byASB Ankiit Singh Book a scoping week
← Work/Tools
INDICATIVE MODELNot a quote

Edge or central? The three-year number.

Per-camera edge inference nearly always looks more expensive on the capex line. Then you price three years of continuous archival against event-triggered retention, and the answer can invert. This model shows you roughly where that happens — with your numbers, not mine.

Directional only. Every figure below is an editable assumption, and no two deployments price the same. Treat the output as a starting position for a real costing exercise, never as a budget.
Assumptions

Deployment

Unit rates

Central GPU

Per-camera edge

Indicative result

—

—

Central GPU — —
Per-camera edge — —

What would have to change to flip this

one input at a time, everything else held
What this model does and does not count

Included

  • Retention policy is independent of where inference runs, and that matters more than anything else here. Keeping events only, archiving everything centrally, and archiving everything on the camera are three different decisions — the model applies whichever you pick to both architectures rather than quietly assuming central hoards footage and edge does not.
  • Storage. bitrate × 86400 ÷ 8 ÷ 1000 GB per camera per day when continuous; event clips only otherwise. Edge-local archiving replaces the storage bill with a per-camera disk.
  • Bandwidth. Central always carries every stream continuously — cameras × bitrate sustained — because the video has to reach the GPU to be looked at. That cost is unavoidable and it is usually the whole story. Edge carries event clips only, averaged over the day and multiplied by 3 for peak headroom, unless you are archiving continuously to the centre, in which case it carries everything too.
  • Compute. Central amortises GPU servers over the horizon. Edge is one device per camera.
  • Power at 24/7 for both, licensing per camera per year, and maintenance — a percentage of capex centrally, and a field failure and swap rate at the edge.

Excluded on purpose

  • The cameras themselves and their cabling — identical in both architectures, so they cancel out of a comparison.
  • Integration and commissioning labour, which is genuinely large and genuinely impossible to model generically.
  • Data-centre floor space, cooling overhead beyond device draw, redundancy and failover. All of these push the central number up, not down.
  • Regulatory and data-protection cost. Where retention rules apply, they usually reshape the architecture before the economics do.

Why the output is a band, not a number

The inputs that dominate the result — real bitrate under motion, actual event rates, negotiated bandwidth and storage rates — vary by more than they look like they should. A model like this is reliable about which architecture wins and roughly by how much. It is not reliable to the nearest thousand, and any tool that tells you otherwise is selling something.

Where it came from

The structure comes from a costing exercise for a 1,400-camera ANPR and face-recognition deployment — edge inference per camera against centralised GPU processing, with event-triggered rather than full archival retention. That project is written up here.

The useful next step

Send me your real numbers.

Defaults are guesses. If you're actually costing a deployment, send this scenario over and I'll tell you where the model is wrong for your case — the assumption that usually breaks it is not the one people expect. Your inputs are attached automatically.