Key takeaways
- Edge suits low latency, high-bandwidth signals (vibration, vision), offline operation and data that must stay on site.
- Cloud suits training, fleet-wide analytics, large models and use cases that tolerate seconds of delay.
- Neither edge nor cloud AI should replace deterministic control or safety functions. Those stay in the PLC and SIS.
- Edge devices are OT assets: they need inventory, patching, hardening and a place in the zone model.
- Model lifecycle at the edge (deploy, monitor, roll back, retire) is the hard part. Plan it before the first deployment.
Every industrial AI project eventually hits the same architectural fork: run the model next to the machine, or run it in the cloud. Vendors usually have an answer before you ask the question, and it tends to match what they sell. The right answer depends on the use case, and on what you can realistically govern with a small team.
This page compares edge and cloud AI on the factors that matter in a plant, then turns to the part most guides skip: managing models across many edge devices over years, not weeks.
Definitions, kept simple
- Edge AI: a model runs on hardware at or near the process. That might be a smart camera, an industrial PC on the line, an edge gateway at Level 3, or a server in the plant DMZ.
- Cloud AI: a model runs in a public cloud or central data center, fed with data sent from the plant.
- Hybrid: the common reality. Train in the cloud, infer at the edge. Or pre-process at the edge, analyze in the cloud.
"Edge" covers a range of locations across Purdue levels. A model on a camera at Level 1 has a very different risk profile from one on a server in the DMZ. Be specific about where it runs. See the Purdue model and AI.
The factors that matter
Latency
How fast must the answer arrive? Reject-gate decisions on a fast packaging line need milliseconds. An energy optimization recommendation can take minutes. Internet round trips add delay and variability you cannot control.
Determinism
Control systems need predictable timing. AI inference, especially on general-purpose hardware with operating system updates and variable load, is not deterministic in that sense. This is a key reason AI should advise or operate within bounds set by the PLC, not replace it.
Bandwidth
High-frequency vibration and camera images produce large volumes. Sending raw data to the cloud can be expensive or impractical on a plant's uplink. Processing at the edge and sending features, events or exceptions upstream is usually the answer.
Availability
If the internet link drops, does the use case stop? For a quality inspection gate, that may mean stopping the line or passing uninspected product. Cloud dependence must be a conscious decision with a defined fallback.
Data sovereignty and confidentiality
Process data can reveal recipes, yields and customer volumes. Images can include people. Some customer contracts, export control rules and privacy rules limit where data may be stored and processed. Edge processing keeps raw data on site; cloud processing needs a clear view of regions, sub-processors and vendor access.
Security posture
| Edge AI | Cloud AI | |
|---|---|---|
| Main exposure | Many devices inside OT zones | Data path out of OT, cloud tenancy |
| Key controls | Hardening, patching, inventory, local access control | Outbound-only DMZ path, identity, tenant security |
| Typical weakness | Unpatched devices, vendor remote access, default credentials | Path creep into write-back, over-broad vendor access |
| Who usually manages it | OT, often with OEM involvement | IT/cloud team, vendor |
Neither is automatically safer. Edge AI devices are OT assets and belong in your OT asset inventory. Cloud AI needs the governed outbound path described in moving OT data to a cloud historian. For the wider picture, see securing AI in OT networks.
Watch out: Vendor edge appliances often include a remote management agent that connects out to the vendor's cloud and accepts updates and commands. That is a remote access path into your OT zone. Review it as one.
Decision table
| Use case characteristic | Lean edge | Lean cloud | Notes |
|---|---|---|---|
| Response needed in under a second | Yes | Inline vision, fast reject gates | |
| Must run during internet outage | Yes | Quality gates, safety-adjacent alerts | |
| Raw data volume is very high | Yes | Pre-process at edge, send features | |
| Data must stay on site | Yes | Contract, confidentiality, privacy | |
| Training on months of history | Yes | Heavy compute, many sites | |
| Cross-site comparison or fleet models | Yes | Benchmarking, shared failure patterns | |
| Large language or generative models | Usually | Edge options exist but are constrained | |
| Few devices, small OT team | Often | Fewer boxes to patch | |
| Recommendation can wait minutes | Yes | Energy, scheduling, planning |
Most plants land on hybrid: train centrally, deploy compact models to the edge for inference, and send results and samples back for monitoring.
Cloud / central
training, fleet monitoring, registry
^ results, samples | signed models
| v
DMZ: model distribution, broker, logs
^ |
| v
Level 3/2: edge inference (gateway, IPC)
^ data | advisory output
Level 1: sensors, cameras, PLC (limits)
Model lifecycle at the edge
Deploying one model to one device is easy. Running twenty models on forty devices across three plants for five years is where programs struggle. Plan these before the first deployment:
Inventory and versioning
- Know which model version runs on which device, with what configuration.
- Keep a model registry: training data period, validation results, approver, deployment date.
- Record each model in the AI use case inventory.
Deployment
- Sign models and verify signatures on the device.
- Push updates through the DMZ, not directly from the internet to Level 2 or Level 1.
- Deploy in maintenance windows, with a tested rollback. You cannot always update a camera on the line mid-shift.
- Treat model changes that affect operations as changes under MOC.
Monitoring
- Monitor input data (drift in distributions, missing signals) and output behavior (alert rates, reject rates).
- Send a sample of inputs and outputs back for review, within your data rules.
- Set thresholds that trigger review, not silent retraining.
Retirement
- Decommission models and devices deliberately. Remove credentials, wipe data, update the inventory.
Plant-floor reality: A model on an edge device often outlives the project team that deployed it. If nobody owns its lifecycle, it keeps running on outdated assumptions long after the product mix has changed.
Common mistakes
- Choosing edge or cloud based on the vendor's platform rather than the use case.
- Putting AI in the control path without bounds in the PLC.
- Edge devices left out of patching and asset inventory.
- No fallback for cloud-dependent use cases during outages.
- Updating models over the air without MOC or rollback.
- Sending raw video or high-frequency data to the cloud and discovering the bill later.
Where to start
- For each AI use case, score latency, availability, bandwidth and data sensitivity using the decision table.
- Mark where inference would run on your Purdue model diagram.
- Check edge candidates against your patching and inventory capability. If you cannot patch them, reconsider or choose a managed option with clear responsibilities.
- Define the update path and rollback before deployment.
- Use the AI risk triage worksheet for any use case where AI output can affect the process.
- Add edge AI questions to your OT/AI vendor security questionnaire.
Effort: the decision itself takes a short workshop per use case. The ongoing effort is lifecycle management, which scales with the number of devices and models. Budget for it from the start.
What good looks like
Each AI use case has a documented reason for where it runs. Edge devices are inventoried, hardened and patched. Cloud paths are outbound-only from a DMZ. Models are signed, versioned and monitored, with rollback tested. AI outputs stay advisory or bounded by the controller. And when an outage hits, everybody already knows what the line does without the model.
Note: Data residency, privacy and export control requirements vary by country, contract and data type. This page is not legal advice; confirm specific obligations with counsel.
Sources and further reading
- NIST SP 800-82 Rev. 3, Guide to Operational Technology (OT) Security
- NIST AI Risk Management Framework
- ISA-95 Standard (ISA)
- MQTT Version 5.0, OASIS Standard
This guide is general information for manufacturers, not legal advice. Confirm regulatory obligations with qualified counsel.
FAQ
Frequently asked questions
When should AI run at the edge in a factory?
Run it at the edge when decisions are needed in milliseconds or a few seconds, when raw data volumes are too high to send upstream (such as high-frequency vibration or camera images), when the use case must keep working during an internet outage, or when data must not leave the site for contractual or confidentiality reasons.
Can edge AI be used for real-time control?
AI can advise or adjust within tight bounds, but deterministic, safety-related control should remain in PLCs and safety instrumented systems. Most edge AI hardware and software is not designed or certified for deterministic or safety duty. If an edge model influences control, enforce limits in the controller and run a full risk assessment and MOC.
What is MLOps at the edge?
It is the practice of deploying, versioning, monitoring, updating and retiring models running on devices close to the process. It covers how a model gets to the device, how you know which version is running where, how you detect performance degradation, and how you roll back safely during a maintenance window.
Is cloud AI less secure than edge AI for OT?
Not inherently. Cloud AI requires a governed path for data to leave the plant, and must never allow inbound connections into control zones. Edge AI keeps data local but adds many devices inside OT that must be patched and hardened. Each has a different attack surface. Choose based on the use case and your ability to manage that surface.
Related guides
Data Foundations
Moving OT data to a cloud historian: architecture, security and governance
Move plant data to the cloud through a one-way, outbound-only path from a DMZ, add context before it leaves the site, and treat every new data flow as a governed conduit with a named owner.
Guide · 7 min read
OT Security
Securing AI in OT networks: a practical guide for manufacturers
Treat every AI system that touches OT as a new conduit and a new piece of software you have to patch.
Guide · 8 min read
OT Security
The Purdue model and AI: where analytics and cloud fit in your plant
The Purdue model is still the clearest way to explain where things sit in a plant network and which way data should flow.
Guide · 7 min read
OT Security
OT asset inventory: the foundation for OT security and AI governance
An OT asset inventory is a maintained list of every device, system and connection in your control environment, with the attributes you need to make security decisions.
Guide · 6 min read