NVIDIA shows how European retailers are shifting from AI “discovery” to physical operations with open models and digital twins

NVIDIA is pitching a new playbook for retail technology in Europe, one that connects AI model performance directly to physical operations in warehouses and stores. In a blog update published Tuesday, the company described how retail and commerce companies are pairing open models and NVIDIA-accelerated computing with “digital twins,” or virtual replicas of real-world environments, to move from purely digital improvements to changes that can be tested and executed in operations.

A central example is Shopify, which NVIDIA says used open-model post-training to reach accuracy it described as surpassing a leading frontier model on Shopify’s own internal benchmark. NVIDIA also reported that Shopify reduced serving costs by 90 percent and increased response speed by 7x for the Sidekick merchant assistant, where a GraphQL subagent retrieves store data to power features including Sidekick Pulse, which reviews a merchant’s store and recommends improvements.

NVIDIA added that Shopify’s work extended beyond Sidekick. With NVIDIA CCluster vLLM inference optimization software, the company said Shopify achieved 4.5x faster performance and up to 12x lower inference cost across five additional e-commerce use cases, including catalog enrichment, merchant assistant capabilities, shopping simulation, product taxonomy classification, and recommendations for merchants and buyers.

In product discovery and retail catalog enrichment, NVIDIA cited Zooplus, which it said rebuilt its product discovery stack for the agentic era. The blog described a production-ready prototype deployed in Zooplus’s AWS environment that reached 86 percent search accuracy using NVIDIA’s Nemotron Retriever library for semantic search, alongside NVIDIA Blueprint technology for generating search-ready product attributes.

NVIDIA also pointed to image and creative workflows. It said Photoroom’s image editing API standardizes primary product images, processing about 40,000 images per day across 70 countries at the point of listing. NVIDIA reported that listings processed by Photoroom sold at a 5 percent higher rate of liquidity, with 3 percent to 7 percent more unique views per product page, while running on NVIDIA GPUs using TensorRT and Triton Inference Server, and using NVIDIA Dynamo Snapshot to help respond to demand spikes. In a separate example, NVIDIA described Omi’s ProductDrop workflow that “anchors” generated images to an approved digital twin to scale creative output across catalogs without introducing product errors, citing L’Oréal’s use through its CreAItech program.

The company’s blog tied these creative and discovery improvements to scientific and materials research used by some large consumer brands. It described L’Oréal accelerating ingredient-combination screening by integrating NVIDIA ALCHEMI into its research workflow, shifting from months of experimentation to hours of computational screening. It also described testing an AI-based molecular simulation approach, CurieOS paired with Orbital Industries’ OrbMol-v2 open model, saying the workflow delivered faster runtimes and lower error relative to another model, supported by NVIDIA ALCHEMI tooling and accelerated computing for identifying promising combinations before laboratory testing.

Beyond consumer-facing AI, NVIDIA said the same model-and-twin approach is being used to simulate and optimize physical operations. It cited Otto Group, which NVIDIA said advanced an autonomous warehouse under human supervision initiative built with Reply using NVIDIA Omniverse libraries and running on Google Cloud, deploying new capabilities at the group’s largest fulfillment center in Haldensleben, Germany. NVIDIA framed the effort as letting logistics teams design, test, and optimize warehouse operations in simulation before physical implementation.

NVIDIA also highlighted new digital twin infrastructure built for retail and brand use cases. It described ALLSIDES launching a “3D Digital Twin Factory,” an automated pipeline that transforms physical objects into SimReady 3D digital twins in minutes, capturing detailed geometry and physical properties for accurate simulation under different lighting. The blog said the resulting twins can underpin simulations of store layouts in 3D and support tracking products reliably throughout the shopping journey. It added that the platform is built on NVIDIA CUDA, OptiX, and TensorRT, and that it already serves brands including adidas, Amazon, Nike, Zalando, and Zara.

Rounding out the update, NVIDIA cited fashion and sizing optimization deployments, including CATCHES. NVIDIA said CATCHES produced results from its RealFit deployment with AMIRI, including a 55 percent uplift in conversion, 80 percent higher add-to-cart rates, and a 56 percent gross sales increase after seasonal adjustment. NVIDIA also described CATCHES extending physics-based virtual try-on through PredictiveFit and launching CASI, a personal fashion agent that shops, styles, and sizes for customers over WhatsApp, powered by open-weight models including NVIDIA Nemotron 3, built on NVIDIA Omniverse and the Newton Physics Engine. The blog also mentioned that at GTC Berlin, Nebius is showcasing three NVIDIA Blueprints for retail, available in early access and intended to be run on Nebius infrastructure, accelerated by NVIDIA.

While NVIDIA’s examples provide specific performance figures, the blog does not offer independent verification or consistent third-party benchmarks across the different deployments. Several results are described as internal benchmarks or production outcomes within the specific partner environments, and the company does not detail contract terms, data provenance, or how results might vary by retailer size, product mix, or regional traffic. Still, the overall throughline is clear: open and frontier models are being optimized not only for accuracy and cost, but also for integration into workflows that connect digital replicas and simulations to real operational decisions.

In the near term, retailers and logistics operators will likely watch for two things: whether these open-model stacks can be sustained at scale without operational overhead, and how quickly digital twin-based simulation translates into measurable improvements in throughput, inventory handling, and customer conversion. NVIDIA’s emphasis on digital twins and accelerated inference suggests competition will focus on time-to-deploy and total cost of ownership, not just model quality. For partners, the next step is likely expanding these pilots into broader rollout programs across catalogs, fulfillment networks, and store channels, while tracking whether the promised cost and speed gains hold up as workloads grow.