The architecture question I get asked most often by maintenance teams evaluating PdM software isn't about algorithms or sensor types. It's about where the computation happens. Edge inference or cloud inference? The honest answer is that it depends — but the deciding factors are more operational than technical, and most vendors aren't being straight with you about the tradeoffs.
This article breaks down the four dimensions that actually determine which architecture fits your plant: latency tolerance, network reliability, data governance requirements, and total cost of ownership over a 36-month horizon. We'll be direct about where edge wins, where cloud wins, and where a hybrid approach is the only defensible answer.
What "Edge AI" and "Cloud AI" Actually Mean in PdM
Let's establish definitions before the comparison, because the terms get used loosely.
Edge AI: The ML model runs on a compute device physically located at or near the monitored asset — typically a gateway appliance mounted in the control panel or on the machine frame. Raw sensor data is processed on-device; only anomaly scores, alert payloads, and compressed feature vectors are transmitted to a cloud backend for aggregation and model retraining.
Cloud AI: Raw sensor data is continuously streamed to a cloud platform where inference runs. The edge device (if any) is a protocol bridge and data buffer, not a compute node. Detection latency is bounded by the round-trip network path, not local compute speed.
Most platforms marketed as "edge" today actually run lightweight threshold rules at the edge and full inference in the cloud. Ask your vendor specifically: where does the anomaly detection model execute, and what is the minimum set of conditions required for an alert to fire locally?
Latency: How Fast Does Your Plant Need to Know?
For most rotating-equipment failure modes — bearing fatigue, gear tooth wear, imbalance progression — detection latency in the 1–5 second range is more than adequate. These faults develop over hours to days. A 3-second edge detection versus a 12-second cloud round-trip makes essentially no practical difference for maintenance scheduling.
Where latency becomes critical is in high-speed or high-consequence applications: a spindle bearing on a CNC machining center running at 12,000 RPM, a stamping press ram at 60 strokes per minute, or a pump in a food processing line where a sudden seal failure creates a contamination event. In these cases, an alert that fires 8–10 seconds faster can be the difference between a controlled shutdown and a safety incident or significant scrap event.
In our work deploying across facilities with mixed asset classes, we've found that roughly 70% of the monitored assets in a typical discrete manufacturer's fleet have failure modes where cloud inference latency is acceptable. The remaining 30% — high-speed rotating equipment, precision machining spindles, and line-critical drives — are where edge inference matters enough to justify the hardware cost.
Network Reliability: The Argument Most Cloud Vendors Skip
This is the dimension that most cloud-first PdM vendors underweight in their comparisons. Plant floor networks are not enterprise IT networks. They were designed to carry PLC ladder logic and SCADA polling traffic, not continuous 10 kHz vibration streams. Wireless coverage in metal-framed manufacturing facilities is inconsistent. Scheduled IT maintenance windows — typically occurring during off-shift hours — interrupt connectivity at exactly the time when lightly staffed overnight operations might be most at risk.
A cloud-only architecture has a specific failure mode: if the sensor stream can't reach the inference engine, no detection happens. This is distinct from a hardware sensor failure, which any system will miss. This is a connectivity gap that produces a false sense of coverage — the system is "running" but not detecting.
Edge inference eliminates this failure mode for real-time detection. The model runs on the local device regardless of network state. When connectivity restores, buffered data and alert payloads sync to the cloud backend. You don't lose detection coverage during network maintenance windows, Wi-Fi dead zones, or cellular backhaul congestion.
The counterargument from cloud-first vendors is that modern plant networks are reliable enough. For facilities with recently upgraded OT infrastructure, that may be true. For the median mid-size US manufacturer running a mix of legacy PLCs and patched-together wireless access points, it is not.
Data Governance: What Stays on Premises
A growing number of discrete manufacturers — particularly in defense supply chains, automotive OEM programs with IP-sensitive tooling, and medical device manufacturing — have explicit requirements about what sensor data can leave the plant network. Raw vibration data from a production line contains embedded information about machine operating parameters, cycle rates, and product characteristics that some customers treat as proprietary.
Edge inference architectures address this directly: raw sensor data never leaves the facility. Only anomaly scores and compressed feature summaries are transmitted to the cloud layer. For facilities that need to satisfy customer data governance requirements or internal IT security policies, this architectural property can be determinative — it's not a feature, it's a qualification requirement.
Cloud-only architectures require raw data upload by definition, which may require additional contractual review with OEM customers or IT security sign-off that can extend deployment timelines by weeks.
Total Cost of Ownership: The 36-Month View
Edge compute hardware — typically an industrial gateway appliance in the $800–$1,400 range per monitored asset cluster — adds upfront cost that cloud architectures avoid. Cloud inference platforms often have lower initial hardware spend but carry higher per-asset monthly subscription rates that compound over time.
A rough comparison for a 100-asset deployment over 36 months:
| Cost Component | Edge-First Architecture | Cloud-Only Architecture |
|---|---|---|
| Gateway hardware (per 8-asset cluster) | $1,200 × 13 = $15,600 | $0 (thin edge bridge only) |
| Software/subscription (per asset/month) | $18 × 100 × 36 = $64,800 | $28 × 100 × 36 = $100,800 |
| Data egress (estimated) | Low (feature vectors only) | Higher (raw stream upload) |
| 36-month total (approximate) | ~$80,400 | ~$100,800+ |
These are illustrative figures, not a vendor price list — actual costs vary by platform. But the pattern holds: edge-first architectures have higher upfront hardware cost and lower recurring subscription rates, which typically inverts around month 18–24 to favor edge on a total cost basis. Cloud-only is cheaper in year one and more expensive over a full contract term.
The Hybrid Recommendation
For most mid-size manufacturers in the US, the practical answer is a tiered architecture: edge inference for all assets with real-time alerting requirements, supplemented by cloud aggregation for cross-asset pattern analysis, model retraining, and ROI reporting. This is the architecture Gearcadence deploys by default, and in our experience it balances the latency and reliability benefits of edge compute against the multi-site analytics advantages that a cloud layer provides.
If your plant has consistent, well-maintained OT network infrastructure and no data residency requirements, a cloud-first architecture is a defensible choice for assets where latency tolerance is high. If your facility has aging wireless infrastructure, maintenance windows, or IP governance requirements, edge inference isn't optional — it's the only architecture that delivers the coverage guarantees you need to trust the system.
Ask the right questions before you sign a contract: Where does inference run? What happens during a 4-hour network outage? What raw data leaves the plant, and who governs it? The answers tell you more about fit than any feature matrix will.
