On-prem is the blocker
Whether the model runs in the building or in someone else's data center is now a procurement gate. In the sales calls it was raised by a meat plant, a clinic, a laundromat chain, an integrator in Nigeria and a telecom reading the EU AI Act.
The title is a line from a partner's own recap, written by the VP of partnerships at a national integrator group after a working session on a food-and-beverage manufacturer's project. The customer's head of manufacturing was aligned, phase two was approved, the metrics were clear and the controls team was engaged. Then: "On-prem is the blocker." The customer, in the partner's words, "would move today if it were fully local."
That sentence is the one buyers keep writing in different words. Where the compute runs decides whether the deal moves. The reasons vary by industry, and they are different enough to take one at a time.
The plant that keeps 90 to 95 percent on site
The senior OT lead planning a greenfield meat-processing plant told the vendor, according to the meeting notes, that roughly 90 to 95 percent of the company's compute runs on-prem and that the plant would keep 90 days of video retention locally. He runs an industrial-automation platform and wants to reach the video agents' API from inside it. His stated goal was a single pane of glass, where the alternative is five to ten separate dashboards.
That is a description of how an OT team, with robots and automated vehicles on the same network, has decided its plant will work.
The clinic that blocks external LLMs
A behavioural-health provider, an existing customer, uses a different gate. Its monthly touchpoint recorded that the organisation blocks external large language models outright and wants explicit safeguards for protected health information before any cloud-based AI touches its cameras. What the account team had to explain, in detail, was which processing happens on the edge device and how the cloud tenancy is isolated. Without that explanation, the summary says, there is no buy-in.
"No cloud upload of production video unless approved"
The VP of business strategy at a global animal-health manufacturer put the requirement in writing, as item four of a ten-point evaluation list: "On-premise or edge options, with no cloud upload of production video unless approved." Production video from a pharmaceutical-adjacent line shows processes, equipment and people the company would rather keep out of someone else's storage bucket. The word "approved" leaves a door open, with the default closed.
The integrator whose pipes cannot carry the stream
Bandwidth forces the same architecture for a different reason. The CEO of a retail loss-prevention integrator in Nigeria wrote:
Nigerian bandwidth can't sustain continuous per-camera cloud streaming, so full-VMS deployments don't work here. Our architecture solves that, an on-site edge device triggers on candidate events and ships only short clips for cloud analysis.
His question to the vendor followed from that: "could your AI agents analyze short event clips we send and return a judgment on retail theft behavior (shelf, concealment, walkouts)?" He is shopping for a model he can send a few seconds of footage to, and nothing else.
A monitoring provider that hosts thousands of sites gave the numbers behind the same instinct: 500 to 700 kilobits per second per stream, and for sites with 60 or more cameras a preference, per the notes, to host the server on-prem.
The laundromat that wants the camera to do the thinking
At the small end, an operator of self-service laundromats in ten European countries wrote that the company wants to "Work with built-in AI analytics on the camera whenever possible (without a dedicated AI server)." The job is to "detect people who enter the laundromat shortly before closing time and remain inside for an extended period, potentially staying overnight." For an unstaffed 24-hour site, an extra box on the wall is an extra thing that can fail.
The telecom asking about the EU AI Act
A European telecom's investment and business teams brought a legal frame. Per the summary, they needed any AI they deployed to comply with EU AI Act guidance, including restrictions on facial recognition and on combining licence-plate data with other records, and they asked where exactly the on-premises appliance would sit and what the edge device does.
What these have in common
Six buyers, six reasons: compute policy, health-data rules, production secrecy, bandwidth, hardware simplicity, and law. The answer that seems to satisfy most of them has the same shape. Detection runs on the box, video stays where it was recorded, and what crosses the network is a clip or a count that someone has approved. The partner's recap ended with the next step, a call to "work the approval path." That path runs through whoever owns the rule, and none of these buyers described that person as sitting in the security department.