Apsara Conference 2026: When AI Begins to “Work in Teams”

From chips and models to cloud, from agents to super-individuals, an industry-wide mobilization to put intelligence to use

Hangzhou in September was once again the focus of the global technology community. The 2026 Apsara Conference was held from September 22 to 24 at the Hangzhou International Expo Center, with a second venue in Yunqi Town. Unlike previous years, many people walking into the venue were not only there to listen to speeches and visit exhibition booths—they also brought their own agents with them. This small change reflects a deeper shift taking place at the Apsara Conference and across the entire AI industry: AI is no longer merely an object to be demonstrated and discussed. It is becoming a “digital colleague” beside attendees, developers, and enterprises.

This year’s Apsara Conference took “Putting Intelligence to Use” as its theme. The theme is not abstract. It ran through the main forum, exhibition areas, breakout sessions, and industry matchmaking almost everywhere. From computing infrastructure to model capabilities, from agent platforms to embodied AI robots, from e-commerce workbenches to automotive cockpits, the conference tried to answer an increasingly urgent question: after large models have moved beyond the “can chat” stage, how can they truly enter industries, work for people, collaborate with people, and create measurable value?

A Conference of “Historic Scale”

In terms of scale, the 2026 Apsara Conference set multiple records. The number of exhibition halls increased from three to four, breakout sessions expanded to 140, participating companies jumped from more than 500 to more than 1,000, total exhibition area exceeded 50,000 square meters, and more than 2,000 speakers took the stage. Whether measured by the number of exhibitors, exhibition area, or the breadth of topics covered, this was the largest Apsara Conference in history.

Behind this growth is a shift in the center of gravity of the AI industry. In the past, the Apsara Conference narrative revolved more around cloud computing, big data, and the Internet of Things. Later, large models became the absolute protagonist. This year, the keywords became “implementation,” “agents,” and “industry co-creation.” The newly added Industry Co-creation Pavilion focused on real-world industrial applications and became the core driver of growth in both exhibitor numbers and exhibition area. What was displayed here was no longer just laboratory model parameters, but real-scenario solutions for manufacturing, retail, logistics, finance, healthcare, education, and other industries.

This year’s conference also pioneered a native-agent way of attending. Each attendee could bring their own agent into the venue, obtain a dedicated digital identity, read the agenda in real time, and receive intelligent recommendations based on preferences. The Apsara Conference thus became the world’s first Agent Native technology event. Attendees no longer needed to manually filter through a massive agenda. Agents could automatically plan routes based on interests, profession, and historical behavior, and even book exhibition experiences in advance, organize meeting notes, and generate matchmaking lists. This experience itself was a live demonstration of “putting intelligence to use.”

Three Cornerstones: Chips, Models, and Cloud

In his keynote at the main forum, Alibaba Group CEO Eddie Wu made a judgment: machines are becoming the main force of thinking, and intelligence is becoming a scalable commodity. He noted that in the future, total machine thinking will reach more than 1,000 times that of humans, while today that ratio is less than 3 percent. Behind this number is an entire industrial layout spanning chips, models, and cloud.

Eddie Wu made it clear that AI models, AI chips, and AI cloud are the three cornerstones of the machine intelligence era, and Alibaba will firmly invest in these three infrastructure areas. This year’s conference showcased the latest achievements around these three cornerstones and sent a clear signal: AI competition is no longer competition among individual models, but competition among full-stack capabilities.

On the chip side, the computing foundation continues to thicken. Alibaba’s T-Head released the domestic AI chip Zhenwu V900. Its computing power is three times that of the previous generation M890, it is equipped with 216GB of memory, and its inter-chip interconnect bandwidth reaches 1,200GB/s. A single cluster can scale to 500,000 cards, with mass production planned for the first quarter of 2027. For large-model training and inference, memory capacity, interconnect bandwidth, and cluster scale determine whether it can support larger parameters, longer contexts, and higher-concurrency tasks. The release of Zhenwu V900 means domestic AI chips are moving from “usable” to “easy to use,” especially in ultra-large-scale cluster scenarios, where interconnect capability is often more critical than peak single-card computing power.

T-Head also announced for the first time its roadmap for Yitian server CPUs. In 2027, it will launch two generations, Yitian 720 and 730. In the future, Yitian 750 will support direct interconnection with Zhenwu AI chips through a self-developed ICN bus, further improving CPU-AI chip collaboration efficiency. The significance of this roadmap is that cloud vendors are no longer merely purchasing general-purpose chips to build computing power. Instead, through self-developed CPUs, AI chips, and interconnect buses, they are building a tighter software-hardware collaboration system. CPUs handle general scheduling, data preprocessing, and some inference tasks, while AI chips handle large-scale matrix computation. The efficiency of data movement between the two directly affects the utilization and cost of the entire cluster.

On the model side, the Qwen team made substantive progress in recursive self-improvement. Recursive self-improvement means a model can generate training data, evaluate results, identify problems, and continue iterating on that basis, forming a “self-evolution” loop. Qwen3.8-Max continued iterating for more than one month with zero human involvement, completing 33 effective iterations. Its score on Artificial Analysis rose from 40 to 45, placing it in the same top tier as the strongest Claude and GPT models. The reason this progress attracted attention is not only the score improvement, but also the extremely low level of human participation during iteration, indicating that models are gaining a certain degree of autonomous optimization capability.

Qwen3.8-Flash open-sourced the next-generation Qwen large-model architecture in advance, cutting training costs by nearly 90 percent. Qwen4, based on the new architecture, is already in training, while Qwen4.5 and Qwen5 will expand to 5 to 10T parameter scale. At present, Alibaba has open-sourced more than 460 Qwen large models, with downloads exceeding 3 billion and more than 300,000 derivative models, firmly ranking as the world’s No. 1 open-source model family. The ecosystem flywheel effect brought by the open-source strategy is becoming apparent: the more developers there are, the richer the application scenarios, the more feedback data, the faster model iteration, and the more developers are attracted in turn. For a large number of small and medium-sized enterprises and entrepreneurs, open-source models have greatly lowered the threshold for AI applications and made the trend of “models as platforms” more obvious.

On the cloud side, Alibaba Cloud launched a series of new products including AgentCore, Agent Sandbox, and the new-generation high-performance storage CPFS. CPFS achieved 100TB/s-level throughput, increased a single file system to 100PiB, and reduced AI storage costs by up to 69 percent. Elastic model startup was compressed from 1,200 seconds to 70 seconds, first-token latency was reduced by 38 percent, and Prompt Caching costs were reduced by up to 95 percent. These indicators may seem like technical details, but they directly determine the user experience and commercial viability of AI applications. Slow model startup, high first-token latency, and expensive caching all make agents appear sluggish and costly in actual business. Through optimizations in storage, scheduling, caching, and sandboxes, Alibaba Cloud is trying to turn agents from “demos” into “production tools.”

Eddie Wu announced that Alibaba will join partners in fully investing in AI infrastructure, with the goal that by 2032, the global data center scale operated by Alibaba Cloud will exceed 20GW. What does 20GW mean? It means huge investments in electricity, land, servers, and networks, and it also means AI cloud will become a society-level infrastructure similar to power grids and transportation networks. In the future, competition among cloud vendors will not only be competition in model capabilities, but also competition in energy, supply chains, data center layout, and global service capabilities.

AI That “Can Work” Is Replacing AI That “Can Chat”

If the main forum on the first day depicted a vision of the machine intelligence era, the scenes in the exhibition halls brought the topic back to the present. This year’s Creation Pavilion had 11 themed zones, covering 17 professional roles and bringing together more than 200 directly reusable Skills, kits, and templates. A Skill can be understood as a skill module that an agent can call. Kits and templates allow developers to avoid building from scratch and instead assemble business assistants like building blocks. This idea of “skillization” and “modularization” is moving AI applications from custom projects toward scalable replication.

Qwen Office released its latest AI Agent hardware product on site. Users can assign tasks by voice, or turn on device listening and transcription, allowing the agent to further understand, process, and generate content. It is no longer just a voice assistant, but more like a portable workbench: automatically recording during meetings, organizing minutes afterward, extracting to-dos, generating emails, and even calling other tools based on context. The emergence of hardware forms means agents are looking for more natural entry points. Phones, earphones, glasses, recorders, and desktop devices may all become carriers for agents.

Alibaba’s global e-commerce AI workbench Accio brought real e-commerce tasks to the site. From product selection and market research to product detail page design and ad-spend analysis, one workbench can complete the entire chain of operations. For small and medium-sized merchants, work that previously required collaboration among operations staff, designers, ad buyers, and data analysts can now be completed by one or two people, or even one person, with AI assistance. This is not simply an efficiency improvement; it is a change in organizational form.

A noteworthy signal is the rise of the “super individual.” The conference specially set up a Super Individual Operations Room. Relying on tools such as large models, agents, and cloud services, a single person can complete a full business loop covering product development, design, marketing, and customer service, achieving “one person as an army.” In the past, starting a business required a team, capital, and a relatively long startup period. Now, a person with industry knowledge and good use of AI tools can quickly validate ideas, build products, and reach customers. AI has not eliminated jobs, but it is redefining job boundaries and forcing organizations to rethink collaboration.

In the Industry Pavilion, embodied AI robots demonstrated assembly production in a simulated workshop environment. Unitree’s humanoid robots staged a boxing “championship.” The world’s first transformable robot switched freely between bipedal and quadrupedal states. BYD displayed a vehicle equipped with the AI super agent “Didixia.” With a single wake-up, users can have the agent operate the in-car system, order takeout, and book hotels. Cars are no longer just transportation tools, but mobile intelligent spaces. Robots are no longer just performers on exhibition stands, but are beginning to enter factories, logistics, inspection, and home service scenarios.

These demonstrations point to a common trend: AI is moving from “can generate” to “can deliver.” Generating a piece of text, an image, or a video is only the first step. What is truly valuable is understanding tasks, breaking down steps, calling tools, executing operations, checking results, and replanning when failures occur. This is the fundamental difference between agents and chatbots. Chatbots wait for questions; agents actively work. Chatbots give answers; agents deliver results.

Industrial Implementation: From Demonstration to Production

The bustle of the Apsara Conference must ultimately return to industrial efficiency. The reason the newly added Industry Co-creation Pavilion is important is that it pulls AI from technical narrative back to the business front line. Manufacturing companies care about yield, scheduling, energy consumption, and equipment maintenance. Retail companies care about inventory, conversion, customer complaints, and repurchase. Financial institutions care about risk control, compliance, investment research, and customer service. Healthcare institutions care about diagnostic assistance, medical record quality control, and research efficiency. Different industries have different pain points, and the ways AI is implemented are completely different.

In manufacturing scenarios, embodied AI robots and visual inspection systems are trying to replace repetitive labor and high-risk operations. Large models can read equipment manuals, maintenance records, and sensor data to help engineers quickly locate faults. Agents can coordinate multiple robots to complete material handling, assembly, and inspection. In retail scenarios, AI workbenches can automatically generate product materials, analyze competitor prices, optimize advertising strategies, and adjust budgets based on real-time data. In financial scenarios, agents can assist analysts in organizing announcements, generating summaries, and tracking public opinion, but final decisions still need human review.

The difficulty of industrial implementation has never been only about model capabilities. Whether data is available, whether processes are connected, whether permissions are clear, whether results are explainable, and whether input-output returns are reasonable are all key factors determining project success. Many breakout sessions at the Apsara Conference discussed exactly these issues: how to inject industry knowledge into models, how to build enterprise-specific knowledge bases, how to let agents securely access internal systems, and how to evaluate the ROI of AI projects. These topics lack flashy concepts, but they determine whether AI can truly enter production systems.

Security, Governance, and Open Source: Calm Thinking Behind the Bustle

Beyond the lively exhibition areas and passionate speeches, this year’s Apsara Conference also featured much calm discussion. As agents begin to have memory, tool-calling capabilities, and a certain degree of autonomy, security and governance issues become more complex. If an agent can read emails, operate databases, and call payment interfaces, then its identity authentication, permission boundaries, behavior auditing, and failure rollback become crucial. The emergence of products such as Agent Sandbox is precisely to run agents in isolated environments and prevent misoperations and malicious calls.

The route debate between open source and closed source also continues. The Qwen family insists on open source, bringing a huge developer ecosystem. But open source also means model capabilities spread faster, and safety alignment and abuse prevention become more challenging. How to strike a balance between open innovation and risk control is a problem the industry must face. At the same time, collaboration protocols among agents, tool-calling standards, and identity and permission management are becoming new infrastructure topics. In the future, whether agents from different vendors can “work in teams” with one another may be more important than how strong a single agent is.

Energy consumption is also an unavoidable topic. The 20GW data center target means enormous electricity demand. Green energy, liquid cooling technology, chip energy efficiency, and computing scheduling optimization will all affect the long-term cost of AI cloud. In the end, AI competition is not only competition in algorithms and chips, but also competition in energy and engineering capabilities.

The Market’s Response

The capital market gave positive feedback on this year’s Apsara Conference. Goldman Sachs maintained its forecast for Alibaba Cloud revenue growth above market consensus, expecting year-on-year growth of 53 percent, 55 percent, and 55 percent for the September 2026 quarter through the March 2027 quarter, respectively. UBS estimates that by 2032, Alibaba Cloud’s external cloud revenue is expected to reach about $170 billion, corresponding to a compound annual growth rate of about 40 percent. Citi’s estimate is close, believing that 20GW can correspond to about $160 billion in external cloud revenue. Morgan Stanley maintained its “Overweight” rating and $180 target price for Alibaba, listing it as a top pick, and believes that more than 20GW of global cloud capacity by 2032 can provide a path to about $240 billion in cloud revenue capacity.

Whether these forecasts are too optimistic still needs time to verify. But the logic of the capital market is not difficult to understand: if AI becomes the new productivity infrastructure, then computing power, models, and cloud services will have long-term and stable demand like water, electricity, and coal. Alibaba Cloud simultaneously possesses chips, models, a cloud platform, and rich business scenarios, and this full-stack layout gives it certain advantages in competition. Of course, risks also exist: huge capital expenditure, long return cycles, uncertainty in chip supply, possible commoditization of model capabilities, and increasingly fierce international competition. What the Apsara Conference shows is the direction; the real test comes after the conference.

Conclusion: Putting Intelligence to Use Is Only the Beginning

From engineering breakthroughs in computing infrastructure to large models moving toward autonomous evolution, and then to wearable devices and super-individual workbenches taking root in real scenarios, this year’s Apsara Conference outlined a clear industrial path. AI is moving from “can generate” to “can deliver.” It is no longer limited to competition over algorithmic parameters, but is taking root downward in industry scenarios and extending upward to full-stack collaboration at the computing foundation.

The four words “Putting Intelligence to Use” are both a response to the large-model boom of recent years and a judgment on future industrial competition. No matter how strong a model is, if it cannot reduce costs and increase efficiency for enterprises, or complete real tasks for individuals, it is just an expensive toy. No matter how smart an agent is, if it cannot safely, reliably, and scalably work in teams, it cannot become part of production systems.

Apsara Conference 2026 shows that AI has already begun to “work in teams”: chips team up with models, cloud teams up with agents, machines team up with robots, and people team up with digital colleagues. The next question is no longer “Can AI do it?” but “At what cost, with what reliability, and at what speed can AI do it?” This industry-wide mobilization starting from Hangzhou has only just begun.