云栖大会观察：阿里的未来预判与正在落地的三块拼图

The recent Apsara Conference laid out core judgments on the era of machine intelligence, while also unveiling a number of tangible AI use cases relevant to people’s daily lives, offering a clear view of where AI development actually stands today.

1. Stage assessment: AI is currently in an early stage comparable to the invention of the electric light bulb, and no mass-market, life-changing killer application has yet emerged. In the future, machines will handle 99.9% of the world’s cognitive workload, with total machine thinking volume reaching more than 1,000 times that of humans. Today that figure is less than 3%, leaving tens of thousands of times room for growth. Ultimately, humans will be freed from repetitive, labor-intensive tasks to focus on creativity and high-quality experiences.

2. Tangible, real-world progress: AI simultaneous interpretation latency has been cut to under 2.5 seconds, with the technology already integrated into AI glasses, smart earbuds, AI smartphones and other end devices. Image generation tools can compress hours of design work into seconds, and can produce complete audio tracks with dialogue and ambient sound. A next-generation video generation model with director-level narrative capabilities is slated for release in November.

3. Key future milestones: Next-generation AI chips will enter mass production in 2027; by 2032, Alibaba Cloud’s global data center footprint will exceed 20 terawatts, making access to AI as convenient as plugging into an electrical outlet.

The AI deployment updates unveiled at this year’s Apsara Conference provide clear directional references for brands to improve marketing efficiency, optimize products and identify upcoming consumer trends.

1. Proven efficiency gains for marketing: Multimodal AI has evolved from generating content to generating immersive experiences. At present, ordinary consumers can only correctly identify AI-generated content 51.2% of the time, an accuracy rate on par with a coin flip. Sixty-three percent of video marketing professionals already use AI video tools, and the AI image and video generation market is projected to reach $60.8 billion by 2030. Alibaba’s image generation model can cut hours of visual design work down to seconds; its next-generation video model, launching in November, will have director-level creative reasoning and full narrative understanding, drastically reducing the cost and turnaround time of brand visual assets and marketing videos.

2. Future consumer experience trends: A native unified omnimodal model will emerge within three years, eliminating modality-specific experience boundaries. AI capabilities will be broadly embedded into everyday devices including smartphones, glasses and earbuds, allowing brands to lay early groundwork for omnimodal interaction scenarios and build new user engagement pathways.

3. Cost trajectory: LLM inference costs have already fallen to one quarter of previous levels, and omnimodal application costs are down more than 90% from the prior generation, continuously lowering the barrier for brands to implement AI applications.

The AI technology updates announced at the Apsara Conference point to tangible opportunities for e-commerce sellers to cut costs, boost efficiency and tap new growth, while also signaling emerging competitive pressures.

1. Efficiency gains in content production: AI-generated content is now virtually indistinguishable to average consumers, and more than 60% of video marketing practitioners already use AI video tools. Alibaba’s commercially optimized image generation model compresses hours of e-commerce visual design work into seconds, while its next-generation video model launching in November features full narrative capabilities, enabling sellers to quickly produce product main images, detail pages and short marketing videos, significantly shortening new product launch cycles and cutting content costs.

2. Growth opportunities in new scenarios: AI capabilities have already been integrated into AI glasses, smart earbuds, AI smartphones and other end devices, and features such as real-time voice interaction, low-latency simultaneous interpretation and cross-app complex task processing are maturing rapidly. Sellers can adapt early to the new consumer scenarios enabled by voice interaction and multi-device reach to capture new traffic entry points.

3. Competitive implications: With LLM inference costs now at one quarter of previous levels and omnimodal application costs down more than 90%, sellers that proactively adopt AI tools will gain clear cost and efficiency advantages, while those that lag behind risk widening competitive gaps.

The AI technology and chip roadmap unveiled at the Apsara Conference provides clear guidance for manufacturing factories to identify new business opportunities, optimize R&D and design, and advance digital and intelligent transformation.

1. Efficiency improvements in R&D and design: The latest Qwen large model has demonstrated the ability to iterate continuously with zero human input. In chip design scenarios, given only a bus module specification, it can independently complete the full-chain iteration of front-end design, verification and back-end design. After running for more than 60 hours and calling EDA tools over 10,000 times, it achieved a 42% reduction in chip physical area. This capability can eventually be replicated across industrial product design and production process optimization, significantly shortening R&D cycles and reducing design costs.

2. New supply chain opportunities: Alibaba’s T-Head Zhenwu V900, a next-generation AI chip unifying training and inference, will enter mass production in the first quarter of 2027, with single-cluster scalability up to 500,000 cards. The Zhenwu chip family already serves more than 650 enterprise customers across sectors including autonomous driving, finance, manufacturing and energy, creating clear incremental demand for supporting supply chains tied to AI hardware and computing infrastructure.

3. Transformation guidance: In the future, AI computing power will be as accessible as electricity. Factories can prepare early to connect to cloud-based AI capabilities and steadily advance intelligent upgrades across production processes.

The technology trends and product rollouts announced at the Apsara Conference offer clear direction for enterprise service providers and AI service providers on technology iteration roadmaps, customer pain points and available foundational infrastructure support.

1. Defined long-term technology trends: The core driver of machine intelligence advancement is recursive self-improvement, which will eventually form a self-reinforcing flywheel of co-evolution between models and chips. A native unified omnimodal model will emerge within three years, breaking experience boundaries across modalities. Through architectural innovation, LLMs have already cut training costs by nearly 90%, reduced inference costs to one quarter of prior levels, and lowered omnimodal application costs by more than 90%, enabling service providers to build cost-effective industry solutions on a cheaper computing base.

2. Mature, reusable technology foundation: The Qwen model family has open-sourced more than 460 large models, with over 3 billion downloads and more than 300,000 derivative models, making it the most popular open-source LLM family on global open-source platforms. T-Head’s next-generation AI chip delivers three times the performance of its predecessor, enabling thousands of cards to operate in coordination as seamlessly as a single super chip. Alibaba Cloud’s plan to build a global data center network exceeding 20 terawatts by 2032 will provide service providers with stable model and computing power support.

3. Clear customer demand scenarios: E-commerce, creative design, smart hardware and other sectors all have well-defined AI adoption needs, including second-level visual asset delivery, low-latency voice interaction and cross-app task processing, all of which represent high-priority service entry points.

The Apsara Conference outlined Alibaba’s core approach to building infrastructure for the machine intelligence era, providing a reference for all types of platforms on AI-era construction, operational logic and ecosystem strategy.

1. Core platform capability roadmap: Platforms in the AI era need to build an intelligence supply system analogous to the power grid, allowing onboard merchants, developers and users to access AI capabilities as conveniently as plugging into an electrical outlet. Alibaba’s implementation path centers on an integrated foundation of models, chips and cloud: on the model side, it continues to iterate the Qwen family and make it broadly available via open source; on the chip side, it has launched a unified training-inference chip supporting 500,000-card clusters; on the cloud side, it continues to expand global data center capacity, with a target of exceeding 20 terawatts of total scale by 2032.

2. Core levers for ecosystem operation: Platforms should align with the trend of multimodal technology evolving from content generation to experience generation, integrating full-scenario AI capabilities across image, video, audio and real-time interaction to help merchants on the platform cut content production costs and shorten development cycles. Examples include enabling merchants to generate e-commerce visual assets in seconds and providing developers with open-source model interfaces to lower development barriers.

3. Risks and opportunities: With human accuracy in identifying AI-generated content at just 51.2%, platforms need to establish governance rules for AI content early. They should also build out interaction pathways across AI glasses, earbuds, smartphones and other end devices in advance to cover new user reach scenarios.

Alibaba’s overall AI layout unveiled at the 2026 Apsara Conference offers a clear view of the latest development stage, technology iteration path and emerging business models of the machine intelligence industry, carrying significant value for industry research.

1. New assessment of industry development stage: Machine intelligence is currently in an early stage analogous to the completion of Edison’s first power station, and no representative mass-market killer application that reshapes daily life has yet emerged. In the future, machines will handle 99.9% of the world’s cognitive workload, and intelligence will become a universally scalable commodity. At present, total machine cognitive output accounts for less than 3% of human thinking volume, leaving tens of thousands of times of growth potential ahead.

2. New technology iteration path: The core logic of AI evolution is recursive self-improvement coupled with a chip-model co-evolution flywheel. For example, Qwen 3.8-Max is capable of continuous iteration for over a month with zero human intervention, and can independently complete chip adaptation and full-process chip design optimization, forming a positive loop in which stronger models improve chip design efficiency, and better chip performance in turn powers stronger models. Through architectural innovation, LLM training costs can be cut by nearly 90%, providing the cost foundation for intelligence to become a standardized utility.

3. New exploration of business models: The future computing power supply system will operate in a power-grid-like model, where “chips produce tokens and cloud networks distribute tokens.” Multimodal technology will evolve from generating content to generating experiences, ultimately delivering a unified omnimodal interaction experience, with AI capabilities embedded across all categories of consumer end devices.