While AI floats in the cloud, factory workers reach for edge terminals

"The terminal will be the last mile of AI popularization"—I heard Lin Songtao say this exact phrase on the production line, except the speaker was replaced by a factory automation supervisor, with a tone carrying three parts expectation and seven parts helplessness. What they really want to ask is: How much does it cost to pave this last mile?

This morning, I stood next to a welding line, watching robotic arms repeatedly grab car doors with a cycle time of 44 seconds. A screen nearby displayed real-time AI visual inspection results, but the image had a 1.2-second delay. Workers said this 1.2 seconds is enough for a defective part to slip into the next process. This isn't a cloud AI problem; it's a glitch in the last mile from cloud to terminal.

Cloud AI vs. Terminal AI: Two Realities on the Production Line

Let me compare these two solutions on the production line from an integrator's perspective.

Dimension	Cloud AI Solution	Terminal AI Solution (Edge/Side)
Inference Latency	600-1200ms (including network + queuing)	30-80ms (local inference)
Single Point Deployment Cost	Hardware is cheap (no GPU), but requires dedicated lines and renting cloud compute	Hardware cost is high (terminal NPU/TPU), but no ongoing cloud fees
Reliability	Affected by network fluctuations, stops when disconnected	Can run offline, resumes after power outage
Model Update	Direct update in cloud, centralized	Requires OTA distribution, complex version management
Adaptation to Line Cycle Time	Usable if cycle >45s, completely unusable if <30s	Stable tracking even if cycle <15s

Key Numbers: I've seen a production line where the cycle time was compressed from 40 seconds to 28 seconds, and the cloud AI solution crashed immediately—because the closed loop of camera capture-upload-inference-return simply couldn't complete within 28 seconds. Finally, we switched to industrial cameras with NPUs, running YOLOv8n locally. Inference time dropped from an average of 800ms in the cloud to 42ms. The cost was jumping from 3,000 to 8,000 per camera, but it saved 120,000 yuan annually in cloud compute rental fees.

Have you calculated the integration costs?

Lin Songtao says "the terminal will be the last mile of AI popularization." I fully agree, but the accounts integrators need to calculate go beyond just equipment costs.

• Deployment Cost: For a cloud AI visual workstation, from laying network cables, configuring firewalls, adjusting cloud APIs to on-site debugging, our team averages 3 working days. For a terminal AI solution, just burn the model onto an SD card, screw it in, calibrate the camera, and it's done in 1.5 days. Labor costs reduced by 50%, which is more sensitive to integrators than the difference in equipment price.

• Maintenance Cost: Cloud solutions depend on the network. Once the factory network fluctuates (very common on production lines; EMI interference from starting welders can knock out WiFi), the line must stop waiting for AI recovery. Terminal AI solutions keep running as long as the local chip doesn't burn out and the model doesn't crash. Unplanned downtime reduced by 70%, according to data from a visual renovation project we did for an OEM last year.

• Model Update Cost: Cloud solutions only require changing the cloud API to update models, while terminal solutions need device-by-device OTA updates. But in actual production line conditions, model update frequency is far lower than imagined—most line AI models iterate only 2-3 times a year. Deploying an OTA system for this actually costs more than manually updating SD cards. So for most workstations, the maintenance cost of terminal AI is actually lower.

But Terminal AI is not a universal key

I've seen cases where all computation was stuffed into the terminal, resulting in chip overheating, frequent throttling, and slower cycle times. The bottleneck of terminal AI is not computing power, but heat dissipation and environmental tolerance. Industrial sites have temperatures of 40-50 degrees, dust, and vibration; consumer-grade chips simply can't withstand this. When doing integration, we must choose industrial-grade edge modules, such as the NVIDIA Jetson Orin NX industrial edition. A single module costs over 5,000 yuan, but it's the only solution guaranteed to run stably 24/7.

In contrast, although cloud solutions lose on latency, computing resources can scale elastically, suitable for scenarios with complex models requiring frequent retraining, such as production lines where defect detection categories change often.

Now back to Lin Songtao's judgment

If Tencent's terminal AI strategy stays only in consumer scenarios like phones and smart homes, I'm not optimistic. But if there are industrial-grade edge inference chips and accompanying model compression toolchains, that is the true solution capable of finishing the last mile. Currently usable industrial-grade edge AI solutions still have relatively high costs. The total cost per workstation (including camera, computing module, bracket, debugging) generally ranges from 15,000 to 30,000 yuan, whereas a cloud solution with just one industrial PC + camera can keep total costs under 8,000 yuan.

Reality is: Production line bosses will choose the cheaper solution until the cheap solution causes them to stop the line once. The loss from stopping the line once could be tens or hundreds of thousands of yuan. Only then do they remember the reliability premium of terminal AI.

When AI applications bloom everywhere, our production lines are still running 2018-era visual models in the cloud. It's not that we don't want to upgrade; it's that the cost-effectiveness of upgrading hasn't been clearly calculated. One final question for fellow integrators: If we push terminal AI solutions now, are customers willing to pay an extra 15,000 yuan per workstation in exchange for reducing downtime risk by 80%? I guess the answer depends on—how much is stopping their line once actually worth?

Original link: https://www.ithome.com/0/978/568.htm