Alibaba's 10-Trillion-Parameter Gamble: The AI Endgame Is Here

At the 2026 Yunqi Conference in Hangzhou, Alibaba didn't just show up—it dropped a bombshell. CEO Wu Yongming took the stage to announce the company's most audacious plan yet: training a new generation of AI models with 5 to 10 trillion parameters. That's up to four times the size of today's largest models, including Alibaba's own Qwen 3.8 Max, which already packs 2.4 trillion parameters. If this sounds like a leap into the unknown, that's because it is.

The Model That Teaches Itself

What's really intriguing isn't just the size—it's how these models will learn. Wu revealed progress on recursive self-improvement (RSI). In plain English, the model can spot its own weaknesses, design experiments, generate data, and optimize itself in a loop. Think of it as a student who not only studies but also writes their own curriculum. Alibaba believes this self-evolution capability is the bedrock for artificial superintelligence (ASI). So, it's not just about building a bigger brain; it's about building one that can upgrade itself.

Chips, Clusters, and a Whole Lot of Computing Power

You can't train a 10-trillion-parameter beast on just any hardware. Enter the Zhenwu V900, Alibaba's new AI chip. Wu called it the most powerful AI chip in China, boasting three times the computing power of its predecessor. A single cluster built around these chips can coordinate up to 500,000 accelerator cards—enough to train and run the most cutting-edge models. Mass production is slated for early 2027. For context, the existing M890 super node already handles inference for models over 2 trillion parameters, a feat Wu says very few companies worldwide can match.

The infrastructure plan is equally staggering. Alibaba aims to expand its global data center capacity to over 20 gigawatts by 2032. That's a lot of power—and a lot of cooling. But here's the catch: the global supply chain is strained. Equipment, power, and infrastructure can't keep up, limiting how fast Alibaba can build. Wu admits demand still far outpaces supply, and the company is racing to deploy AI super nodes starting this quarter.

The Industrial Revolution, but for Machines

Wu painted a bigger picture. He compared this moment to the Industrial Revolution, calling it the dawn of the era of machine intelligence. Today, machine thinking is less than 3% of all human thinking. In the future, he predicts it could balloon to 1,000 times the total human thinking. The products that define this era? They probably don't exist yet. Wu likened today's popular AI coding apps to the early light bulbs of the electrical age—just the beginning.

As model capabilities grow, AI will take over complex decision-making and long-term tasks in more fields. Meanwhile, with U.S. export restrictions tightening, Chinese tech firms are racing to build an independent AI supply chain. Alibaba's simultaneous push on super-large models, self-developed chips, and data centers is a critical step toward a complete AI ecosystem. The global competition over super-models, computing infrastructure, and next-gen AI chips just entered a new stage.

Key Points

Alibaba plans 5-10 trillion parameter models, up to four times larger than Qwen 3.8 Max.
Recursive self-improvement (RSI) will let models design experiments and optimize themselves.
Zhenwu V900 chip triples computing power; mass production expected Q1 2027.
Data center capacity to reach 20 GW by 2032, but supply chain constraints slow expansion.
Machine thinking could grow 1,000x human thinking, marking the era of machine intelligence.
U.S. export restrictions push China to accelerate its independent AI supply chain.

Alibaba's 10-Trillion-Parameter Gamble: The AI Endgame Is Here