Preface:

At the 2026 World Internet Conference Wuzhen Summit, production rhythm, task closed loops, stable operation, and supply-demand matching began to vie with model capability for center stage.

The evaluation function of AI competition is being rewritten; looking to implementation has thus become the natural outcome after the technology race entered deep waters.

Author | Fang Wensan

Image source | Internet

From "models flexing their muscles" to "systems doing the work"

This shift is not a topic the conference organizers actively created; it is closer to a reality that has already taken hold in the industry.

The industry has reached a scale where it needs to find a path to fulfillment. The Ministry of Industry and Information Technology disclosed in October this year that in 2025 the scale of China's core AI industry exceeded 1.2 trillion yuan; across roughly 20 industries including steel, automobiles, and pharmaceuticals, more than 1,000 high-value application scenarios have been distilled, more than 5,000 enterprises have been organized for supply-demand matching, and the penetration rate of AI technology applications among manufacturing enterprises above designated size exceeds 30%.

It is therefore increasingly difficult for enterprises to tell a differentiated story simply by saying "we also have a large model." Alibaba's approach is to push competition toward AI infrastructure and Agent runtime environments.

During the 2026 World Artificial Intelligence Conference, Alibaba Cloud launched products such as AgentLoop and AgentTeams, incorporating agent development, deployment, monitoring, collaboration, and governance into a unified platform, while continuing to strengthen inference efficiency and the compute software stack.

This reflects a transformation in which cloud vendors are turning large-model competition into a new platform war: what customers ultimately need may not be a set of model APIs, but a production system that can keep a large number of agents running stably, scheduled, and managed.

Ant Group, meanwhile, has driven its approach down to business closed loops. Its embodied intelligence solution has already been deployed with Guoda Pharmacy, using robots for order receiving, sorting, and delivery, and attempting to use the same "robot general brain" to adapt to robots of different brands and configurations; its public information shows that LingBot-VLA2.0 has been adapted to more than 20 robot configurations from 17 manufacturers.

By the time it reached Wuzhen, the significance of the smart pharmacy demonstration was no longer whether the robot could grasp products, but whether it could enter existing store environments, work according to real orders, and integrate into existing business processes.

These corporate actions point to the same issue: the AI industry is moving from an oversupply of models into a stage where the ability to fulfill scenarios is scarce.

More and more people can train models, but those who can embed models into business processes, data permissions, payment systems, terminal devices, and organizational management systems have instead become the new scarce role.

Open source pushes scarcity downstream

The Ministry of Industry and Information Technology disclosed in July that the global cumulative downloads of China's open-source AI large models exceeded 10 billion. Joseph Tsai gave Qwen's figures at the opening ceremony: more than 470 models open-sourced, cumulative downloads exceeding 4 billion, and more than 330,000 derivative models. The ModelScope community screen read 270,000 open-source models, 33 million developers, and more than 3,000 institutions and individual contributors.

Before the summit opened, DeepSeek announced further comprehensive open-sourcing of its internal inference framework and simultaneously opened its codec components. Zhejiang Lab's 021 scientific foundation model completed training with 236 billion parameters and went online and open, pushing open-source supply from general capabilities to scientific research infrastructure.

In May 2026, the Cyberspace Administration of China, the National Development and Reform Commission, and the Ministry of Industry and Information Technology jointly issued the "Implementation Opinions on Standardized Application and Innovative Development of Agents," with four aspects of measures specifically listing the building of standard protocols, naming the Agent Interconnection Protocol AIP, and also listing 19 typical application scenarios.

The "Opinions on Developing New Quality Productive Forces" issued in October wrote "fully implement the AI+ action" into the industrial integration chapter. The same document contained a blunt sentence: it must not become detached from the real economy and drift toward the virtual, seek greatness and completeness, or create bubbles; those who cause major losses through rushing headlong and blind investment will be seriously held accountable.

The State Council executive meeting on October 16 directly explained the methodology: focus on using implementation applications to drive technological innovation, advance pilot-scale verification, and increase scenario cultivation. The Ministry of Industry and Information Technology subsequently released high-value scenarios for AI empowering new industrialization, covering 20 sub-sectors and more than 1,000 scenarios.

When foundation models become public goods that can be picked up at will, the room to collect rent from weights themselves is compressed, and scarcity will run downstream and toward both ends along the industrial chain.

At the summit's "AI Wuzhen Dialogue," many guests judged that as foundation model supply becomes easier to obtain, the competitive track is extending outward, and the compute foundation, development tools, industry solutions, and service ecosystem have become new battlegrounds.

As of the end of 2025, among the 748 generative AI services in China, 83.56% were industry models and vertical models for manufacturing, healthcare, and transportation, while general large models accounted for only 16.44%. The model layer has retreated to the position of the foundation, and the center of the battlefield has switched to scenarios and delivery.

Open source here is more like an industrial screwdriver: it wrings profit out of weights and twists it into places where the work can actually be completed, and conveniently pushes global competition from single-point chips and single-point models toward system-level engineering and ecosystem organization.

Three accounts turn "implementation" from an adjective back into a verb

① The token account: this expo placed supernodes from four vendors together for the first time: Alibaba Zhenwu M890 and Panjiu AL128, Huawei Ascend 950, ZTE Matrix cluster, and Sugon scaleX series.

Large-model inference is split into two stages, Prefill and Decode, with the bulk of cost concentrated in Decode, constrained by memory bandwidth, communication latency, and VRAM capacity. The peak compute table is least useful here; unit token cost is the yardstick.

Hardware vendors have begun comparing "how much of the compute in these cards is actually used," comparing effective compute utilization and total cost of ownership.

The other side of the ledger is the payback period. According to Q2 2026 financial reports, Tencent's capital expenditure was 52.78 billion yuan, up 176% year on year, and free cash flow turned negative.

Alibaba management gave a timetable on the earnings call: based on the current average gross margin, AI-related capital expenditure is expected to break even in roughly 3 years, with hope of compressing it to 2.5 years.

Baidu's core AI business revenue was 12.5 billion yuan, accounting for half of general business revenue; GPU cloud revenue grew 283% year on year, while online marketing revenue fell 19% over the same period.

The valuation anchor is changing. Beyond compute and bandwidth, Harness, scenarios, and ecosystems have begun to be priced in.

② The integration account: in a September 2026 survey by Deloitte China and the University of Hong Kong covering 205 enterprise executives, 64.4% of enterprises had entered production deployment, 48.5% reported that actual returns were below expectations, and 44.4% could not clearly calculate the ROI of AI projects at all.

The most prominent deployment bottleneck is integration with existing systems, at 56.6%. When evaluating results, 87.3% focused on processing speed and work-hour savings, while only 21.5% tracked revenue and customer-level metrics.

Gartner predicts that by the end of 2027, more than 40% of Agentic AI projects may be terminated due to unclear value and excessive cost.

Pilot-scale bases are public goods prepared for this account. On the first floor of the Wuzhen World Internet Science and Technology Museum, Zhejiang's national AI application pilot bases in six directions were exhibited together for the first time: healthcare, earth science, embodied intelligence, spatiotemporal geographic information, petrochemicals, and cultural tourism. Their locations are basically close to the most industrially dense places: embodied intelligence is in Hangzhou Binjiang, spatiotemporal geographic information is next to the Moganshan Geoinformation Laboratory in Deqing, and petrochemicals is in Ningbo.

The National Development and Reform Commission disclosed that more than 30 national AI application pilot bases have been laid out, promoting central and state-owned enterprises to open more than 1,000 application scenarios. If enterprises bear this trial-and-error cost alone, the vast majority of projects will die at the hurdle from the demonstration environment to the production environment.

③ The uptime days account: in the China Academy of Information and Communications Technology's "Embodied Intelligence Development Report (2026)" there is a rather blunt sentence: shipping does not equal taking up a post; for a considerable proportion of devices, average monthly uptime is less than 10 days, and order quality is low. The report calls the current stage the peak of inflated expectations and predicts that the first round of industry consolidation will be completed in the next 2 to 3 years.

Counterpoint statistics show that in the first half of 2026, global humanoid robot shipments were about 22,000 units, with entertainment performances, science education, and data collection together accounting for more than 60%, guided tours and shopping guidance about 19%, intelligent manufacturing 13%, and warehousing and logistics 5%.

Among Tongxiang's "Wuzhen Seven," Shuhan Technology connected its control large model into a coal chemical workshop, reducing raw coal consumption by 6% and increasing comprehensive annual benefits by about 170 million yuan.

Freetech's ODIN intelligent driving platform has established cooperation with 55 OEMs, with 432 designated projects accumulated and more than 300 projects achieving mass production; Sainade's autonomous loading and unloading robot has a maximum single-unit load of 80 kilograms and handles thousands of boxes of goods per hour.

Ant Lingbo's LingBot series models support robots of 17 brands and more than 20 configurations and have been put into use in multiple pharmacies in Pudong, Shanghai, with data from on-site debugging fed back to model iteration.

The three accounts point to the same issue: for every completed task, can the cost and effect be confirmed by signature?

The three hardest things in the convention: interfaces, metering, and responsibility

The forum "Global Cooperative Innovation for Agent Mutual Trust, Interconnection, and Interoperability" released the "Self-Discipline Convention on Agent Development," jointly initiated by the China Academy of Information and Communications Technology and 20 units including Huawei, Tencent, Ant Group, Alibaba, Baidu, Xiaomi, Volcano Engine, and Douyin, with 8 measures.

This is an engineering checklist: hierarchical and verifiable agent identities, protocol compatibility and mutual recognition, classified and tiered authorization and behavioral boundaries, rights, obligations, and responsibility allocation, and users' right to know, right to choose, and ultimate control.

The six words "commercially operable" weigh much more than the two words "mutual trust." The technical path for multi-agent collaboration has been drawn, but who bears the cost, how benefits are distributed, and who is responsible for errors still have no answers.

Without a risk pricing mechanism, implementation can only choose low-risk scenarios, so the tasks agents are most proficient at over the long term remain booking flights, checking reports, and organizing meeting minutes.

The positions in the industrial chain are therefore rearranged. Upstream is shifting from selling cards to selling system efficiency; supernodes, compute-power coordination, and unit token cost have become the subjects of quotations.

The midstream model layer has been compressed into thin profits by open source and price cuts, earning ecosystem rent through toolchains, developer communities, evaluation, and distribution entry points.

Downstream integration delivery, data governance, and industry knowledge encapsulation have become the thickest-profit yet most fragmented segment. What customers buy has changed from software licenses to business results, and pay-per-effect models have begun to emerge in contracts.

For competitors, the capability premium of the closed-source camp is being continuously squeezed, and the moat is moving to scenario data feedback and write-back into business systems.

After weights become free, compute power, evaluation standards, distribution entry points, and high-quality data become scarce again, and concentration will occur at another layer. Open source is more like moving the competition venue from upstream to downstream; it itself is not responsible for fairness.

6G end-to-end prototypes, satellite internet, quantum communications, and space-air-ground-sea collaborative networking are exhibits in the short term, but in the long term they are the connectivity layer laid for massive numbers of agents.

Conclusion:

The AI value chain is being restructured into new layers. The closer a segment is to industrial processes, the more it demands industry expertise, proprietary data, organizational access permissions, and long-term service capabilities.

These resources are not easily replicated at the speed of model weights, which suggests that AI competition may gradually shift from a phase driven purely by technological scale effects to one in which technological scale effects and industry depth go hand in hand.

AI is entering an operational phase, and the industry is moving toward "proving the business case"—and this round of competition will be slower and more brutal.

Some materials referenced: World Internet Conference: "Feel the 'Light of the Internet' Illuminating the Future in the Water Town of Wuzhen"; CCTV.com: "2026 Wuzhen Summit · Observation | The Future Is Here, 'AI Partners' Light Up Life!"; CCTV News: "Three Departments Issue a Document with Four Aspects of Measures to Promote Standardized Application and Innovative Development of Agents"; CNR: "Higher AI 'Content'! Come to the Wuzhen Summit and Join the Appointment with the Digital-Intelligent Future"; Zhejiang Daily: "Wuzhen Summit, Why 'Open Source'".