Computer vision
Edge or cloud for video analytics: what it actually costs
The cloud version looks cheaper until you price the bandwidth. A worked comparison for a small multi-site deployment.
6 min readByteWeave Studio
The default instinct on a video analytics project is to send frames to the cloud. The tooling is better, deployment is a container rather than a site visit, and nobody has to think about hardware in a stockroom.
Then you price the egress and the always-on GPU, and the picture inverts. Video is the one workload where the arithmetic reliably favours doing the work where the data is.
The bandwidth nobody budgets for
A single 1080p stream at a usable frame rate is on the order of 2 to 4 Mbps continuously. Eight cameras at one site is roughly 25 Mbps sustained, every hour the store is open. Most retail sites have an asymmetric connection with far less upstream than that, shared with the point-of-sale terminals.
You can reduce it by sampling frames rather than streaming, and that is a legitimate design. But sampling at one frame every two seconds loses anything about movement or dwell time, which is usually most of what was being asked for.
What the two architectures actually cost
Cloud puts the cost in recurring spend: a GPU instance that has to stay warm because the workload is continuous, plus egress, plus the connectivity upgrade at each site. It is small at one site and scales linearly with every site you add.
Edge puts the cost up front. A small accelerated box per site is a one-off in the low hundreds of dollars, plus a site visit to install it and a modest ongoing cost for management and remote access. It is the more expensive first site and the cheaper tenth, and for most multi-site retail deployments the crossover arrives within the first year.
The reasons that are not about money
Edge keeps footage on site, which turns the privacy conversation from a negotiation into a statement of fact. That matters more than the cost argument in most retail and workplace deployments, and it is often what actually decides the design.
Edge also degrades better. When the connection drops — and at a retail site it will — an on-premise box keeps analysing and buffers its events. A cloud-dependent pipeline stops seeing anything, and the gap in the data is exactly during the incident somebody will later ask about.
Where cloud still wins
Keep training, model management and reporting central. Nothing about running inference at the edge requires you to also operate models by hand at each site; the box should pull a signed model bundle and report its version, and everything else stays where the tooling is good.
Cloud inference is also the right call for genuinely bursty or low-volume work — a few cameras, occasional analysis, or a proof of concept where you want an answer this week rather than a hardware order. The mistake is carrying that choice into a multi-site rollout without redoing the arithmetic.
- Computer vision
- Edge inference
- Architecture
- Cost
More reading
Computer vision
Running computer vision on CCTV cameras you didn’t choose
Nobody replaces sixty cameras to try analytics. What it takes to get useful results from mixed, badly angled, years-old hardware.
6 min read
Document AI
Why your invoice OCR works in testing and fails in production
The test folder is clean exports. The real intake is phone photographs. What changes between the two, and how to find out before a client does.
5 min read
LLM engineering
Why your RAG demo works and your RAG product doesn’t
A retrieval demo on twenty documents proves almost nothing about the same system on twenty thousand. The four things that break in between.
6 min read
Have a problem
worth solving?
Tell us what you're building. We'll help you figure out what's possible — and say so if we're not the right people for it.