Moving processing from a distant cloud region to a gateway on the same site cuts a typical control loop’s round trip from 20 to 100 milliseconds down to 1 to 10 milliseconds. That is the “seconds saved” in practice: not one big delay, but thousands of small ones that stop stalling machines, retrying radios, and queueing camera frames. The cost side is simpler. Raw sensor and video data stays on site, so you stop paying to store, process, and move data that nobody will ever look at. On a filtering-heavy workload such as camera monitoring, that routinely removes 80 to 95 percent of cloud data volume. The exact savings depend on how chatty your devices are, how far your cloud region is, and whether your loop actually needs sub-10 ms response. The rest of this article shows how to work those numbers out for your own site.
How much latency does edge actually remove?
Latency is mostly distance plus hops. Light in fibre travels about 200 km per millisecond, so a plant in Mexico talking to a cloud region in Virginia spends roughly 30 ms on pure propagation each way before any router, load balancer, or TLS handshake adds its share. Measured round trips from a factory floor to a public cloud region commonly land between 20 and 100 ms, and they jitter, which matters more than the average for a control loop.
A gateway on the same switch answers in under 1 ms on wired Ethernet and 2 to 10 ms over Wi-Fi or private 5G. Deterministic timing on top of that comes from IEEE 802.1 Time-Sensitive Networking, which schedules traffic so worst-case delay is bounded rather than merely low on average. Mobile operators expose the same idea through ETSI Multi-access Edge Computing, placing compute at the base station instead of the core network.
One correction to figures that circulate widely, including in earlier versions of this piece: NIST has not published field benchmarks showing “65 ms to 12 ms” reductions. Its SP 500-325 fog computing conceptual model defines the architecture and its properties. The percentage claims come from vendors, and the honest range is “from tens of milliseconds to single digits,” which you should measure yourself rather than quote.
Which workloads need it and which do not?
- Motion control and robotics: loops of 1 to 10 ms. Cloud is unusable here. Edge is mandatory, usually a PLC or industrial PC speaking OPC UA over TSN.
- Machine vision and safety: 10 to 50 ms. A camera that must stop a press before a hand enters it cannot wait for a cloud inference call.
- Predictive maintenance and dashboards: seconds to minutes. Cloud is fine for the analytics. Edge still helps by pre-filtering vibration or video data so you upload features, not raw streams.
- Billing, reporting, model training: hours. Leave these in the cloud. Edge adds cost without benefit.
If your slowest acceptable response is above about 200 ms and the data volume is small, you probably do not have a latency problem. You may still have a bandwidth or data residency problem, covered below.
Where does the cost saving come from?
Not where the old article said. Major cloud providers charge nothing for data coming in. On AWS, for example, inbound transfer is free, while outbound transfer costs around 9 cents per GB and standard object storage about 2.3 cents per GB per month, per the published on-demand pricing. So edge does not halve a “per-gigabyte upload” bill. It shrinks three other things: what you store, what you run compute against, and what you have to pull back out.
Video is the clearest case, and it is the one we deal with daily. A single 1080p camera at 4 Mbps produces about 1.8 GB per hour, or 43 GB per day. Ten cameras recording continuously push roughly 13 TB a month into storage that accumulates indefinitely. A gateway that runs motion or object detection locally and keeps only flagged clips typically retains 2 to 10 percent of that. The uplink is the other constraint: many industrial sites have 20 to 50 Mbps of upstream capacity shared with everything else, and ten continuous camera streams will saturate it.
| Item | All-cloud | Edge-filtered | What it depends on |
|---|---|---|---|
| Monthly video to cloud, 10 cameras | ~13 TB | 0.3 to 1.3 TB | Event rate, clip length, retention rules |
| Cloud storage at ~2.3¢/GB-month | ~$300/month, growing | $7 to $30/month | Retention period, storage tier |
| Site uplink needed | 40+ Mbps continuous | Bursty, usually under 5 Mbps | Whether raw retrieval is ever required |
| Gateway hardware | None | $250 to $3,000 one-time | Inference load, temperature rating, redundancy |
Downtime savings are real but site-specific, and the multi-million-dollar figures in circulation are not traceable to a published study. Price your own: line output per minute, times the minutes lost to stalls, times the share of stalls that trace to control latency. For a line making $2 per cycle at a cycle every few seconds, a 90-second stall is tens of dollars, and a bad shift has dozens of them.
Why do battery-powered sensors last longer?
The radio dominates power. A LoRa or cellular module draws tens to hundreds of milliamps while transmitting and microamps asleep, so battery life is set almost entirely by how often and how long it transmits. When the gateway is close, the link budget improves, packets succeed on the first try, and retransmissions disappear. When the gateway does the filtering, the node can send a summary every hour instead of a sample every second.
The gain therefore depends on the duty cycle. A node already reporting hourly gets almost nothing from edge processing. A node streaming raw vibration data can double its life. “Up to 40 percent” is a reasonable midpoint for a chatty sensor, not a law.
Does edge make compliance easier?
Yes, when it is designed for it. Chapter V of the GDPR restricts transfers of personal data outside the EU; Brazil’s LGPD has parallel rules. If faces, licence plates, or worker IDs are processed and discarded on site, and only counts or anonymised events leave the building, the cross-border question often never arises. For health data in the US, the HHS de-identification guidance describes what has to be stripped before data falls outside HIPAA’s reach, and a gateway is a natural place to strip it.
The trade-off is inventory. Fifty gateways in fifty buildings are fifty assets that hold regulated data, need patching, and can be stolen. Compliance gets simpler per record and harder per device.
Is local processing more secure?
It removes hops, and each hop is a place to intercept or misconfigure. But a gateway in an unlocked cabinet on a factory floor is a physical target the cloud never was. Treat it accordingly: hardware root of trust, signed firmware, encrypted storage, and per-device identity rather than shared credentials. NIST IR 8259 lists the baseline capabilities an IoT device should have, and NIST SP 800-207 describes the zero-trust model that stops a compromised gateway from becoming a tunnel into everything else. Use mutual TLS on the uplink; MQTT 5 over TLS is the common transport.
What hardware and software do people actually deploy now?
For light filtering and protocol translation, a Raspberry Pi 5 class board in an industrial enclosure costs under $150 and handles dozens of sensors. For on-site vision inference, the NVIDIA Jetson Orin family runs several camera streams of object detection on 10 to 25 watts, with developer kits starting around $250 and ruggedised units from integrators at $1,000 to $3,000. Fanless industrial PCs from Advantech, Moxa, or Siemens fill the gap when you need rated temperature ranges and DIN-rail mounting.
On the software side, the cloud providers’ own runtimes, AWS IoT Greengrass and Azure IoT Edge, handle deployment and fleet updates if you are already in that cloud. Lightweight Kubernetes such as k3s is the usual choice when you want to stay vendor-neutral. Models are quantised to 8-bit integers before deployment, which cuts inference compute by roughly three quarters at a small accuracy cost; measure the accuracy loss on your own data rather than trusting the headline.
Does 5G make edge unnecessary?
No. 5G cuts radio latency to a few milliseconds, but the packet still has to reach a server. If that server is in a cloud region 1,500 km away, you are back to 30 ms round trips. 5G and edge are complementary: the radio gets the packet off the device fast, and a local compute node answers it. Private 5G on site is itself a form of edge deployment.
What goes wrong in practice?
From the field, three failures recur. First, teams buy gateways before measuring and discover the bottleneck was a saturated uplink or a slow PLC scan cycle, not cloud distance. Second, nobody owns firmware updates, and two years later there are 40 gateways on 11 different versions. Third, filtering rules are set once and never revisited, so the edge quietly discards events that a new process needs.
Our own version of this problem is live and multi-camera video. Encoding on site and sending one compressed programme feed, instead of raw camera streams, is edge computing by another name. The lesson transfers: the hardest part is never the gateway, it is deciding what to throw away and being able to defend that decision afterwards.
Where should you start?
- Measure. Put timestamped probes at the sensor, the gateway position, and the cloud endpoint, and record P99 round trip for a week, not the average.
- Pick one loop that fails above 50 ms and one data stream that costs the most to store. Those are your two pilots.
- Deploy a single gateway, run both paths in parallel for a month, and compare stall counts and cloud invoices.
- Write the filtering rules down and assign an owner before expanding.
Do the measurement before the purchase order. A week of P99 data will tell you whether you have a latency problem, a bandwidth problem, or a residency problem, and each one points at a different piece of hardware.
