·8 min read·

Edge installations: when on-prem is a feature, not a compromise.

Some systems belong on the edge. The cloud-first reflex is wrong more often than the industry will admit. Here is the framework we use to decide, the hardware we deploy, and the conversation we always have with clients before specifying.

Cloud-first has been the default for so long that 'we'll run it locally' sometimes sounds like a step backwards. It often isn't. For the kind of systems we build — retail installations, brand activations, kiosks, live events — edge is the right answer more often than the industry would have you believe. After ten years of installation work, we have a framework.

When edge is the right call

  • 01Privacy — pixels and PII never leaving the building.
  • 02Latency — interactive installations that need <50ms response budgets.
  • 03Resilience — venues where the internet is a suggestion, not a guarantee.
  • 04Cost — when the cloud bill for streaming raw data dwarfs the cost of a small box on-site.
  • 05Regulation — financial, medical, education environments where data residency is non-negotiable.
  • 06Compliance — venues with their own IT security policies that ban outbound traffic by default.

The edge-vs-cloud decision matrix

ConcernCloud-firstEdge-firstHybrid
Latency-critical interactionRiskyDefaultEdge for interaction, cloud for analytics
Identifiable / sensitive dataCompliance riskCompliant by designAnonymise at edge, send aggregates
Venue network is patchyBadDefaultEdge with cloud sync when available
Heavy data, light interactionDefaultWastefulEdge processing, cloud archive
Many distributed sitesEasier opsHeavier opsEdge per site, cloud control plane
Frequent software updatesEasierHarderContainerised at edge, push from cloud
Decision matrix — edge vs cloud vs hybrid for installation work.

Hardware we deploy

  • 01Mac mini M2/M4 — our default for any installation that needs MacOS-tied software, Final Cut workflows, or simple admin via Apple Remote Desktop.
  • 02Intel NUC / Beelink mini PC with Linux — when the workload is Linux-native or the client's IT has a Windows / Linux preference.
  • 03Raspberry Pi 5 for low-spec installations — signage, kiosk shells, IoT bridges.
  • 04Jetson Orin for on-prem computer vision — when we need GPU-class inference without the cloud round-trip.
  • 05Arduino / ESP32 for the bridge layer — the silicon that turns API calls into LED flips.

What the conversation looks like

We start every installation engagement with the same question: what has to leave this building, and what doesn't? The answer determines everything else — hardware spec, network design, what we even need an account in the cloud for. It also tends to make the legal team much happier.

On-prem isn't a step backwards. Sometimes it's the entire reason the project is possible.

— Internal installations playbook

Hybrid patterns we keep reaching for

  • 01Edge processing + cloud archive — heavy lifting happens locally, anonymised metrics ship up overnight.
  • 02Edge inference + cloud training — models train in the cloud, deploy to edge devices once a month.
  • 03Edge realtime + cloud control plane — devices run autonomously, ops team manages them from a central dashboard.
  • 04Edge capture + cloud transcode — raw video stays local, processed previews uploaded for review.
  • 05Edge fallback for cloud-first products — local cache means the UX survives a network blip.

Why teams resist edge

  • 01Operational pain. A box on-site needs visiting when it breaks.
  • 02Software update story is harder. SSH + containerised deploy is the answer.
  • 03Hiring and skill profile — fewer engineers have run real on-prem infrastructure in the last decade.
  • 04'The cloud is the default' — three years of conference talks have made cloud-first feel mandatory.
  • 05Asset management — keeping track of devices across many sites is a discipline most teams haven't built.

We solve most of these with operational discipline. Devices have a known build, a remote management agent, a clear update channel, and an inventory. Edge is harder to operate — and dramatically better for the user experience when it's the right call.

Talk to Remiam about a system like this.