How ZKH and Intel Are Solving the Million-Dollar Bearing Problem with Edge AI

SHANGHAI – November 02, 2026 — The modern factory floor is a marvel of automation, yet it remains vulnerable to the simplest of human errors. Consider the intake of maintenance, repair, and operations (MRO) supplies. A worker receives a shipment of industrial bearings. To the naked eye, two bearings may look completely identical, yet a micro-millimeter difference in a single specification dictates whether a part will smoothly operate a multi-million-dollar machine or cause a catastrophic production halt. For decades, verifying these minute details against delivery documents has been a tedious, manual process deeply prone to error.

This is the exact friction point that ZKH Group Limited, a leading Chinese MRO procurement platform, is targeting with its latest technological leap. Unveiled at Intel Connection 2026 in Suzhou, ZKH's AI subsidiary, LingLong Intelligence, has partnered with Intel to launch LingLong Insight 1.0. This 4-billion-parameter multimodal AI model is designed specifically for industrial edge computing, bringing advanced image and text understanding directly to the worksite.

By deploying AI locally on factory workstations rather than relying on distant cloud servers, the collaboration addresses a critical vulnerability in global supply chains. It also signals a broader shift in how industrial giants are leveraging proprietary data, moving from traditional e-commerce models to highly specialized, AI-driven hardware and software solutions.

The High Cost of Human Error in MRO

In the complex ecosystem of industrial manufacturing, the intake of goods is a notorious bottleneck. An undetected mismatch during the receiving process can allow the wrong part to slip into inventory. Weeks later, when that part is pulled for critical maintenance, the error is finally discovered—often resulting in delayed repairs, prolonged equipment downtime, and disrupted production schedules that can cost facilities tens of thousands of dollars per hour.

LingLong Insight 1.0 aims to eliminate this risk by combining multimodal AI with deep industrial knowledge. The system utilizes a visual verification workflow powered by industrial cameras. When a shipment arrives, the cameras capture high-resolution images of the delivery documents, packaging, and part labels. The AI then cross-references these images, extracting text and visual data to identify the items and check their specifications against the required documentation.

If the model detects a discrepancy—such as a serial number that is off by a single digit or a subtle visual difference in the part's casing—it immediately flags the issue alongside the relevant images. This allows workers to catch potential mismatches before confirming receipt, effectively neutralizing the error before it enters the facility's ecosystem.

"Industrial AI is only as good as the knowledge behind it," said Yang Liu, Chief Technology Officer of ZKH Group and CEO of LingLong Intelligence. "Our industrial data and close understanding of customer needs give us a strong foundation for developing AI that delivers practical value. With Intel's edge computing technology, we can now take that intelligence beyond the cloud and onto the factory floor, where every decision counts."

Intel's Core Ultra Play: Winning the Industrial Edge

The technological backbone of this initiative is the "LingLong Agent Box," a ruggedized edge computing device powered by Intel Core Ultra processors. This hardware choice underscores a massive strategic pivot occurring in the semiconductor industry. While hyperscalers dominate cloud AI, chipmakers are aggressively competing to power the "edge"—specialized, on-premise AI boxes that operate without internet connectivity.

Intel's Core Ultra processors are uniquely suited for this environment. They integrate a Central Processing Unit (CPU), a Graphics Processing Unit (GPU), and a Neural Processing Unit (NPU) into a single architecture capable of delivering up to 180 total platform Tera Operations Per Second (TOPS). For a multimodal model like LingLong Insight 1.0, this tri-architecture is crucial. The NPU handles the smaller matrices and complex neural network tasks efficiently, while the integrated GPU excels in larger matrix multiplications required for high-resolution image processing.

The economics of edge AI further justify this localized approach. While edge solutions typically require a higher upfront hardware investment—often ranging from $500 to over $1,000 per device for capable AI accelerators—they drastically reduce long-term operational costs. Cloud AI requires continuous expenditures for compute, storage, and data transfer. In contrast, high-volume, latency-sensitive workloads processed at the edge can drive the cost per inference down to fractions of a cent, with some industry implementations operating at roughly $0.000066 per inference in their first year.

Furthermore, processing information locally at the workstation eliminates network latency and reduces reliance on the cloud. In an industrial setting where speed is paramount and internet connectivity can be spotty or restricted due to security protocols, this responsive, on-site verification is a game-changer. For Intel, which reported $52.85 billion in revenue in 2025 amidst a highly competitive semiconductor market, capturing these industrial IoT workloads is a critical step in expanding its AI footprint beyond the traditional PC and data center markets.

From E-Commerce to AI Vendor: The Data Moat

While Intel provides the computational horsepower, the true differentiator of LingLong Insight 1.0 is its training data. ZKH Group has spent nearly 30 years operating as a leading MRO procurement service platform in China. Over three decades, the company has accumulated more than 27 million product records and approximately 4 billion product parameters.

In the AI arms race, generic computer vision models often struggle in specialized industrial environments because they lack domain-specific context. A standard AI might recognize a bearing, but it cannot inherently understand the complex taxonomy of industrial specifications required to distinguish between thousands of nearly identical variants. By standardizing its massive proprietary catalog and connecting it with expert knowledge, ZKH has transformed its historical sales and product data into a formidable strategic asset: structured, AI-ready knowledge.

This transition from a B2B marketplace to an agentic AI solution provider represents a strategic evolution that Wall Street is watching closely. Following the announcement of the Intel partnership on November 2, ZKH's stock experienced a significant surge, closing up 8.54%. Investors are clearly recognizing the high-margin potential of turning proprietary product catalogs into defensible software and hardware solutions.

The competitive landscape for industrial edge computer vision is already populated by established automation veterans like Cognex, Keyence, and Siemens. However, ZKH's approach—bundling a lightweight, 4-billion-parameter model trained specifically on its unparalleled MRO database—creates a unique value proposition. It bypasses the need for manufacturers to train their own models from scratch, offering an out-of-the-box solution deeply fluent in the language of industrial supplies.

Scaling the Autonomous Supply Chain

The deployment of LingLong Insight 1.0 at the receiving dock is merely the first phase of a much broader vision for autonomous industrial operations. LingLong Intelligence has already outlined plans to extend the model's capabilities to other critical bottlenecks, including warehouse inbound and outbound checks, production-line material verification, and equipment operation and maintenance.

This expansion aligns with a growing industry consensus that the future of manufacturing relies on hybrid architectures. While the cloud will continue to host massive, centralized data analytics and model training, the edge will become the domain of real-time, responsive execution. By bringing AI-powered support to more stages of industrial operations, facilities can systematically engineer human error out of their most repetitive and high-stakes processes.

Intel and ZKH intend to deepen their collaboration, combining Intel's computing platforms and global ecosystem resources with LingLong Intelligence's data and model capabilities. Their joint exploration of global market opportunities suggests that this localized, edge-driven approach to MRO verification could soon become a standard operating procedure in factories worldwide. As supply chains grow increasingly complex, the ability to instantly and accurately translate vast industrial knowledge into practical, on-site decisions will separate the resilient manufacturers from those left vulnerable to the costly consequences of a single mismatched part.