Shopify cut model serving costs by 90% and accelerated response times sevenfold after post-training open artificial intelligence models, NVIDIA said. The e-commerce provider used the NVIDIA Megatron Bridge library and NeMo RL reinforcement learning software to adapt Nemotron 3.5 Lightning for a GraphQL subagent inside its Sidekick merchant assistant. NVIDIA added that its CCluster vLLM software lowered inference costs by up to 12 times across five other Shopify applications, including catalog enrichment and product taxonomy classification.
Online pet retailer Zooplus rebuilt its product discovery system inside Amazon Web Services using Infosys services, the NVIDIA Nemotron Retriever library, and retail catalog blueprints. The resulting prototype achieved 86% search accuracy. Concurrently, marketplace photo provider Photoroom standardizes roughly 40,000 product images daily for Vestiaire Collective across 70 countries. Running on NVIDIA GPUs with TensorRT and Triton Inference Server, the pipeline produced a 5% increase in inventory liquidity and up to 7% more unique page views.
Catalog tools and chemistry
Software firm Mirakl onboarded over 70 million products through its Catalog Transformer by pairing fine-tuned open models with NVIDIA chips hosted on its infrastructure and cloud provider Scaleway. Creative workflow developer Omi introduced ProductDrop, an image system used in L’Oréal’s CreAItech initiative that matches generated visuals to pre-approved digital twins. Omi deployed its TRELLIS 3D asset generation on NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs, speeding inference tenfold and cutting latency by four times.
Cosmetics manufacturer L’Oréal also integrated NVIDIA ALCHEMI into research workflows to screen chemical formulations computationally in hours instead of months. The company is evaluating the setup alongside Orbital Industries’ CurieOS and the OrbMol-v2 molecular model to identify ingredients that solubilize ultraviolet filters. According to NVIDIA, OrbMol-v2 operates five times faster and produces 21% lower error than comparable models.
Warehouse simulation and sizing
Otto Group deployed an autonomous warehouse project under human supervision at its Haldensleben, Germany, fulfillment center, NVIDIA said. Built with systems integrator Reply using NVIDIA Omniverse libraries on Google Cloud, the system enables operators to design and evaluate facility logistics entirely in digital simulations before altering physical operations. Digital twin startup ALLSIDES similarly builds 3D models of retail stores and products for brands including adidas, Amazon, Nike, Zalando, and Zara using NVIDIA CUDA, OptiX, and TensorRT.
Apparel technology company CATCHES recorded an 80% rise in add-to-cart actions, a 55% jump in conversions, and a 56% increase in gross sales following an AMIRI deployment of its RealFit virtual try-on software. CATCHES also introduced PredictiveFit for retailers alongside CASI, a WhatsApp personal shopping agent powered by NVIDIA Nemotron 3 models and the Newton Physics Engine.
Cloud operator Nebius is demonstrating three early-access retail software blueprints running on NVIDIA hardware at GTC Berlin. The blueprints include an agentic store manager assistant, an inventory simulation tool for replenishment forecasting, and a fine-tuning pipeline that continuously feeds real-world data back into open models.
