FB Pixel no scriptAlibaba Cloud lays out vision for artificial superintelligence at Apsara Conference 2025
MENU
KrASIA
News

Alibaba Cloud lays out vision for artificial superintelligence at Apsara Conference 2025

Written by 36Kr English Published on   6 mins read

Share
The company says it wants to become “the Android of the LLM era.”

Even as drizzle lingered over Hangzhou, the energy around artificial intelligence inside Yunqi Town was undiminished when this year’s Apsara Conference opened on September 24. Eddie Wu, CEO of Alibaba Group and chairman and CEO of Alibaba Cloud Intelligence Group, delivered the keynote speech, outlining his vision for artificial superintelligence (ASI).

The 2024 edition had marked Wu’s first appearance since taking over Alibaba Cloud more than a year earlier. At the time, he argued that the greatest potential of generative AI was not in building one or two new superapps for smartphones, but in “taking over the digital world and transforming the physical one.”

Then, his words were aspirational. This year, they have evolved into a concrete roadmap backed by aggressive action.

A barrage of releases

This year, Alibaba Cloud unveiled a slate of new releases, led by its flagship model Qwen3-Max, the most advanced in the Qwen family. The company claims its performance surpasses GPT-5 and Claude Opus 4, placing it among the global top three on LMArena. Other launches include Fun, a new family of speech models.

What stood out most, however, were Wu’s two bold assertions.

First, he argued that large models are the next operating system, absorbing traditional software and enabling anyone to create applications through natural language. In the future, he suggested, nearly all digital interactions may come through model-driven agents rather than conventional software.

This conviction is why, in recent years, Alibaba Cloud has been rebuilding its stack—across compute infrastructure, middleware, and cloud services—to align with the changes large models bring to the technology base.

Second, Wu proposed that a “super AI cloud” is the next computer, and Alibaba Cloud’s goal is to build it. Drawing parallels to earlier eras of computing, he explained that in the AI era, natural language is the new programming language, agents are the new software, context is the new memory, and large models form the operating system-like layer that orchestrates interactions among users, software, and compute resources.

Back in February, Alibaba announced a RMB 380 billion (USD 53.2 billion) three-year investment plan in AI infrastructure. Wu added that by 2032, compared with 2022, Alibaba Cloud’s global data center energy consumption will grow tenfold to meet the arrival of the ASI era.

He also laid out Alibaba Cloud’s development strategy: not artificial general intelligence (AGI), which has dominated industry debate, but rather ASI.

Wu broke down the path to ASI into three stages:

  • First, emergent intelligence, when AI develops reasoning by learning from humanity’s collective knowledge.
  • Second, autonomous action, when AI acquires tool use and programming abilities to assist humans, and that’s roughly where the industry stands today.
  • Finally, self-iteration, when AI connects to raw data from the physical world, learns autonomously, and ultimately surpasses humans.

The global large model race in 2025 remains uncertain. After releasing GPT-5, OpenAI faced criticism for lackluster results and stalled innovation. Meanwhile, Meta and OpenAI have doubled down on investment, unwilling to miss out on this technological shift.

Alibaba Cloud, however, has signaled that it is not just investing, but doing so aggressively. The market reacted accordingly. Alibaba’s Hong Kong-listed shares jumped more than 9% intraday following the announcement, reaching their highest level since October 2021.

Demonstrating scale and intent

Before the conference, Lin Junyang, head of the Qwen team, teased on X that Alibaba will release at least six new models, and none of them are “small.”

When the announcements came, the number was indeed higher. Zhou Jingren, CTO of Alibaba Cloud, moved through slides at breakneck speed during his presentation, but still ran over time.

In total, Alibaba Cloud introduced seven major releases, each delivering significant advances in size or performance:

  • Qwen3-Max, its flagship model, is trained on 36 trillion tokens with more than a trillion parameters. It delivers significantly stronger coding and agent-tooling capabilities. In blind tests such as LMArena, preview versions ranked third globally on the Chatbot Arena leaderboard.
  • Qwen3-Next introduces a new architecture with 80 billion total parameters but only three billion active at a time. This design cuts training costs by more than 90% compared with dense models of similar performance.
  • Qwen3-VL, a vision model that can interpret charts and images, and even convert design mockups into frontend code.
  • Qwen3-Coder, an upgraded programming model that offers faster generation, higher-quality output, and improved safety.
  • Qwen3-Omni, a multimodal model able to process and integrate text, audio, and video.
  • Wan2.5-Preview, a vision foundation model for text-to-video, image-to-video, and editing.
  • Fun, a family of speech models that provides preset voices for applications ranging from e-commerce to audiobooks and children’s entertainment.
Photo and header photo source: Alibaba Group.

The launch came amid a surge in open-source competition, triggered by DeepSeek, which reignited global debates about open versus closed models. Unlike last year’s closed approach, 2025 has seen nearly all model makers expand their open-source commitments. Among China’s tech giants, Alibaba has been the most aggressive.

Its early start in open-sourcing and ecosystem building is now paying off. Alongside DeepSeek, Qwen is one of the few Chinese models with global recognition. To date, Alibaba has open-sourced more than 300 models across formats, with more than 600 million downloads and 170,000 derivative models worldwide.

Alibaba also introduced a new agent development framework, ModelStudio-ADK, which lets agents plan and invoke models autonomously, though at higher compute costs. Daily model calls on its Model Studio platform have reportedly increased 15-fold in the past year.

This open-source push has already translated into revenue. Alibaba Cloud’s quarterly revenue rose 26% year-on-year, with AI-related income growing at triple-digit rates for eight consecutive quarters. According to Omdia, China’s AI cloud market reached RMB 22.3 billion (USD 3.1 billion) in the first half of 2025, with Alibaba Cloud holding 35.8%, more than the next three players combined.

“The Android of the LLM era”

In 2024, OpenAI’s video generator Sora and the lukewarm reception to GPT-5 briefly dampened industry optimism. That lull did not last. Just days before Apsara, Nvidia announced a USD 100 billion investment into OpenAI. Wu predicted that global AI investment will exceed USD 4 trillion over the next five years.

In a media interview after the conference, CTO Zhou noted that while companies agree broadly on technical direction, execution can differ greatly. “Model competition is now about systems versus systems,” he said. “Innovation isn’t about holding back for a big reveal. Models, infrastructure, and cloud are all interdependent.”

The “systems” in question point to strategic choices. After DeepSeek shifted the debate, major players have increased spending on compute power, cloud services, and open-source ecosystems.

Tencent has focused first on applying AI to its existing businesses before expanding outward. ByteDance, meanwhile, has followed an iOS-style path, with tight integration from model to application, though it prefers to perfect closed versions before selectively open-sourcing.

Alibaba Cloud, after Wu’s arrival in 2023, took a different turn. It shed low-margin projects and doubled down on AI, investing in startups, scaling in-house model development, pushing open source, and rebuilding infrastructure. Its approach resembles Google’s full-stack strategy, spanning compute, cloud, and models.

Alibaba’s ASI roadmap isn’t entirely novel, echoing Google DeepMind’s framework, which outlines six levels of AGI, ending with a stage “superhuman.”

Wu’s push underscores the inseparability of AI and cloud solutions. As he put it: “Tokens are the electricity of the AI world.”

The AI era is still in its early stages. Model usage accounts for only a small share of enterprise cloud consumption, but the shift is accelerating. Xu Dong, general manager of Qwen, noted that while most requests a year ago involved offline data labeling, online tasks have now surged by several orders of magnitude as companies embed models into their workflows.

For 16 years, Alibaba Cloud has described itself as the “water and electricity” of the digital world. Its new tagline, “the Android of the LLM era,” extends that metaphor into a platform strategy for the age of AI.

KrASIA Connection features translated and adapted content that was originally published by 36Kr. This article was written by Deng Yongyi for 36Kr.

Share

Loading...

Loading...