NVIDIA and Palantir Technologies announced a collaboration to deploy an AI stack inside NVIDIA's own supply chain operations. The focus is on materials allocation decisions that determine how quickly components move from silicon production to a finished data center rack.

The system puts NVIDIA Nemotron open models inside Palantir Foundry, grounded in the Palantir Ontology, a governed data layer that connects materials, manufacturing sites, commits, and supplier signals into one unified view.

Jensen Huang, founder and CEO of NVIDIA, described the scope in the joint announcement:

"Supply chains are the operating system of the physical economy, and AI factories are among the most complex systems ever built. From wafers and components to manufacturing, systems and customer delivery, hundreds of companies and trillions of dollars of global economic activity come together to deliver AI infrastructure. NVIDIA and Palantir are transforming this vast operational graph into sovereign intelligence, combining NVIDIA Nemotron models with Palantir's Ontology to reason, plan and orchestrate the journey from wafer to token."

What makes the allocation problem hard

Each Grace Blackwell NVL72 rack draws on millions of parts across thousands of suppliers. A single compute tray requires two Grace CPUs, four Blackwell GPUs, and thirty-two HBM3e stacks. NVIDIA's next platform is called Vera Rubin, named after American astronomer Vera Rubin, whose work on galaxy rotation curves provided the primary observational evidence for dark matter. The announcement noted its supply chain is twice as large as Grace Blackwell's.

Contract manufacturers cannot begin assembly until every required component is on site. NVIDIA tracks the interval from material receipt to finished product shipment as Time of Ownership. Before the current system, the weekly allocation problem was reworked manually.

NVIDIA cuOpt, an open-source GPU-accelerated optimization library, now solves this as a mixed-integer linear program. It reports which constraints are binding, so planners can see whether a given week's output was limited by GPU supply or memory availability rather than manufacturing capacity in a specific region.

Where the math ran short

When the team back-tested historical allocation decisions against real outcomes, the results revealed a gap. Planners had access to signals the solver could not see: emails with manufacturing partners, weather conditions in key regions, supplier debrief transcripts, and years of accumulated judgment about how much trust to extend to a given manufacturer's commitment.

NVIDIA and Palantir built a post-training workflow to capture that judgment. They selected Nemotron 3.5 Lightning, a 30-billion-parameter model with roughly 3 billion active parameters per forward pass. The model uses a mixture-of-experts architecture, where each input is routed through a subset of specialized sub-networks rather than the full model. That design keeps inference costs well below those of a dense model with an equivalent total parameter count.

Fine-tuning used LoRA, or Low-Rank Adaptation, a parameter-efficient method developed by Microsoft researchers and published in 2021. LoRA freezes base model weights and introduces small trainable matrices, which cuts the number of parameters that require updates during training by orders of magnitude. The full run finished on two NVIDIA B200 GPUs in minutes.

The gap between a general model and a specialized one

On a point-in-time backtest that replays each historical decision using only information available on that day, the post-trained Nemotron 3.5 Lightning reached 86.7% allocation-decision accuracy. The base version of the same model scored 17.5%. Nemotron 3 Ultra, a significantly larger general-purpose model, scored 55.5%.

Alex Karp, co-founder and CEO of Palantir, described the deployment in the joint announcement:

"NVIDIA has arguably the most valuable, intricate, and complex supply chain in the world. Our sovereign stack, powered by Nemotron models and Ontology, is delivering capabilities that exceed the frontier while providing alpha protection qualities unavailable otherwise."

The result points to something specific about post-training: on a bounded allocation task, a specialized 30-billion-parameter model outperformed a general-purpose model more than an order of magnitude larger.

Where the deployment runs

The stack operates on NVIDIA reference architectures and the jointly developed Palantir Sovereign AI Operating System Reference Architecture. On-premises configurations are supported by Dell Technologies and Cisco. Rackspace and Nebius cover cloud and co-location options.

Palantir plans to demonstrate the full AI stack at AIPCon 11, its flagship customer conference, where applications across manufacturing, energy, healthcare, automotive, and government sectors will be covered. The system is built to keep proprietary supply-chain data and model weights inside a single governed compute boundary, with inference also contained within that environment.

Disclaimer: All materials on this site are for informational purposes only. None of the material should be interpreted as investment advice. Please note that, despite the nature of much of the material created and hosted on this website, HODL FM operates as a media and informational platform, not a provider of financial advisory services. The opinions of authors and other contributors are their own and should not be taken as financial advice. If you require advice, HODL FM strongly recommends contacting a qualified industry professional.