Shopify Cuts AI Costs 90% Using Open Models on NVIDIA Hardware

European retailers are adopting open artificial intelligence models and physical intelligence tools to significantly reduce computing costs and improve operational speed, according to a blog post published by NVIDIA. This shift allows commerce platforms to achieve frontier-level accuracy without the high price tags typically associated with proprietary large language models.

How it unfolded

• Shopify engineers used the NVIDIA Megatron Bridge library and NeMo RL to post-train the Nemotron 3.5 Lightning model for its Sidekick assistant.
• The resulting subagent retrieves store data for features like Sidekick Pulse, surpassing leading frontier models in accuracy on internal benchmarks.
• Shopify reported a 90% reduction in serving costs and seven-times faster response times compared to previous configurations.
• Additional use cases, including catalog enrichment and merchant recommendations, saw up to 12x lower inference costs using CCluster vLLM software.

Who is affected

Zooplus, a major European online pet retailer, rebuilt its product discovery stack with Infosys and NVIDIA technologies. The prototype, running on Amazon Web Services, used the NVIDIA Nemotron Retriever library for semantic search. This deployment achieved an 86% search accuracy rate by generating search-ready product attributes through the NVIDIA Blueprint for retail catalog enrichment.

Photoroom partnered with Vestiaire Collective to standardize approximately 40,000 primary images daily across 70 countries at the point of listing. The pipeline runs on NVIDIA GPUs using TensorRT and Triton Inference Server. Listings processed through this system sold at a 5% higher liquidity rate and received between 3% and 7% more unique views per product page.

Omi launched ProductDrop, a workflow that anchors generated images to approved digital twins to prevent product errors across entire catalogs. L’Oréal uses this approach within its CreAItech program. The system runs TRELLIS 3D asset generation on NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs, speeding up inference by 10x and reducing pipeline latency by 4x.

L’Oréal is also accelerating ingredient discovery using NVIDIA ALCHEMI, moving from months of experimentation to hours of computational screening. Additionally, researchers are testing CurieOS with Orbital Industries’ OrbMol-v2 model. Integrated with the NVIDIA ALCHEMI Toolkit, this open-source molecular simulation model is five times faster than comparable models and delivers 21% lower error rates.

Mirakl, a software company for intelligent commerce, runs open models on NVIDIA GPUs across its own infrastructure and European cloud providers like Scaleway. This strategy maintains technology sovereignty, avoids vendor lock-in, and delivers the speed and scale required by its customers. As Circuit Observer reported on October 21, 2026, NVIDIA committed $10 billion to European AI infrastructure, highlighting the foundational layers required for such industrial-scale computing partnerships.

What happens next

These deployments demonstrate that open models can compete with frontier proprietary systems on both cost and performance metrics when accelerated by specialized hardware. Retailers are increasingly relying on digital twins and semantic search to enhance product discovery and operational efficiency. The integration of reinforcement learning and semantic retrieval libraries allows companies to fine-tune models for specific commerce tasks, such as catalog enrichment and merchant assistance, without incurring the high costs of standard large language model inference.

The background in plain language

Open models are AI systems whose code and weights are publicly available, allowing companies to modify them for specific tasks rather than relying on fixed, black-box services. Digital twins are virtual replicas of physical processes or products, used to simulate outcomes before implementation. Inference refers to the computational process of running an AI model to generate answers or predictions. Teams usually evaluate open models to balance performance needs against licensing and hosting costs, checking whether customized training yields better accuracy for their specific data sets than general-purpose alternatives.

Background: Large language models

Large language models act as probabilistic optimizers that predict the next most likely token rather than verifying logical correctness. This mechanism means they excel at pattern matching but fail at rigorous reasoning. You must treat their output as a draft requiring strict validation, not a final solution.

Read the full guide: Large Language Models as Optimizers: Why Your Code Changes Break

What to do and how to stay safe: NVIDIA

• Audit current AI inference costs to identify opportunities for switching from proprietary models to open-source alternatives that can be fine-tuned for specific tasks.
• Evaluate whether digital twin simulations can reduce physical testing cycles for product development or operational workflows before full deployment.
• Review cloud infrastructure strategies to ensure technology sovereignty and avoid vendor lock-in by leveraging open models across multiple providers.
• Monitor performance benchmarks of post-trained open models against leading frontier models to determine if accuracy meets business requirements at lower costs.

Step-by-step guide: Mobile Device Security for Small Business: Control Without Cost

General security guidance from the Circuit Observer newsroom. It is not confirmed advice from the organisations named in this story.

Frequently asked questions

How much did Shopify reduce its AI serving costs?

Shopify cut serving costs by 90% while delivering responses seven times faster on its internal benchmarks.

What accuracy rate did Zooplus achieve with its new search stack?

Zooplus achieved an 86% search accuracy rate using the NVIDIA Nemotron Retriever library in its AWS environment.

Which GPU hardware did Omi use to speed up inference?

Omi used NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs to speed up inference by 10x.