Box and Nvidia say new AI engine cuts token use by 90 percent

At BoxWorks 2026 in San Francisco, Box and Nvidia announced Box Context Retrieval, an engine using Nvidia Nemotron Retriever embedding models. It is meant to help AI agents pull relevant material from enterprise content—including scanned documents, images, audio and video—while limiting access to what each user is allowed to see. Box says the system can improve search accuracy by 20 percent, cut token use by 90 percent and make question-answering four times faster. Nvidia Nemotron 3 Ultra models are already available in Box AI for Enterprise Advanced customers; the retrieval engine is slated for those customers early next year. The companies are linking Nvidia OpenShell's secure runtime with Box Mount for policy enforcement on agent workflows and both belong to the Open Secure AI Alliance. Box CEO Aaron Levie said enterprise AI reaches its full potential when models and agents can securely access a business's unique context, and that agents can get the most relevant document chunks instead of reading whole files. Nvidia CEO Jensen Huang said the pair are connecting agents to enterprise knowledge with fast, accurate retrieval, Nemotron open models and OpenShell. Kari Briski of Nvidia told TechCrunch she expects the fourfold speed gain, 20 percent fewer errors and 90 percent fewer tokens. No contested claims identified. Wider availability beyond Enterprise Advanced early next year is not detailed.