Walmart uses AI to plan what 4,700 stores need on the shelf

Walmart is using AI to help decide what fresh food its stores need, and it cut the cloud bill for that work from more than $100,000 a month to about $10,000. Engineer Aditya Mulik of Walmart Global Tech described the setup in a talk published by InfoQ. The forecasts run overnight in batches for about 4,700 U.S. stores.

Why the old forecasts fell short

Walmart already had machine learning models that study one or two years of sales history. They work well for steady patterns, such as Thanksgiving. They struggle with surprises. Mulik gave a storm as an example, when shoppers suddenly want water. Fresh items like rotisserie chicken and fresh beef spoil fast, so ordering too much means waste and ordering too little means empty shelves. The new AI layer is meant to add real-world context on top of those forecasts.

Where the savings came from

The team says the drop in cost came from switching to pay-as-you-go pricing instead of reserved capacity. Reserved capacity means paying to hold computing power whether it is used or not. A job that runs once a night uses little of it. The talk also covers problems the team hit, including data in the wrong format, mismatched data between departments and dropped network connections.

What is not known yet

The material available does not say how much more accurate the forecasts became, or how much of the work stores rely on today. The full details of the fixes are also not covered here. The cost figures come from the Walmart engineer's own account.