Nvidia and Palantir are combining their stacks for AI-run supply chain management, and the first customer is Nvidia itself, across operations spanning millions of parts. Palantir is integrating Nvidia's open Nemotron models into Foundry, where companies can fine-tune them on their own operational data, while Nvidia's cuOpt handles scenario planning and optimization. The stated goal is to surface bottlenecks earlier, speed up allocation of materials, and compare options against each other. Decisions and their outcomes are fed back into the models to improve them over time.
Manufacturing, pharmaceutical and energy companies will be able to run the entire stack on their own hardware or in the cloud through the Sovereign AI OS reference architecture. Infrastructure partners named are Dell, Cisco, Rackspace and Nebius. Palantir says it will give further detail at AIPCon 11.
The Nemotron line is Nvidia's answer to Chinese open models. According to Artificial Analysis, the flagship Nemotron 3 Ultra was the strongest open model out of the US when it launched in June. It has since been passed by Inkling from Thinking Machines Lab.
That detail is worth sitting with, because it is the vendors' own scoreboard. A model that led its category in June and no longer does is being embedded as the reasoning layer inside enterprise supply chain systems that companies will fine-tune, deploy on their own metal and operate for years. This reads less like a bet on Nemotron being the best model and more like a bet that, for this class of work, model choice is not where the value sits — the operational data and the optimizer are.
Which is the more honest version of the pitch. Alex Karp, Palantir's chief executive, has been arguing that companies should run their own models rather than hand their institutional knowledge to outside vendors; Mistral founder Arthur Mensch recently made a similar case. It is a convenient argument for both parties here. Nvidia gives away open weights it does not monetize directly and in exchange keeps its silicon at the center of the deployment; Palantir gets a model layer it did not have to train. The customer gets independence from OpenAI and Anthropic, and a different dependency in its place.
Notably absent from the announcement is any customer other than Nvidia, any accuracy claim for the supply chain task itself, and any result from the deployment that is supposedly already running. The language is all forward-looking: the system "should" find bottlenecks sooner. Nvidia running it on its own millions-of-parts operation is the strongest thing in the release precisely because it is the only deployment with a name attached.
The feedback loop is the part that will be hardest to hold to account. A system that writes its own decisions and their outcomes back into a model fine-tuned on one customer's private data produces a supply chain whose logic exists nowhere in writing. That is a reasonable trade in exchange for finding a shortage three weeks early. It is a worse one the first time a regulator, an auditor or a board asks why the system allocated the way it did.