Alibaba Plans 5- to 10-Trillion-Parameter Model in Full-Stack AI Push

Alibaba Group CEO Eddie Wu said the company plans to train a new artificial intelligence model at the scale of 5 to 10 trillion parameters, presenting a full-stack AI roadmap spanning models, chips, and cloud infrastructure in a keynote at the 2026 Apsara Conference in Hangzhou on September 22, 2026.

Apsara Conference is Alibaba Cloud’s annual flagship event. The official conference agenda placed Wu’s address in the Opening Ceremony and Apsara Keynote Session, which also featured sessions dedicated to Qwen and to T-Head’s silicon work. The organizer describes the 2026 edition as spanning three premier keynote forums, more than 120 parallel forums, and a 50,000-square-meter technology exhibition bringing together over 1,000 enterprises.

Plans for a 5 to 10 Trillion-Parameter Model

Wu said the technical pathway toward artificial superintelligence, or ASI, has become increasingly clear, pointing to Recursive Self-Improvement, or RSI, a process in which models engage with real-world tasks and feedback, identify their own limitations, and autonomously design experiments, synthesize data, and evaluate outcomes in a continuous cycle of self-evolution. He said Alibaba’s Qwen team is exploring RSI and has made meaningful progress, while continuing research on model architecture and data optimization.

The planned model, at 5 to 10 trillion parameters, is aimed at completing more complex, longer-horizon tasks and advancing toward ASI, Wu said. He said multimodal models that unify understanding and generation are another core research direction, arguing that models need profound multimodal abilities to comprehend and communicate human signals such as voice, visuals, facial expressions, and gestures and to align with human culture, aesthetics, and values.

Wu framed the past year as one in which AI moved from artificial general intelligence as a starting point toward self-iterating ASI, with the “vibe coding” paradigm maturing into full-spectrum “vibe working” and what he described as decisive breakthroughs in long-horizon task execution. He set out two views of the era: that machines will produce more than 1,000 times the thinking of all humanity combined, and that the products that will truly define the era have not yet arrived. Machine thinking currently amounts to less than 3 percent of all human thinking, he said, while machine power already drives 99.9 percent of the world’s physical work, and he predicted machine thinking will ultimately shoulder 99.9 percent of all thinking.

Wu called AI coding “simply the light bulb of the Machine Intelligence era,” arguing that merely automating existing human work will never define a new era. Drawing an analogy to electrification, he noted that Thomas Edison’s Pearl Street Station in 1882 powered barely 400 lightbulbs, while air conditioning arrived in 1902 and the first digital computer in 1946.

Wu said the Machine Intelligence era rests on three infrastructure cornerstones: AI models, AI chips, and the AI cloud. He said Alibaba remains committed to building that core infrastructure across all three as a long-term strategic priority.

Zhenwu V900 Chip and M890 AI Supernodes

Wu said Alibaba’s T-Head semiconductor unit is building a data center chip portfolio covering the Zhenwu series of GPU chips, the Yitian series of CPU chips, the Panmai series of smart network interface cards, and ICN interconnect chips, together spanning the core silicon required to build ultra-large-scale AI clusters.

At the conference, Alibaba introduced the next-generation Zhenwu V900, which Wu called “the most powerful AI chip in China today.” He said the V900 delivers three times the performance of its predecessor, the Zhenwu M890, and that a single cluster built on it can support up to 500,000 cards for frontier model training and inference. Citing the maturity of T-Head’s product lines and adoption across clients, he said Alibaba anticipates significant growth in annual AI chip shipment volumes.

Wu said Alibaba’s proprietary M890 AI Supernode already delivers high-efficiency inference for foundation models above 2 trillion parameters, a capability tier he said only a handful of companies globally hold, and that Alibaba Cloud is bringing AI Supernodes online at commercial scale. He said the company is pursuing end-to-end co-optimization across chips, servers, supernodes, networks, models, and inference engines to maximize token throughput and cost efficiency across its AI cluster infrastructure.

A 20-Gigawatt Cloud Target and a Qwen Device Push

Wu described the AI cloud as the grid that delivers tokens wherever they are needed, saying it demands hyper-scale, globally distributed infrastructure that co-optimizes GPUs, CPUs, networking, storage, and databases. He said customer demand for AI remains exceptionally robust, with the industry’s mid-to-long-term demand far outpacing Alibaba’s supply capabilities and global shortages across the AI data center supply chain currently limiting the speed at which the company can scale compute infrastructure. He said provisioning AI compute is driving accelerating revenue growth at Alibaba Cloud.

Wu set a target for Alibaba Cloud to operate more than 20 gigawatts of global data center capacity by 2032, describing the build-out as a response to exponentially rising industry demand for AI.

Wu also said Alibaba’s open-source Qwen-27B has emerged as the most popular model among developers worldwide on desktop, and that the company will continue refining such models so developers and enterprises can deploy intelligence locally. At the conference, Alibaba launched Qwen Intelligence, an end-to-end solution built for partners that enables mobile devices to reason through and execute complex tasks.

Wu closed the address by restating Alibaba’s conviction that the more capable AI becomes, the more powerful humanity will be, describing that conviction as the ultimate purpose of the company’s commitment to the Machine Intelligence era.