The Connectivity Tax Nobody Budgets For in Edge AIIssuing time:2026-11-10 18:08
The demo always works. A model runs on a dev board, detects the event, prints a confidence score, and everyone agrees the hard part is over. Then the device goes into a hospital corridor, a factory aisle, or a utility cabinet three kilometres from the nearest access point. It has to stay online for years, often on a battery, in a building full of other radios, without anyone physically visiting it. That is where most embedded-AI projects quietly stall. Not at the model. At the connection underneath it. The four constraints, restated as a billThe usual case for edge inference lists four constraints: latency, bandwidth, privacy, and availability. They are usually presented as architectural arguments for putting the model on the device. They are also, less obviously, a cost sheet. Latency is the cost of a round trip that some decisions cannot afford. A fall-detection event that arrives 800 ms late is a different product from one that arrives in milliseconds. Bandwidth is the cost of sending frames nobody will ever look at. Uploading everything from a 16-channel camera system is not a technical choice; it is a recurring invoice. Privacy is the cost of moving data across a boundary that regulators, hospital IT, or a factory’s own rules may not permit at all. In those projects the cloud is not a slower option, it is simply unavailable. Availability is the cost of a single point of failure. If the WAN drops and the local logic lives in the WAN, the site goes blind. Put the four together and the conclusion is not “edge is better.” It is that moving inference to the device moves the bill from the network to the hardware. The network invoice shrinks; a power budget, a radio plan, a key-management scheme and a fleet-update mechanism appear in its place. That substitution is the connectivity tax, and it rarely appears in the project estimate. Where the tax actually shows upFour line items, in the order they tend to bite. Power. A coin-cell device does not get to keep a radio awake. Every always-on listening window is drawn from the same budget as the inference itself. This is why a module’s sleep/wake behaviour matters more than its peak throughput in battery-powered products. Density. One device on a bench behaves nothing like sixty devices in one ward reporting at the same moment. Contention and queueing, not raw link speed, set the tail latency once deployment scales. Security. If the main processor is already running inference, adding TLS handshake and key handling in software steals cycles from the very task the product exists to do. Hardware crypto offload is not a checkbox here; it is capacity that gets returned to the application. Lifecycle. A model that works at launch has to be updated, rolled back, and versioned across a fleet that nobody wants to touch. Field devices make “we’ll SSH in later” an expensive assumption. None of these are model problems. All of them are decided at the module and gateway layer, which is usually chosen last and specified least.
What Wi-Fi 6 actually buysIt is worth being specific, because “supports Wi-Fi 6” says almost nothing about an edge-AI product. OFDMA and uplink/downlink MU-MIMO reduce contention when many devices report at once — the density line item. 1024-QAM lifts short-range throughput, leaving headroom for inference traffic where the link is good. TWT (Target Wake Time) lets a device agree on when it will be awake instead of listening continuously — the power line item. And a WPA3 / EAP hardware crypto engine, paired with an on-module MCU, keeps security and link management off the host processor — the security line item. Three tiers, one chainEdge AI hardware usually settles into three tiers, and they are cumulative rather than alternatives. The module keeps the device connected, powered down when idle, and authenticated. The edge box runs local inference and produces an immediate response — a 4K camera paired with a 1 TOPS NPU can decide locally and trigger a hardware screen-lock in under 300 ms, without a server round trip. The industrial server handles many channels at once, per-channel algorithms and regions of interest, and alert grading. Scale a project and it typically grows along this chain. It does not replace one tier with another. Designing only the middle tier is a common reason a pilot cannot be replicated across a site.
Five places the tax collectsThe same four line items appear differently across deployment types:
Different sectors, same conclusion: the architecture is chosen by which line item is largest, not by which model is most accurate. A four-question pre-mortemBefore the model is selected, four questions tend to predict whether the product will ship:
If the answers are vague, adding accuracy to the model will not fix the deployment. ClosingEmbedded AI is often described as a compute problem that moved closer to the sensor. In practice it is a connectivity problem that became visible. The inference happens in milliseconds; the device has to remain connected, powered and trustworthy for years. The teams that budget for the second part are the ones whose pilots become products. |