The largest general-purpose AI model is not necessarily the best model for a company’s most specialized decisions.
Palantir and Nvidia have announced an AI stack that combines customizable Nvidia Nemotron open models with Palantir Foundry and its Artificial Intelligence Platform. Palantir’s Ontology grounds the models in an organization’s data, operational relationships and decision-making rules.
The stack is being deployed first within Nvidia’s supply-chain operations, but the companies plan to extend the approach to Palantir customers in industries including manufacturing, retail and technology. For channel partners, that could create opportunities involving model customization, data integration and AI infrastructure deployment.
“Supply chains are the operating system of the physical economy, and AI factories are among the most complex systems ever built,” said Jensen Huang, the CEO of Nvidia.
Post-training pushes smaller model ahead
In a technical post accompanying the announcement, Nvidia said its supply-chain team worked with Palantir to evaluate Nemotron 3.5 Lightning and Nemotron 3 Ultra on a materials-allocation task. The test compared the general-purpose models with a version of Lightning post-trained using company-specific operational data.
The general-purpose Lightning model achieved 17.5% allocation-decision accuracy, while the larger Ultra model reached 55.5%. After post-training, Lightning’s accuracy rose to 86.7%, according to Nvidia.
For context, Nemotron 3.5 Lightning is a smaller mixture-of-experts model designed for high-volume, low-latency execution, while Nemotron 3 Ultra is intended for more complex planning and reasoning. In this test, post-training allowed the smaller model to outperform Ultra on one narrowly defined business task.
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.
However, the results were not independently verified. Although they demonstrate a substantial improvement on the tested allocation task, they do not establish that other companies, workloads or production environments will see the same gains.
What this means for enterprises
Enterprises could post-train smaller models to improve their performance on narrowly defined, company-specific tasks.
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. Nvidia and Palantir characterize this approach as sovereign AI.
Palantir Foundry and AIP also provide the organizational context needed to specialize the models. Palantir’s Ontology connects data, business objects, operational relationships and permitted actions so the AI system can reason within a company’s specific operating environment.
The jointly developed Palantir Sovereign AI Operating System Reference Architecture supports deployments on premises, in colocation facilities or in the cloud. The announcement names Cisco and Dell for on-premises infrastructure and Rackspace and Nebius for colocation and cloud deployments. 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. The benchmark is vendor-reported, but it illustrates why smaller, specialized models could be more practical than using the largest available model for every enterprise task.