NVIDIA Transforms European Retail, From Digital Discovery to Physical Operations

NVIDIA published a roundup of European retail deployments built on its accelerated computing stack, spanning open-model post-training, digital twins and warehouse simulation. The announcements arrive alongside GTC Berlin. The flagship numbers come from Shopify, which post-trained NVIDIA Nemotron 3.5 Lightning for the GraphQL subagent inside Sidekick, its merchant assistant. On Shopify's internal benchmark, that model beats a leading frontier model on accuracy while cutting serving costs by 90% and answering 7x faster.

The Shopify work relies on NVIDIA Megatron Bridge for fine-tuning and NVIDIA NeMo RL for reinforcement learning. The GraphQL subagent fetches store data for several Sidekick features, including Sidekick Pulse, which researches a merchant's store and suggests improvements. Separately, NVIDIA CCluster vLLM inference optimization software lifted performance 4.5x and cut inference cost by up to 12x for Shopify across five other workloads: catalog enrichment, a merchant assistant, e-commerce shopping simulation, product taxonomy classification, and merchant and buyer recommendations. The pattern here is worth noting: post-training a smaller open model on a narrow task, rather than routing everything to a large hosted frontier model, is increasingly how teams hit cost targets on high-volume agent workloads. The trade-off is that such gains are usually task-specific and verified on internal benchmarks, so they do not automatically transfer to other workloads.

Zooplus, described as Europe's leading online pet retailer, rebuilt product discovery with Infosys and NVIDIA. A production-ready prototype inside Zooplus's AWS environment reached 86% search accuracy, using the NVIDIA Nemotron Retriever library for semantic search and NVIDIA's retail catalog enrichment blueprint to generate search-ready product attributes. Semantic search means matching on meaning rather than exact keywords, which matters when catalogs use inconsistent vendor vocabulary.

Photoroom's Image Editing API now standardizes every primary listing image for Vestiaire Collective, roughly 40,000 per day across 70 countries, at the point of listing. Listings processed by Photoroom sold at a 5% higher rate of liquidity, with 3% to 7% more unique views per product page. The pipeline uses NVIDIA GPUs with TensorRT and Triton Inference Server, and NVIDIA Dynamo Snapshot is positioned to help respond to demand spikes. For resale marketplaces, image consistency is a ranking and trust problem as much as a creative one, which makes the lift in views plausible.

Omi launched ProductDrop, a workflow that ties each generated image to an approved digital twin so brands can scale creative across full catalogs without misrepresenting products. L'Oréal is applying this within its CreAItech program, pairing Omi's digital twins with generated brand environments. Omi runs TRELLIS 3D asset generation on NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs, reporting 10x faster inference and 4x lower pipeline latency.

L'Oréal is also using NVIDIA ALCHEMI in its research workflow to screen ingredient combinations, moving from months of experimentation to hours of computational screening. It is testing the same approach with CurieOS, Orbital Industries' operating system for science and engineering, and OrbMol-v2, Orbital's open source molecular simulation model. Integrated with the NVIDIA ALCHEMI Toolkit on NVIDIA accelerated computing, OrbMol-v2 is five times faster than comparable models and has 21% lower error than the next-best model of its kind, according to the announcement. The workflow helps identify promising ingredient combinations for solubilizing UV filters before committing to lab testing.

Mirakl, which sells intelligent-commerce software, runs open models on NVIDIA GPUs on its own infrastructure and through European cloud providers such as Scaleway, an arrangement it frames as keeping technology sovereign and avoiding vendor lock-in. Its Catalog Transformer combines fine-tuned open models with proprietary services and has onboarded more than 70 million products, converting messy vendor data into organized catalogs.

On the physical side, Otto Group extended its autonomous warehouse under human supervision initiative, built with Reply on NVIDIA Omniverse libraries and running on Google Cloud, to its largest fulfillment center in Haldensleben, Germany. New capabilities let logistics teams design, test and optimize warehouse operations in simulation before anything is built. This is the digital-twin argument in concrete form: simulating multiple robotic fleets together is cheaper than discovering coordination problems on a live floor.

ALLSIDES launched a 3D Digital Twin Factory, an automated pipeline that turns physical objects into SimReady 3D digital twins in minutes with sub-millimeter geometric accuracy, measured physical properties and relightable materials that hold accuracy under any lighting. Those twins support simulated store layouts in 3D and reliable product tracking through the shopping journey and at checkout. The platform is built on CUDA, OptiX and TensorRT and already serves adidas, Amazon, Nike, Zalando and Zara.

CATCHES reported production results from a RealFit deployment with AMIRI: conversion up 55%, add-to-cart up 80% and gross sales up 56%, after seasonal adjustment against like-for-like styles. It launched PredictiveFit, extending physics-based virtual try-on to retailers, and CASI, a personal fashion agent that shops, styles and sizes for customers over WhatsApp using open weight models including NVIDIA Nemotron 3, built on Omniverse and the Newton Physics Engine.

At GTC Berlin, Nebius is showing three NVIDIA retail blueprints, all in early access and running on Nebius infrastructure: an Agent Harness for Employee Productivity that consolidates scheduling, task management, inventory alerts and operational insights for store managers; an Inventory Simulation and Optimization blueprint for modeling demand, testing replenishment strategies and tuning stock levels before committing; and an Agentic Finetuning Flywheel that generates training data through agentic workflows, fine-tunes open source models and feeds real-world results into the next cycle.

Why it matters: Teams building commerce and retail systems now have concrete, if vendor-reported, reference points for post-training small open models against frontier alternatives and for using simulation ahead of physical deployment. Developers evaluating cost per query, catalog quality or virtual try-on should treat these figures as vendor benchmarks to reproduce on their own data rather than guaranteed outcomes. The early-access blueprints signal that NVIDIA wants this work to run through its libraries and hardware, so lock-in and portability are worth weighing, in this reading, alongside the headline speedups.