Private on-site vs managed cloud: where should VLM video analytics run?

Two honest options, not one-size-fits-all

Vision-language models can evaluate a camera scene against a plain-language rule. The question every IT lead and data protection officer then asks is the right one: where does that evaluation happen, and what leaves the building to make it happen?

Wyse Envision offers two deployment models and is explicit about the difference. Model A — Private On-site runs inference on a GPU-equipped edge device at your premises; inference media stays on site. Model B — Managed Cloud uses a CPU-only edge device that sends only approved frames or clips through Wyse Cloud to an approved cloud model, and returns the result to the site. Model B is offered as a controlled beta under explicit data-transfer approval.

The four questions that decide it

Question	Points to Model A (on-site)	Points to Model B (managed cloud)
Privacy and data flow	Footage must not leave the site; sensitive areas; contractual or customer restrictions on transfer.	Transfer of approved frames can be documented, approved and audited; provider and region disclosed per deployment.
Connectivity	Weak or metered links; sites that must keep checking when the WAN is down.	Stable link and low frame volumes; outlets without a server room.
Equipment	A place to house an edge GPU server and someone to own it locally.	No local GPU; a small device is enough.
Operating cost	Continuous, sustained checks across many cameras — local GPU is the economical path over a multi-year horizon.	Trials, bursts and small sites where convenience matters more than unit cost. Never positioned as cheaper than local GPU.

Data-flow diagrams for the two deployment models. The same diagrams appear in the privacy section of the homepage and in every readiness study.

What the PDPA conversation actually needs

Malaysia’s 2024 amendment to the Personal Data Protection Act added biometric data to the sensitive-data category, introduced processor security duties and breach-notification obligations. Two practical consequences for video analytics:

• Biometrics are off by default. Facial recognition and biometric search are not part of the standard Wyse Envision offer. Any biometric scenario would need its own legal basis and separate approval.
• Model B needs a documented transfer path. Before any managed-cloud pilot, the data-flow sheet records what leaves the site, which provider and region process it, how long it is retained and how it is deleted. Compliance is shared legal and operational evidence, not a feature label — which is why the privacy map is produced in the readiness study, not promised in a brochure.

Model A keeps inference media on site. Model B makes every transfer explicit, governed and auditable. Neither option claims that “all video stays in Malaysia” on your behalf; the study documents what is true for your site.

A worked way to decide, site by site

Multi-site operators rarely need one answer. A flagship factory or mall with a proper equipment room and continuous checks is a natural Model A site. A small outlet with no IT presence that needs one or two daily checks may be better served by Model B, with the transfer documented. The readiness study produces the recommendation per site, together with the N/L/A sizing: how many cameras are recorded, how many are watched live, and how many need continuous AI.

Either way, evidence, human approval and reporting work the same. The deployment choice changes where the model runs and what data moves; it does not change the operating discipline.