NVIDIA Transforms European Retail, From Digital Discovery to Physical Operations

NVIDIA rounded up how European retailers and their tech partners are using open models, digital twins and simulation, from search and product images to warehouses and virtual try-on. The headline claim comes from Shopify: a post-trained NVIDIA Nemotron model powering a data-retrieval subagent inside its Sidekick merchant assistant beat a leading frontier model on Shopify's internal benchmark, while cutting serving costs by 90% and answering seven times faster.

If that pattern holds, the economics of AI in commerce shift away from renting the biggest model available and toward tuning smaller open ones for narrow jobs.

We think this is the most practically important trend in the whole GTC Berlin slate, even if it's buried in a retail roundup. Narrow tasks, like pulling store data through a query interface, don't need a frontier generalist. A smaller model trained on the actual job can be cheaper, faster and, by Shopify's account, more accurate. That's good news for merchants who pay for these features indirectly, and good news for anyone worried about the energy bill of AI.

But every number here is self-reported. Shopify's comparison uses its own internal benchmark. CATCHES reports conversion up 55% and gross sales up 56% for AMIRI after seasonal adjustment, and Photoroom says listings it processed for Vestiaire Collective sold at a 5% higher rate. These are promising, and they are also exactly the kind of figures that show up in a vendor's showcase post. I'd want to see them hold across more brands and more seasons.

There's a consumer angle we shouldn't skip. Generated product imagery is now standardizing tens of thousands of listings a day, and brands are scaling creative output across whole catalogs. Omi's approach of anchoring every generated image to an approved digital twin is the responsible version, because the product shown matches the product sold. Without that discipline, the same tools make it trivially easy to sell a prettier fiction.

And then there's the shopping agent. CATCHES launched CASI, a fashion agent that shops, styles and sizes for customers over WhatsApp. Agents acting on a buyer's behalf are useful, but they also move persuasion into a private chat where disclosure and comparison are harder. We'll be watching how those agents are labeled and whose interests they rank first.

On the operations side, Otto Group extending simulated warehouse design to its largest fulfillment center in Haldensleben, Germany, under human supervision, is the sober, unglamorous use of AI that tends to stick.

Post-training means taking an existing open model and further training it, through fine-tuning or reinforcement learning, on a specific task. Shopify used NVIDIA's Megatron Bridge and NeMo RL libraries for that work. Other participants named include Zooplus with Infosys on semantic product search, L'Oréal on computational ingredient screening, and Mirakl running open models on its own infrastructure and European clouds like Scaleway to avoid lock-in.