Why the edge box is winning the argument over the cloud
The cloud was supposed to eat everything. For video agents it lost on arithmetic: a facility with 64 cameras produces 30 terabytes a month, and a lock has to fire in under a second.
The pitch was that everything would move to the cloud, cameras included. Ship the pixels up, run the models in a data center, send the answer back. For dashboards and search, that worked. For agents that have to act, it ran into two numbers.
The first number is 30 terabytes
Spot AI's engineering team gave the figure in a technical post: a facility with 64 cameras generates over 30 terabytes of video a month. Multiply by a few hundred sites and you are looking at a bandwidth bill and a storage bill, both of which grow with every camera you add, before a single model has run.
The edge box turns that around. Video stays on site. What leaves the building is metadata, an event or a count, and the occasional clip. Spot AI's VP of engineering, Amrish Kapoor, describes the company's design as "edge-first, cloud-native," with on-site devices that "detect events on site and surface metadata for cross-site reporting." When the company launched its agents in 2024, it said the edge hardware carried three times the compute of a traditional AI camera, specifically so newer foundation models could run there.
The second number is one second
An agent that locks a door or stops a conveyor cannot wait for a round trip to a region a thousand miles away. The useful way to think about it is a latency budget: the time from capture to encode to inference to decision to actuation to network. For a robot arm, that budget is tens of milliseconds. For most security and safety work it is one to two seconds, still far tighter than a person watching forty screens.
Those budgets decide where code lives. Low-latency control belongs on the edge. Long-horizon reasoning, the part that weighs a scene against a policy or summarises a shift, can tolerate seconds and belongs in the cloud, where the large models are.
This closely aligns with Qualcomm's vision of implementing AI on edge to solve real-world challenges.
That was Tushar Gupta of Qualcomm Ventures explaining the chip company's investment in Spot AI. Qualcomm sells edge silicon, so the investment tells you where the company expects the next decade of inference to run.
The job splits
The cloud keeps half the work. Kapoor's post argues for exactly this hybrid at industrial sites, and the same split is now common across the category: a small model on the box for cheap, always-on detection, and a large model upstream for the questions that need context. The edge answers "is a person in the forklift lane right now." The cloud answers "how did lane incidents trend across nineteen plants this quarter."
The consequence I keep coming back to is for the people who install cameras for a living. Once the value sits in the box and the model, the integrator's job shifts from running wire to tuning agents, which I think is a bigger change to the industry than any product launched this year. It follows from 30 terabytes a month and one second.