-- There is a line item in many 2026 retail AI budgets that, according to Mitchell Goldman, Chief Operating Officer at Edgify, most finance teams have not yet modelled. It is the cloud egress charge on video streamed from in-store cameras to a vendor's data centre for inference.
Goldman says a single self-checkout lane streaming 1080p video at 15 frames per second to a cloud inference endpoint produces roughly 50 gigabytes of egress per day. Across four lanes per store and 12,000 stores, he says the figure reaches the petabyte-per-day range. At public cloud egress list pricing, Goldman estimates the annual run rate reaches the high seven figures to low eight figures before a single inference decision is made.
Why Cloud-First Made Sense in 2019 and Does Not in 2026
Goldman says cloud-first retail AI was the right architecture in 2019 because it was the only architecture available. The compute did not exist at the lane, edge AI research was not mature, and privacy regimes had not yet made customer video a regulated input. In 2026, according to Goldman, every one of those constraints has changed.
On hardware, most POS terminals now ship with Intel i3 and i5 generation processors (or equivalent) with sufficient headroom to host a production computer vision model. ARM-based alternatives including ELO Backpack and Qualcomm Robotics platforms run inference at single-digit-watt power envelopes. New generation Single Board Computers (SBCs), for older devices than an i3, brings the cost of a per-lane inference box to approximately $300, according to Goldman.
On regulation, Goldman says CCPA, CPRA, Illinois BIPA, the Texas biometric law, and the EU AI Act's biometric data provisions all tighten restrictions on centralised vendor-cloud models. Federated edge AI, where raw video never leaves the lane and only model updates aggregate to a central server, is in Goldman's assessment the only architecture that maps cleanly onto the regulatory direction of travel.
Three Points Where Cloud AI Breaks at Fleet Scale
Goldman identifies three specific points where cloud-based AI fails at fleet scale in retail.
On bandwidth, Mitchell says a modern grocer's WAN is provisioned for POS, inventory sync, and corporate traffic, not continuous video streaming. He says the infrastructure upgrade required to support cloud computer vision at 12,000 stores rivals the entire AI software budget. Edge inference, he says, removes the bandwidth question.
On latency, Goldman says lane interventions need to happen in under 200 milliseconds for a cashier or self-checkout shopper to act before the basket closes. He says a cloud round trip with inference cannot reliably achieve that at scale, particularly in rural store geographies on consumer-grade connectivity.
On per-store data drift, Goldman says every grocery store sells a different long-tail SKU mix. A central model trained on aggregated retailer data averages over local variation and ends up, in his assessment, specifically good at none of it. Federated learning, he says, allows each store's model to reflect its own local distribution while contributing to and benefiting from the global model.
Edgify's founder Nadav has described the architectural commitment as follows: by 2030, the right place for a retail vision model is the same place the camera is.
Production Results
At a top-five UK grocer, Edgify says its federated stack achieves 99.9 percent recognition accuracy on the produce long tail with no customer video leaving the device. At a US natural-grocery banner, in-lane shrink interventions reduced loss-prevention incidents from 60 per week to two.
Goldman acknowledges that edge deployment carries its own costs, including hardware refresh cycles every three to five years and operational requirements around federated orchestration, version drift management, and secure aggregation. He says the total cost of ownership still favours edge over cloud at fleet scale.
Edgify's federated edge AI platform powers self-checkout, staffed lanes, and self-service scales across loss prevention, and produce recognition across top-tier US and European grocery retailers. For more information visit https://www.edgify.ai/.
Contact Info:
Name: Gianmarco
Email: Send Email
Organization: Xraised
Website: http://xraised.com
Release ID: 89197651
Google
RSS