banner

Wu Aibin: For Industrial AI, Value Delivery Is What Matters Most

Offline review 2026.09.11

Focusing on Value Delivery: Driving the High-Quality Evolution of Dual-Cross Platforms with AI Agents— Full Text of Wu Aibin's Keynote Speech at the 2026 Global Industrial Internet Conference

At present, the integration of the industrial internet and artificial intelligence has entered the deep-water zone. The industry has officially bid farewell to the extensive development model that "prioritized construction over real results" and entered a new stage of high-quality development where "AI values real results, and value speaks for itself."

At the 2026 Global Industrial Internet Conference, Wu Aibin, Chairman of Luculent, delivered a keynote speech titled "Focusing on Value Delivery: Driving the High-Quality Evolution of Dual-Cross Platforms with AI Agents", drawing on the company's hands-on experience in process industries and its full-spectrum AI implementation achievements, and offering insights and recommendations for the high-quality development of the industrial ecosystem.

The following is the full text of Wu Aibin's speech.


Distinguished leaders, fellow industry colleagues:

Hello everyone! I am Wu Aibin from Luculent. It is a great honor to be here today to share with you our practical thinking on the industrial internet and industrial AI.

As a national-level "dual-cross" platform, Luculent's Sunshine platform has long been deeply engaged in process industry scenarios that demand extremely high safety and reliability, including energy, coal mining, chemicals, and metallurgy. Over the past year, we have comprehensively advanced our "All in AI" strategy, making AI the core thread running through product iteration and project delivery, and we have reaped a series of achievements: the Sunshine platform earned the highest A-level evaluation for dual-cross platforms from the MIIT; the company was recognized as a National Manufacturing Single-Item Champion; and we secured first place in the leading group category of the National Industrial Internet Competition. Our self-developed Yaoguang vertical industrial large model has incubated more than 150 industrial agents that have achieved large-scale production deployment, delivered multiple group-level benchmark projects for large central SOEs, and we have also participated in the MIIT's high-quality industrial dataset initiative, achieving parallel progress across R&D, products, and implementation.

Drawing on industrial changes and frontline project practice, I would like to share from three perspectives: industry insights, enterprise practice, and industry recommendations.

I. The Industrial Internet Has Entered a New Stage Where "AI Values Real Results, and Value Speaks for Itself"

In the past, the industrial internet focused on connecting equipment and aggregating data, competing on the scale of construction. Today, the industry has entered a stage of high-quality development that pursues data value, AI scenario implementation, and business outcomes.

First, the value logic of dual-cross platforms has shifted — from "aggregating data" to "using data," with AI agents becoming the new carriers of platforms.In the early days, platform construction focused on accessing and aggregating data, but vast amounts of data lay dormant, with their value left unrealized. Now, the competition is no longer about data volume, but about what practical business problems the data can solve. AI agents are gradually replacing traditional report-based analysis, and industrial internet platforms are transforming from data reservoirs into capability foundations for agent-driven operation and decision-making.

Second, industrial AI has returned to rationality. Rejecting the simple bolting-on of general-purpose large models, trustworthy implementation is now a hard threshold.Process industries leave very little room for error. General-purpose large models suffer from hallucinations, unclear logic, and uncontrollable results, and putting them directly into production poses safety risks. Industrial AI must take root on the frontline, fully master process mechanisms, and achieve controllability, explainability, traceability, and quantifiability, truly serving production.

Third, the evaluation criteria have fundamentally changed — from valuing construction scale to valuing business outcomes.In the past, platforms were evaluated by the number of connected enterprises, devices, and applications. Today, cost reduction, efficiency gains, quality improvement, safety assurance, and manpower savings are the yardsticks that test the real effect of implementation.

II. Four Technology Foundations: Building a New Paradigm for AI Implementation on Dual-Cross Platforms

Keeping pace with industry changes, Luculent has deeply integrated AI into its platform architecture, product system, and full project delivery process. Relying on the Sunshine platform, we have built a four-in-one layout comprising a cloud-edge-device integrated foundation, an industrial AI operating system, the Yaoguang vertical industrial large model, and high-quality industrial datasets, charting an AI implementation path tailored to process industries.

First, build a solid cloud-edge-device integrated foundation to achieve closed-loop, controllable AI and group-level capability reuse.Process industry AI cannot stop at big-screen displays; it must go down to the frontline, connect with industrial control systems, and form a business closed loop. We confront two major industry pain points head-on. On one hand, process industries are wary of "black-box AI" — models trained purely on data cannot explain their reasoning logic, so plants dare not use them directly in production. On the other hand, the subsidiaries of large groups build and develop separately, resulting in massive duplicated investment, while proven experience is hard to replicate quickly.

To this end, we have built a cloud-edge-device integrated architecture compatible with DCS, PLC, SCADA, and ICS industrial control protocols, opening a secure channel between the management domain and the control domain.

  • At the end side, equipment sensing data is collected for operating-condition perception and on-site early warning;

  • At the edge side, low-latency computing, data cleaning, and on-site anomaly handling are performed, avoiding the risk of cloud network latency;

  • At the cloud side, group-level governance, model iteration, and cross-site situation analysis are focused.

The entire system enables end-side sensing, edge-side intelligent computing, cloud-side orchestration, and global collaboration. The optimization strategies output by agents can reach industrial control systems directly, completing the full closed loop of "data collection – intelligent analysis – decision output – production intervention – iterative optimization." We integrate decades of accumulated process mechanisms into the models, so AI conclusions are evidence-based, dispelling frontline concerns about black-box models. Thanks to the unified foundation, agents and algorithms validated at a single plant can be replicated in batches across all sites of a group, freeing us from the burden of "one project per plant."

Built on this foundation, we have implemented fan energy-saving and intelligent optimization scenarios at China Baowu Group; achieved closed-loop AI agent optimization at power plants under GD Power Development; built a penetrating safety control system for SPIC; and supported the construction of group-level platforms for China Datang Corporation, Huadian Coal, and Inner Mongolia Energy Group — truly achieving large-scale capability reuse across multiple sites of central SOEs.

Second, build an industrial AI operating system to turn fragmented AI into native platform capabilities.In the past, industrial AI was mostly bolted-on, scattered applications — fragmented systems that were hard to reuse. We have built an industrial AI operating system that integrates the time-series data foundation, digital twins, a mechanism model library, agent orchestration, and full-lifecycle model management, connecting the entire chain of data access, governance, reasoning, decision-making, execution, and iteration. AI is no longer scattered plug-ins, but native capabilities of the platform that are schedulable, reusable, and governable. Based on this operating system, we have incubated in batches agents for equipment O&M, safety compliance, and energy optimization targeting power generation, coal mining, chemicals, and steel, and they have already been deployed at scale in multiple large enterprises.

Third, the Yaoguang vertical industrial large model — a coupling of platform foundation, industry mechanisms, and vertical large model.The Yaoguang industrial vertical large model is deeply integrated with the Sunshine platform, more than 2,000 industrial mechanism models, a fault case library, and industry regulations. The platform handles data governance and scenario integration; Yaoguang handles complex task decomposition and multi-turn reasoning; and the mechanism knowledge base performs result verification and correction.

When anomalies occur in generating units or chemical facilities, agents automatically retrieve measurement points, ledgers, and maintenance records, conduct root-cause analysis in conjunction with mechanisms, and output reviewable, actionable handling recommendations — avoiding the hallucination problem of large models and constraining AI with industry know-how to make it trustworthy and usable. At present, Yaoguang has been deployed at multiple leading energy enterprises, with more than 150 agents continuously creating business value.

Fourth, build high-quality industrial datasets to solve the "cooking without rice" problem of industrial AI.In industrial scenarios, algorithms are not the biggest bottleneck; high-quality data is the key to large-scale implementation. Focusing on electric power, coal mining, chemicals, and metallurgy, we build datasets with complete operating conditions, precise labeling, and compliant, controllable governance, connecting the toolchain of collection, cleaning, labeling, and governance. Instead of blindly pursuing data volume, we prioritize the accumulation of effective samples from real operating conditions, providing raw material for large model training and agent iteration.

We participate in the MIIT's high-quality industrial dataset project, took the lead in establishing the Jiangsu Wind Power Equipment High-Quality Industrial Dataset Consortium, and are exploring industry-wide data co-construction and sharing to fill the industry's data gap.

On this basis, we have realized multi-agent A2A collaboration, where equipment, safety, and energy agents dispatch and empower one another, preventing the formation of new AI silos within the platform. At the same time, we have productized the models and solutions honed in competitions and transformed our delivery model — from simply selling software tools to delivering quantifiable business outcomes. In thermal power, coal mine, and steel projects, we have achieved tangible gains such as reduced unplanned equipment downtime, optimized energy consumption, and earlier risk warnings.

III. Jointly Building a Pragmatic, Trustworthy Industrial AI Ecosystem That Truly Delivers

Drawing on frontline practice, I have three recommendations for the industry's high-quality development:

First, uphold a value-oriented approach and drive the industry from "construction delivery" to "outcome operation." De-emphasize scale-based indicators and strengthen effectiveness evaluation; solutions should target real production pain points and be replicable, quantifiable, and commercially viable.

Second, jointly establish unified standards to consolidate the underlying foundation of the industrial AI industry. Industry partners should work together to build standards for high-quality datasets, compliant data circulation, and agent A2A interaction and dispatch, unify the specifications for foundations, models, and capability invocation, reduce duplicated construction, and break down ecosystem barriers.

Third, pool efforts in underlying technology breakthroughs to build a self-controllable, secure, and trustworthy industrial intelligence system. Continue to deepen core technologies including cloud-edge-device collaboration, industrial control adaptation, industrial AI operating systems, vertical large models, and trustworthy security systems, balancing mechanisms, data, and safety, and advance the industrial internet from simple application integration toward a foundation-level, system-level self-controllable upgrade.

Distinguished leaders and colleagues, the industrial internet is now at a critical juncture of deep integration between AI and the real economy. Luculent will stay true to the mission of dual-cross platforms, remain rooted in process industries, continue to deepen the "All in AI" strategy, and spare no effort in advancing the large-scale implementation of industrial AI agents. We look forward to joining hands with all of you to build a pragmatic, trustworthy, and high-value new ecosystem for the industrial internet!

Thank you all!


Return to List Page
Share on WeChat Moments ×
Open WeChat and click on "Discover" at the bottom,
Use 'Scan' to share web pages on your social media circle。