Instacart: Caper smart carts on Jetson and GPU ranking with Dynamo

NVIDIA's case study notes that physical stores still account for about 80 percent of grocery sales, while Instacart's data and recommendations were built for online orders. Smart carts must recognize items dropped into the basket in real time, including stacked or hidden items, across many store formats.

On the online side, ranking ads and items for each page is a latency-sensitive workload that Instacart ran on CPUs before the change described here.¹

On the cart

According to NVIDIA, each Caper Cart has a touchscreen, five cameras (three facing the basket, two facing the shelf), a certified scale, location sensors and an NVIDIA Jetson Orin NX module. Weight readings back up vision when items are blocked or stacked.

In the cloud

Instacart moved ad and item ranking from CPUs to NVIDIA Hopper GPUs with NVIDIA Dynamo, per NVIDIA's case study, which also says vision-language model encoders run asynchronously in the cloud and that a single vision-language model serves all stores with little store-specific retraining (it does not say that model runs on the cart). The Shelby Report names the serving layer NVIDIA Dynamo-Triton. Both describe a combined platform that links cart data with Instacart's online ordering history.

Announced in March 2026

At GTC 2026 Instacart described a grocery world model that connects in-store and online data; the Shelby Report describes it as being built, not finished. Retail partners named by NVIDIA include Kroger, Wegmans, Coles, Schnucks and Wakefern.

Instacart's own blog post on the Jetson carts could not be opened from our location on the review date (it redirects to a regional site), so the cart details here rely on NVIDIA and press sources.¹²

What others can learn: splitting the work between an edge computer on the device (vision plus weight sensing) and GPU serving in the cloud for ranking is a pattern any retailer with connected devices can study, and the Shelby Report credits the latency cuts to moving ranking from CPU to GPU serving, while it ties the click-through gain to a switch to a transformer-based architecture (NVIDIA groups both results with the GPU migration). Limits: Instacart's 1.6 billion past orders give it ranking signals few retailers have, so the sales lift may not carry over. All figures come from NVIDIA or press reports of Instacart statements, without baselines, test duration or absolute latency values, and the sources describe the sales lift slightly differently. Retailers with modest traffic may meet latency targets on CPUs.