Instacart's real bet is not delivery. It is owning the grocery aisle's missing data

Instacart's Chief Connected Stores Officer, David McIntosh, told Forbes the company expects its sensor-equipped Caper Cart to "become the default way that people shop in store" ^[11]. The cart itself is the cheap part; what is being assembled is a store-level behavioural dataset that a grocer either funds and shares in, or watches a partner accumulate on its floor.

The gap being sold against is old and genuine. According to the Forbes account, an online retailer sees searches, abandoned baskets, purchases and repeat behaviour, while a physical store may know a product sold without knowing where the shopper found it, or why an item its system calls in-stock is absent from the shelf ^[6]. The environment is hostile to off-the-shelf vision: a large store holds tens of thousands of products, much of it in near-identical packaging, displays move, suppliers restock their own lines, and customers abandon items in the wrong place ^[7].

The hardware answer is a basket-facing camera set, outward-facing shelf cameras, a certified scale, location signals and an NVIDIA Jetson computer, fused into a single interpretation of what is happening in the basket ^[1]. McIntosh likens it to a small autonomous-driving stack and says inference runs on the cart, because supermarket connectivity is unreliable and a cloud round trip would introduce noticeable lag; responses land within hundreds of milliseconds ^[2]. That is a design decision with a capex consequence: intelligence per cart, not per store.

The defensibility argument rests on data no one else holds. Instacart says it has processed more than 1.6 billion lifetime orders, which it is feeding into grocery-specific models alongside catalogue data and in-store signals ^[8]. "Even frontier models have never seen the inside of a basket in a grocery store," McIntosh said ^[12]. The carts generate millions of sensor inputs a day, and the flywheel now extends to the shelf: Instacart acquired Arpalus to accelerate Store View, its computer-vision shelf intelligence product, while roughly 600,000 Instacart shoppers can capture shelf imagery on their phones ^[3][9]. The claimed payoff is better online fulfilment, fewer substitutions and recommendations that do not decay, since a suggestion for aisle five is worthless once the display moves to aisle seven ^[13].

Now the numbers that are not there. Instacart says thousands of Caper Carts are live across more than 100 cities and that the Caper business tripled year over year ^[4]. Spread thousands of units over a hundred-plus cities and the average is tens of carts per city, which is pilot density inside any national chain, not a default ^[10]. Tripling is also a ratio without a base. The Forbes account does not disclose cart cost, carts per store, or any retailer-side return figure ^[5].

What to watch: whether Instacart or its grocery partners start publishing store counts and penetration rather than city counts, since that is the only way to test the "default" claim ^[4][10]; whether shelf intelligence shows up as a measurable fall in substitution rates, the one metric a grocery operator can already price ^[13]; and how contracts treat ownership of in-aisle behavioural data now that it is the asset rather than a by-product ^[11][8]. The reporting here rests on a single interview with an Instacart executive ^[5], so treat the trajectory as vendor guidance until a retailer confirms the economics.