaba Cloud), the models (Qwen), and the distribution (its e-commerce and device ecosystems). Third, the market read the announcement as credible: Alibaba shares jumped around 3 percent in Hong Kong trading on the news.

The open questions are execution risks the keynote did not address: whether T-Head can hit mass production on schedule in Q1 2027 given constraints on advanced chip manufacturing in China, and whether demand materializes fast enough to fill 20 gigawatts of capacity. But as a statement of intent, Apsara 2026 was unambiguous — Alibaba intends to compete at every layer of the AI stack, at global scale, on its own silicon.