AI Industry Ecosystem Conference in Beijing Urges Shift From Generation to Delivery

Meng Haowei, co-founder of AI Large Model Works, said in his opening speech that the AI industry's scarce resource is not information but the ability to turn information into judgment, judgment into cooperation, and cooperation into deliverable results. He said AI is no longer an isolated software industry: from an AI call in Silicon Valley to power, data centers, copper mines, railways and ports in Africa, models, computing power, energy, resources and infrastructure are nested and linked. On the application side, AI is moving from Copilot to 'digital employees' — humans set goals, AI executes continuously, and humans accept results and take responsibility. He left the industry three questions: what irreplaceable capability a company has, where the hardest problem in its industry is, and whether it can turn that into a result customers will pay for.

Jia Anya, vice president of large model products at SenseTime, said AI solves 'whether it can be done' while people decide 'whether it is worth doing.' She described three shifts over two years: from 'being able to talk' to 'being able to do,' from a single model to a technology stack, and from capability-driven to value-driven. As models, products and open-source code move toward 'equal rights,' companies should first find high-value problems and then consider what measurable incremental value AI can create. Personal efficiency does not automatically equal enterprise efficiency; enterprises care about customer acquisition, growth, operational optimization and ROI. AI adoption cannot stop at providing a tool set. Enterprise data, processes and business needs are highly personalized, and the challenge is converting general AI capabilities into a standardized delivery system that produces personalized results. SenseTime's 'Xiaohuanxiong' has covered more than 20 million individual users and served more than 8,000 enterprises, moving from data analysis to complex task planning, tool invocation and enterprise cloud-edge workflows. She stressed that enterprises want end-to-end, measurable value rather than demos, and that AI should amplify human experience, judgment and core competitiveness rather than replace people.

Li Guang, co-founder of TDengine, said industrial scenarios face three fractures in data, semantics and intelligence. TDengine has formed a data foundation from data collection, time-series storage, industrial ontology and real-time computing to AI agent runtime, and aims to close the loop from equipment anomaly sensing and data analysis to judgment and action. Through a Skills mechanism, it packages industrial analysis algorithms, industry knowledge and third-party system connection capabilities as pluggable modules. Li said industrial AI cannot be completed by one company and will require platforms, industry experts, equipment companies and ecosystem partners.

Wang Yu, founder and CEO of iPollo and chairman of Beijing's 'industry-education-evaluation' AI skills ecosystem chain, said the core changes in the AI era are the replication of capabilities and the parallelization of time. People can directly invoke encapsulated capabilities and let multiple agents work in parallel. What matters will shift from 'what I know' to 'what I can organize and invoke.' iPolloWork aims to build an AI Work platform for next-generation work, enabling new collaboration between agents and humans and among agents. AI can complete most basic work, but the last mile still requires human judgment, editing and calibration. The human role moves from execution toward goal definition, organization and final decisions.

Chen Xiaojun, AI product lead at Tianxiaxiu's Linggandao, said brand AI marketing in 2026 shows three changes: cognitive upgrading, budget restructuring and data infrastructure. As AI becomes a new information and decision entry point, brand competition has moved beyond being seen to being understood and cited by AI and entering user decisions. GEO, Agent and AIGC should not remain experimental items in traditional marketing budgets; companies need independent budgets and new ROI evaluation standards. Past marketing spending was often one-time consumption, while data can become a reusable enterprise asset. Tianxiaxiu is using an influencer database, marketing Agent and enterprise data middle platform to further AI-ify influencer selection, placement, monitoring and review. Chen said AI's significance goes beyond lowering content production costs; it can improve placement results and final ROI.

Zhou Qiuye, vice president of Shumei Technology, called 2026 a stage when AI security needs to truly come to the enterprise 'desktop.' When agents begin to call software, operate accounts, access enterprise data and even execute real actions, the boundaries of traditional cybersecurity change. He summarized enterprise problems as 'cannot see, cannot explain, cannot prevent, cannot manage,' and said companies need to move from 'can use AI' to 'dare to use AI': first see risks, then manage behavior, then build a governance mechanism covering the full lifecycle, and finally implement mandatory standards and compliance systems. AI security is a system engineering of governance, technology and operations rather than only a technical issue. As agents go deeper into business processes, prompt injection, agent poisoning, trust and permission risks will become long-term basic capability building.

Song Xuan, founder of Spoon AI x Spoon Classroom and former vice president of Xibei, brought the discussion into China's 5-trillion-yuan catering market, which is traditional and fragmented. Spoon AI has embedded catering industry know-how into agents and currently has 25 sub-agents covering strategy, brand, marketing, operations, menu and delivery, aiming to give catering practitioners without professional experience near-expert working ability. After entering the industry, Song concluded that 'in industry applications, AI is the least important part.' Catering companies do not care which advanced model is used; they care whether it brings growth and whether investment can calculate ROI. Even if owners understand AI, pushing frontline employees to use it faces resistance. The biggest threshold for vertical AI lies beyond the model in industry cognition, know-how, key scenarios, and finding people willing and able to use AI. 'People are the protagonists of application,' he said. As technical capabilities converge, the real barrier for AI industry applications returns to understanding the industry itself.

Li Jie, founding partner of Weture, focused on the content industry being rapidly restructured by AI. Compared with traditional industries such as catering, AI short dramas and animated dramas are built on large model capabilities from the start. AI has significantly reduced content production costs, shortened production cycles and further standardized the process from creativity to production to distribution. Content supply is expanding quickly, but a new question follows: when producing content becomes cheaper, where does the next stage of competitiveness lie? Li pointed to OPC and interactive content. AI will amplify individual creators' productivity and creativity, creating more 'super individuals'; content will also move from one-way viewing to real-time interaction. Users will not just watch a fixed story but influence plot direction through their feedback, making content truly personalized. In this vision, AI changes not only production efficiency but the form of content itself: from 'I produce, you watch' to 'users participate, AI generates in real time.'

Across the conference, speakers from different fields converged on a core consensus: AI value evaluation is shifting from 'can it generate' to 'can it deliver.' Discussions focused on tasks, processes, costs and results rather than merely displaying AI capabilities. Foundation and scenarios need two-way alignment: models, computing power, data and security form one side of infrastructure, while industry, brand, security, catering and short dramas form the other side of real scenarios; missing either end prevents scale. AI Large Model Works said the conference aimed to connect more than the guests and speeches on stage; it sought to place model vendors, infrastructure, AI platforms, security companies and industry practitioners on the same industrial map to find paths for technology to enter industries. When AI becomes easier to obtain, what is scarce may no longer be mastering a new technology but finding a real problem, organizing suitable capabilities, and turning it into a deliverable, verifiable result that someone is willing to pay for. The next stage of the AI industry is moving from single-point capability competition to more complex industrial collaboration. Models must combine with data, agents must enter real processes, applications must answer ROI, security must support scale, and industry know-how determines how deep technology can go. From 'can generate' to 'can deliver,' from 'how strong the capability is' to 'whether the result runs through,' the measure of AI competition is changing. The next differentiator may not be who is smarter but who can get things done and turn one success into sustainable, replicable industrial capability.