Alibaba Cloud sets out vision for synthetic superintelligence at Apsara Conference 2025

Even as drizzle remained over Hangzhou, the energy around expert system 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, provided the keynote speech, describing his vision for synthetic superintelligence (ASI).

The 2024 edition had actually marked Wu’s very first look because taking control of Alibaba Cloud more than a year previously. At the time, he argued that the best capacity of generative AI was not in constructing a couple of brand-new superapps for mobile phones, however in “taking control of the digital world and changing the physical one.”

His words were aspirational. This year, they have actually progressed into a concrete roadmap backed by aggressive action.

A barrage of releases

This year, Alibaba Cloud revealed a slate of brand-new releases, led by its flagship design Qwen3-Max, the most advanced in the Qwen household. The business declares its efficiency goes beyond GPT-5 and Claude Opus 4, positioning it amongst the international leading 3 on LMArena. Other launches consist of Fun, a brand-new household of speech designs.

What stuck out most, nevertheless, were Wu’s 2 strong assertions.

He argued that big designs are the next operating system, soaking up standard software application and allowing anybody to develop applications through natural language. In the future, he recommended, almost all digital interactions might come through model-driven representatives instead of traditional software application.

This conviction is why, over the last few years, Alibaba Cloud has actually been reconstructing its stack– throughout calculate facilities, middleware, and cloud services– to line up with the modifications big designs give the innovation base.

Second, Wu proposed that a “incredibly AI cloud” is the next computer system, and Alibaba Cloud’s objective is to develop it. Drawing parallels to earlier ages of computing, he described that in the AI period, natural language is the brand-new programs language, representatives are the brand-new software application, context is the brand-new memory, and big designs form the operating system-like layer that manages interactions amongst users, software application, and calculate resources.

Back in February, Alibaba revealed a RMB 380 billion (USD 53.2 billion) three-year financial investment strategy in AI facilities. Wu included that by 2032, compared to 2022, Alibaba Cloud’s international information center energy intake will grow significantly to fulfill the arrival of the ASI period.

He likewise set out Alibaba Cloud’s advancement technique: not synthetic basic intelligence (AGI), which has actually controlled market dispute, however rather ASI.

Wu broke down the course to ASI into 3 phases:

Emerging intelligence, when AI establishes thinking by discovering from mankind’s cumulative understanding.
Second, self-governing action, when AI gets tool usage and shows capabilities to help people, which’s approximately where the market stands today.
Self-iteration, when AI links to raw information from the physical world, finds out autonomously, and eventually exceeds human beings.

The international big design race in 2025 stays unsure. After launching GPT-5, OpenAI dealt with criticism for uninspired outcomes and stalled development. Meta and OpenAI have actually doubled down on financial investment, reluctant to miss out on out on this technological shift.

Alibaba Cloud, nevertheless, has actually indicated that it is not simply investing, however doing so strongly. The marketplace responded appropriately. Alibaba’s Hong Kong-listed shares leapt more than 9% intraday following the statement, reaching their greatest level because October 2021.

Showing scale and intent

Before the conference, Lin Junyang, head of the Qwen group, teased on X that Alibaba will launch a minimum of 6 brand-new designs, and none are “little.”

When the statements came, the number was certainly greater. Zhou Jingren, CTO of Alibaba Cloud, moved through slides at breakneck speed throughout his discussion, however still ran over time.

In overall, Alibaba Cloud presented 7 significant releases, each providing considerable advances in size or efficiency:

Qwen3-Maxits flagship design, is trained on 36 trillion tokens with more than a trillion criteria. It provides considerably more powerful coding and agent-tooling abilities. In blind tests such as LMArena, sneak peek variations ranked 3rd worldwide on the Chatbot Arena leaderboard.
Qwen3-Next presents a brand-new architecture with 80 billion overall criteria however just 3 billion active at a time. This style cuts training expenses by more than 90% compared to thick designs of comparable efficiency.
Qwen3-VLa vision design that can translate charts and images, and even transform style mockups into frontend code.
Qwen3-Coderan updated shows design that uses much faster generation, higher-quality output, and enhanced security.
Qwen3-Omnia multimodal design able to procedure and incorporate text, audio, and video.
Wan2.5-Previewa vision structure design for text-to-video, image-to-video, and modifying.
Enjoyablea household of speech designs that offers predetermined voices for applications varying from e-commerce to audiobooks and kids’s home entertainment.

The launch came in the middle of a rise in open-source competitors, set off by DeepSeek, which reignited worldwide arguments about open versus closed designs. Unlike in 2015’s closed technique, 2025 has actually seen almost all design makers broaden their open-source dedications. Amongst China’s tech giants, Alibaba has actually been the most aggressive.

Its early start in open-sourcing and environment structure is now settling. Together With DeepSeek, Qwen is among the couple of Chinese designs with international acknowledgment. To date, Alibaba has actually open-sourced more than 300 designs throughout formats, with more than 600 million downloads and 170,000 acquired designs worldwide.

Alibaba likewise presented a brand-new representative advancement structure, ModelStudio-ADK, which lets representatives strategy and conjure up designs autonomously, however at greater calculate expenses. Daily design contacts its Model Studio platform have actually apparently increased 15-fold in the previous year.

This open-source push has actually currently equated into income. Alibaba Cloud’s quarterly profits increased 26% year-on-year, with AI-related earnings growing at triple-digit rates for 8 successive quarters. According to Omdia, China’s AI cloud market reached RMB 22.3 billion (USD 3.1 billion) in the very first half of 2025, with Alibaba Cloud holding 35.8%, more than the next 3 gamers integrated.

“The Android of the LLM period”

In 2024, OpenAI’s video generator Sora and the lukewarm reception to GPT-5 briefly moistened market optimism. That lull did not last. Simply days before Apsara, Nvidia revealed a USD 100 billion financial investment into OpenAI. Wu anticipated that international AI financial investment will go beyond USD 4 trillion over the next 5 years.

In a media interview after the conference, CTO Zhou kept in mind that while business concur broadly on technical instructions, execution can vary significantly. “Model competitors is now about systems versus systems,” he stated. “Innovation isn’t about keeping back for a huge expose. Designs, facilities, and cloud are all synergistic.”

The “systems” in concern indicate tactical options. After DeepSeek moved the dispute, significant gamers have actually increased costs on calculate power, cloud services, and open-source communities.

Tencent has actually focused initially on using AI to its existing services before broadening external. ByteDance, on the other hand, has actually followed an iOS-style course, with tight combination from design to application, though it chooses to ideal closed variations before selectively open-sourcing.

Alibaba Cloud, after Wu’s arrival in 2023, took a various turn. It shed low-margin tasks and doubled down on AI, buying start-ups, scaling internal design advancement, pressing open source, and restoring facilities. Its method looks like Google’s full-stack technique, covering calculate, cloud, and designs.

Alibaba’s ASI roadmap isn’t completely unique, echoing Google DeepMind’s structure, which lays out 6 levels of AGI, ending with a phase “superhuman.”

Wu’s push highlights the inseparability of AI and cloud options. As he put it: “Tokens are the electrical energy of the AI world.”

The AI period is still in its early phases. Design use represent just a little share of business cloud intake, however the shift is speeding up. Xu Dong, basic supervisor of Qwen, kept in mind that while many demands a year back included offline information labeling, online jobs have actually now risen by a number of orders of magnitude as business embed designs into their workflows.

For 16 years, Alibaba Cloud has actually explained itself as the “water and electrical power” of the digital world. Its brand-new tagline, “the Android of the LLM period,” extends that metaphor into a platform technique for the age of AI.

KrASIA Connection includes equated and adjusted material that was initially released by 36Kr. This short article was composed by Deng Yongyi for 36Kr.