It is the question I hear in almost every first conversation: Should our AI run locally on the factory floor or in the cloud? The honest answer: it depends. But there are clear criteria that make the decision easier. In this post, I look at both approaches - edge computing manufacturing and cloud - and explain why I recommend a hybrid path in most cases.

The debate: on-premise AI vs cloud

When a manufacturing company wants to introduce AI, an infrastructure decision comes up early. Does inference - the actual evaluation by the AI model - run directly at the machine? Or are images and sensor data sent to a cloud server that returns the result?

Both approaches have their place. But the requirements in manufacturing are often different from a typical software startup. Cycle times, data privacy, and network reliability play a central role.

What edge computing means in practice

Edge computing manufacturing in concrete terms: an industrial PC or compact inference device sits directly in the control cabinet or at the production line. Camera images are captured locally, the AI model evaluates them in milliseconds, and the result goes directly to the PLC or MES system.

There is no sending data back and forth over the internet. All processing happens on-site, on-premise, within the company's own network.

Advantages of edge computing in production

Latency: In real-time inspection, milliseconds matter. When a workpiece passes the camera system at high speed, the evaluation must be completed within a few milliseconds. A round trip to the cloud takes significantly longer even with good connectivity and is not deterministic.

Data security: Image data stays in the plant. This is a central argument, especially for companies with strict compliance requirements. Many manufacturers simply cannot send production images to external servers - due to compliance rules or because the end customer contractually prohibits it.

GDPR compliance: When no personal data leaves the factory premises, the GDPR assessment becomes significantly simpler. This saves time and legal costs.

Independence from internet: Production halls do not always have stable WiFi or a reliable internet connection. Edge systems work completely offline. No internet means no AI downtime.

Deterministic response times: A local system delivers constant response times. This is critical for integration into existing automation workflows.

When cloud makes sense

The cloud has its strengths - but they lie elsewhere than real-time inference:

Dashboards and reporting: Aggregated production data, trend analyses, and quality dashboards can be conveniently deployed in the cloud. Multiple sites can be compared centrally.

Long-term analytics: When I want to analyze defect patterns over months, I need storage and compute capacity that goes beyond what a single edge PC can reasonably provide.

Model training: Training AI models - especially deep learning models - is compute-intensive and benefits greatly from GPU clusters as offered by cloud providers. Once trained, models are then deployed to edge hardware.

Scaling across sites: When a company wants to roll out the same model at ten locations, central cloud management can simplify distribution and versioning.

My approach: inference on edge, training in cloud

In my projects, I follow a clear principle: inference - the part that makes real-time decisions - runs locally on edge hardware. Training and optional long-term analytics can happen in the cloud but do not have to.

This approach is pragmatic because it leverages the strengths of both worlds:

Fast, reliable evaluation directly at the line
No dependency on internet connections during operation
Production data stays in the plant
A cloud backend for analytics can be added when needed

Important: the cloud component is optional. Many clients deliberately start without cloud and add it later when the need arises.

Hardware reality: what an industrial edge setup looks like

Edge computing in production does not mean consumer hardware in a server room. A typical setup includes:

Inference hardware: NVIDIA Jetson modules (e.g., Jetson Orin) or industrial IPCs with integrated GPU. These devices are designed for continuous operation in industrial environments - extended temperature range, vibration resistant, DIN rail mounting.

Industrial cameras: GigE Vision or USB3 Vision cameras, often with specific lighting for the respective inspection task.

Network: Local Ethernet is sufficient. No special cloud gateway needed.

Software stack: Optimized inference frameworks like NVIDIA TensorRT or ONNX Runtime that extract maximum performance from available hardware.

The form factor is compact. A complete inspection system fits into a standard control cabinet and can be integrated into existing lines without major modifications.

Cost comparison: edge hardware vs cloud subscription

I deliberately avoid citing specific prices here because they depend heavily on the use case. But the cost structure differs fundamentally:

Edge hardware:

One-time acquisition costs for hardware (industrial PC, camera, lighting)
Total costs for a single inspection system typically fall in the low to mid five-figure range
Ongoing costs are minimal: electricity, occasional maintenance
No monthly fees for compute

Cloud solution:

Lower initial investment
Ongoing monthly costs for compute time, storage, and data transfer
Costs scale with data volume and number of evaluations
Over several years, cumulative cloud costs can exceed the one-time edge investment

Experience shows: for permanent, high-frequency inference at a production line, edge hardware is often more cost-effective in the long run. For sporadic analyses or model training, cloud is more flexible.

The hybrid approach: the pragmatic middle ground

The best solution is rarely pure edge or pure cloud. The pragmatic middle ground looks like this:

Edge for everything time-critical: Inspection, sorting, real-time decisions
Cloud for everything analytical: Dashboards, trend analyses, model training
Clear data separation: Raw data stays local, only aggregated metrics go to the cloud
Incremental expansion: Start with edge-only, add cloud connectivity when needed

This approach gives manufacturing companies control over their data without having to forgo the benefits of modern cloud analytics.

Conclusion

Edge vs cloud AI is not an either-or question. For real-time evaluation in production, there is hardly a way around edge computing. For training and analytics, the cloud can be a valuable complement. I help find the right mix - tailored to the specific requirements, IT landscape, and data protection needs of each company.

Edge Computing vs Cloud: Where Should Manufacturing AI Run?