A smaller artificial intelligence model trained on company-specific data outperformed a much larger general-purpose model on a specialized business task, according to a benchmark released by Nvidia and Palantir. The two companies announced a joint AI stack that pairs Nvidia's customizable Nemotron open models with Palantir's Foundry platform and Artificial Intelligence Platform. The approach is being rolled out first within Nvidia's own supply-chain operations before expanding to customers in manufacturing, retail and technology sectors.

Nvidia's supply-chain team tested two Nemotron models on a materials-allocation task: the smaller Lightning model and the larger Ultra model. In their baseline form, Lightning achieved 17.5% accuracy on allocation decisions, while Ultra reached 55.5%. After post-training Lightning with company-specific operational data, its accuracy jumped to 86.7%. Nemotron 3.5 Lightning is a smaller mixture-of-experts model built for high-volume, low-latency tasks, while Nemotron 3 Ultra is designed for more complex planning and reasoning. The post-training process allowed the smaller model to surpass Ultra on this narrowly defined business task, though the results were not independently verified and don't establish whether other companies or workloads will see similar gains.

According to Nvidia CEO Jensen Huang, "Supply chains are the operating system of the physical economy, and AI factories are among the most complex systems ever built." The company-run benchmark suggests that adapting a smaller open model to a specific operational task can produce better results than relying on a larger general-purpose model. The announcement states that because Nemotron models are customizable open models, organizations can adapt and deploy them while retaining greater control over their models, proprietary data and deployment environments—an approach Nvidia and Palantir characterize as sovereign AI.

The joint stack works by using Palantir's Ontology to connect data, business objects, operational relationships and permitted actions so the AI system can reason within a company's specific operating environment. This organizational context is what allows smaller models to be specialized for particular tasks rather than relying on the largest available model for every enterprise use case. The jointly developed Palantir Sovereign AI Operating System Reference Architecture supports deployments on premises, in colocation facilities or in the cloud, with infrastructure partners including Cisco and Dell for on-premises setups and Rackspace and Nebius for colocation and cloud environments. Nvidia and Cisco are also expanding AI factory deployments with Equinix, reflecting the broader channel ecosystem emerging around enterprise AI infrastructure.

For channel partners, the opportunity extends beyond providing another AI model. Enterprises adopting the stack may need help preparing proprietary data, building Palantir Ontologies, post-training Nemotron models and deploying the supporting infrastructure across manufacturing, retail and technology industries. The benchmark illustrates why smaller, specialized models could be more practical than using the largest available model for every enterprise task, though the vendor-reported results apply specifically to the materials-allocation test conducted within Nvidia's supply chain. The shift toward customizable, company-specific AI models could reshape how enterprises balance model size against operational precision, particularly when decision-making accuracy on narrow tasks outweighs the breadth of general-purpose capabilities. Partners who can bridge the gap between off-the-shelf AI platforms and the messy reality of proprietary business logic may find themselves in higher demand than those simply reselling compute power.