Cloud vs Edge AI for Factory Cameras in India: Bandwidth, Cost and Downtime

Cloud AI needs sustained upload most Indian plants do not have: sixteen cameras streaming 24/7 is about 32 Mbps uphill and roughly 10 TB a month. Edge inference on a box in your panel room works out near ₹120–₹275 per camera per month for hardware and power, keeps deciding when the link drops, and keeps worker video inside the plant.

If you run a 200–1,000 worker plant in Chakan, Tirupur, Rajkot or Ankleshwar, this decision usually gets made for you by a vendor slide that never mentions your uplink. It is worth twenty minutes of your own arithmetic, because the wrong architecture does not fail loudly — it fails quietly, on the day you needed the alert.

Key points

• Cloud inference is priced in bandwidth, not compute. Sixteen cameras at ~2 Mbps continuous is ~32 Mbps of sustained upload and ~345 GB/day — about 10 TB a month. The arithmetic is fixed and you can check it on paper: 1 Mbps continuous = 10.8 GB/day. Per-camera bitrates by resolution and codec: CCTV bandwidth calculator.
• Edge hardware is cheaper than the monthly it removes. NVIDIA lists the Jetson Orin NX series at up to 157 TOPS, power-configurable between 10 W and 40 W (NVIDIA Jetson Orin). Across 16 cameras over three years, a ₹60,000–₹1,50,000 industrial box amortises to ₹105–₹260 per camera per month, plus roughly ₹230/month of electricity for the whole box (~₹15/camera).
• The AI layer itself runs ₹1,000–₹2,500 per camera per month in India either way; cloud-hosted inference typically adds 20–40% for bandwidth and GPU time on top (see what a factory camera AI system costs in India).
• A dropped link costs you different things. Cloud loses the decision — no alert while the line is down. Edge loses only delivery: it keeps deciding and queues the message.
• Neither survives a power cut without a UPS on the PoE switch. In our Mihnovka pilot, the grid cut of 13 July 2026 ran 11:08 to 12:58 — the recording machine rode it on its internal battery, but the switch, router and cameras had no backup, so for 110 minutes there was nothing to record. Architecture does not fix that; a battery does (PoE budget and power).
• Storage location is now a written question. DPDP Section 16(1) lets the Central Government restrict, by notification, transfer to a notified country, and cross-border obligations commence 14 May 2027 (Section 16, DPDP Act 2023).
• Most mid-size plants land on hybrid — edge filters and decides, cloud stores a thin trail and watches whether the edge box is alive.

The three architectures, plainly

Full cloud. Cameras (or a bridge) push video to a data centre. Models run there. You pay per camera per month, forever, and your uplink is the product.

Full edge. A small GPU box sits in the panel room on your LAN. It reads every camera stream locally, runs detection and reasoning on-site, and sends out only text alerts and thumbnails. Video never leaves the plant.

Hybrid. The edge box does continuous detection and discards the overwhelming majority of frames. It escalates a handful of frames or short clips to a bigger cloud model when it needs judgement, and pushes an event trail off-site so the record survives a fire or a stolen recorder.

The choice is not ideological. It is decided by three numbers: how much you can push uphill, how much that costs per camera per month in rupees, and what you lose in the hours your link is down.

The bandwidth arithmetic Indian plants actually fail

Download speed is irrelevant here. Cameras push up, and most Indian business broadband is asymmetric and contended — a line that tests at 50 Mbps down rarely holds anything like that uphill against everything else the plant is doing. Test your own line with TRAI's MySpeed portal, at shift change, not at 7 a.m.

Three ways to feed a cloud model, for a 16-camera floor:

Planning figures, not measurements: ~15 fps, moderately busy scene, H.265. Smart-codec savings tend to shrink on a factory floor, where conveyors, forklifts and stack-lights keep much of the frame moving — verify against your own recorder's actual bitrate readout.

Row one is what "cloud AI" means when a vendor does not qualify it, and it is unbuyable on a normal Indian plant uplink: a symmetric dedicated line at that capacity is a serious recurring bill in most industrial estates, before you have paid for a single inference. Get that quoted before you believe any cloud proposal. Rows two and three are what the honest cloud vendors actually do — which means they are already half an edge architecture, because something on-site has to decide which frames matter.

That is the real finding: on Indian bandwidth, every workable "cloud AI" design puts a filter on the floor. The only question left is how much intelligence that filter has.

Rupees per camera per month

Planning bands, mid-2026, GST extra. Get live quotes.

Note what the table does not say: edge does not remove the analytics licence. It removes the bandwidth and the cloud premium on top of it. The honest edge number for hardware and power is roughly ₹120–₹275 per camera per month, on top of whatever software you buy.

Two things owners get wrong here. First, they compare hardware to hardware and conclude cloud is cheaper — it is, on day one, and only on day one. Second, they compare a per-instance cloud plan against a per-camera plan without checking how many streams an instance carries. Normalise every quote to rupees per camera per month before you compare anything.

The honest summary, and it is arithmetic rather than a promise: at 4–8 cameras and one shift, an edge GPU box is over-engineering and cloud usually wins. At 16+ cameras, or two shifts, or a second unit down the road, the removed bandwidth and cloud premium typically cover the box within a couple of years — run your own numbers with your own quotes before committing.

What happens when the link drops

Every Indian plant loses its internet. The question is what that costs.

Cloud: the model stops seeing. For the length of the outage you have no PPE alert, no idle-line alert, no gate event — the product is off. Footage may still be recording locally, but the decision did not happen, and nobody is going to review four hours of video afterwards. Many cloud VMS products degrade quietly before they drop — frame rate first, then resolution — so ask in writing whether yours raises an alarm when that happens.

Edge: the model keeps seeing. Detection, counting and alerting continue on the LAN. What breaks is delivery — the WhatsApp or Telegram message. A properly built edge agent spools those messages and drains them when the link returns, so you get the 11:40 alert at 13:10 with its original timestamp. Late is recoverable. Never is not.

Both: die instantly on a power cut unless the PoE switch is on battery. This is the most expensive lesson from our own pilot and it has nothing to do with cloud or edge. Cameras have no battery. When mains drops, they drop, even if the recorder is on a UPS — in our logged 110-minute cut the recording machine stayed up on its internal battery and recorded nothing at all, because one unbacked switch had taken every camera with it. Budget the UPS across recorder and switch.

One more asymmetry worth naming: a cloud service can tell you when a camera goes offline. A lone edge box does not, unless someone built a watchdog into it. That single feature — "who tells you it broke?" — is the strongest argument for keeping a thin cloud channel alive even on an edge-first design.

DPDP: where the video sits is now a written question

Recognisable worker video is personal data, and your factory — not your vendor — is the Data Fiduciary. Two provisions matter for architecture.

Section 16(1) empowers the Central Government to restrict, by notification, the transfer of personal data by a Data Fiduciary for processing to a country or territory outside India. Section 16(2) preserves any other Indian law in force that provides a higher degree of protection or restriction on such transfer (Section 16, DPDP Act 2023). So this is a negative-list mechanism, not a whitelist — but it is a list the government can extend, and your architecture has to survive that.

The timing: the DPDP Rules 2025 were notified 13 November 2025 and took effect 14 November 2025, with the substantive duties — notice, security safeguards, breach reporting, cross-border transfer and Significant Data Fiduciary obligations — commencing 14 May 2027 (Shardul Amarchand Mangaldas). You have runway. You do not have an exemption.

The practical pressure arrives earlier than the legal one: buyer audit questionnaires increasingly ask where footage sits. Separately, and on its own track, BIS/STQC Essential Requirements bear on CCTV sold in India from 1 April 2026 (BIS CRS implementation guidelines) — that is a hardware-certification rule about devices, not a data-location rule, and the two are often wrongly merged in vendor decks.

Edge inference sidesteps the storage question rather than answering it: if only text alerts leave the plant, there is no cross-border video transfer to argue about. That is a genuine architectural advantage. Pair it with a posted notice — see the worker CCTV privacy notice.

Note also that the labour law your consultant may still be quoting has changed: the four labour codes were brought into force on 21 November 2025, with the OSH Code 2020 replacing 13 central labour laws including the Factories Act 1948 (Acuity Law). Neither cloud nor edge changes your duties there, and neither code contains a blanket requirement to install cameras.

Where edge does not help — and where cameras do not help at all

Honesty is cheaper than a returned system.

Edge does not make a bad install good. A dark shed, a camera pointed at a wall, or a lens that cannot resolve a helmet at working distance produces the same wrong answer at 157 TOPS as at zero. Placement decides accuracy; compute does not. Start at how many cameras a factory floor needs.

Small on-box models are weaker than big cloud models. This is the real trade. Bounded tasks — helmet on/off, person in a danger zone, forklift crossing a pedestrian line, count on a conveyor — run well on an edge box. Open-ended judgement — is this line genuinely running or just vibrating, and why did output drop after tea break — is still better in a large model. That is why hybrid wins: filter on the edge, reason in the cloud on a handful of frames a day.

A single edge box is a single point of failure. No redundancy, no vendor NOC, no one watching. Ask what happens when it dies, and how you would know within an hour.

And cameras themselves have hard limits. They cannot see inside a machine cover, cannot detect bearing wear, cannot tell you why scrap rose on a die, and cannot read a mislabelled batch card. They will not fix a supervisor problem — they will only document it, and documentation you do not act on generates resentment without generating output. Cameras also raise a live industrial-relations question in any plant with an active union, and no architecture makes that go away.

How to decide, in one sitting

1. Measure sustained upload at the plant, during a shift. If you cannot hold your camera bitrate uphill for 24 hours, continuous cloud is off the table. Decision made.
2. Count cameras and shifts. Under 8 cameras, one shift: cloud. 16+ or multi-shift or multi-site ambition: edge-first hybrid.
3. Normalise every quote to ₹/camera/month, including bandwidth and amortised hardware.
4. Ask the outage question in writing: what still works when the internet is down for four hours, and what still works when mains is down for two?
5. Ask where video is stored, get the region in the contract, and prefer designs where raw worker video never leaves the plant.
6. Put the PoE switch on the UPS. Whatever else you decide.

This is the shape Mama is built to: an edge agent that reads the floor on-site, sends the owner one short message on WhatsApp, and spools it when the line is down. Before any of that, record a phone walkthrough of the floor and get back a camera plan — because the placement decision outranks the cloud-versus-edge one.

FAQ

Is edge AI cheaper than cloud AI for factory cameras in India? Past roughly 12–16 cameras it usually is, but not because the software is cheaper. An industrial edge box at ₹60,000–₹1,50,000 amortises to about ₹105–₹260 per camera per month over three years, plus ~₹230 a month of electricity for the whole box. The analytics licence — ₹1,000–₹2,500 per camera per month — is payable either way; cloud adds bandwidth and typically 20–40% on top of that licence.

How much internet upload do I need for cloud video analytics? For continuous 2MP H.265 streaming, roughly 2 Mbps sustained per camera — about 32 Mbps for 16 cameras, 24 hours a day, or ~10 TB a month. Event-only upload drops that to around 50 GB a month. Test upload, not download, and test during a shift, not at night.

Does DPDP require me to keep factory footage inside India? Not automatically. Section 16(1) lets the Central Government restrict, by notification, transfer to a notified country; Section 16(2) preserves stricter Indian laws. Cross-border obligations commence 14 May 2027. In practice, customer audit questionnaires ask about storage location well before a regulator does.

What happens to my alerts when the internet goes down? With cloud inference, the analysis stops — no alerts for the whole outage. With edge, detection continues on the LAN and alerts are queued, arriving late with their original timestamps. Neither survives a power cut unless the PoE switch and cameras are on a UPS, not just the recorder.

Can I run AI on my existing CCTV without replacing cameras? Often yes — an edge box can pull RTSP streams from existing IP cameras and add analytics without touching the recorder. The limits are resolution, lens and lighting: if a camera cannot resolve a helmet or a face-shield at working distance, no model fixes it. Note separately that CCTV sold in India from 1 April 2026 falls under BIS/STQC Essential Requirements, which apply per model rather than per brand — check the specific model with your supplier.

Should a 300-worker plant in Tirupur or Rajkot start with cloud or edge? Start hybrid. Put a modest edge box on the floor for continuous detection and keep a thin cloud channel for the event trail and an offline-alert watchdog. That gives you on-site resilience, off-site evidence, an honest answer on data location, and a monthly bill that does not scale with your uplink.