NVIDIA Puts Weekly Chip Allocation on Palantir
NVIDIA has detailed how it replaced a manual weekly decision about which scarce CPUs, GPUs and memory go to which contract manufacturers with a Palantir Foundry data layer and a GPU-accelerated solver.
In a technical blog post, NVIDIA describes an internal workflow it calls Digital Supply Chain Intelligence, built with Palantir to compress "time-to-rack" — the interval from silicon leaving the fab to an assembled system landing on a data center floor.
The scale figures are the most concrete part. A Grace Blackwell NVL72 platform draws on millions of parts and thousands of suppliers, with final assembly spread across dozens of OEMs and ODMs. A single compute tray, one of 18 in a rack, requires two Grace CPUs, four Blackwell GPUs and 32 HBM3e stacks. NVIDIA says the supply chain it built for Vera Rubin is twice the size of Grace Blackwell's.
The recurring bottleneck is what NVIDIA calls the critical material allocation problem: deciding how much constrained material goes to which manufacturing site each week, a task previously reworked by hand. Palantir Foundry's Ontology links materials, sites, commits, capacity and allocations into one governed layer; NVIDIA cuOpt, its open-source optimization library, poses allocation as a mixed-integer linear program minimizing Time of Ownership and reports which constraints are binding.
Back-testing historical decisions against outcomes showed planners were using inputs the solver could not see — supplier emails, weather, geopolitics — which motivates the Nemotron post-training work the post's title promises.
The headline accuracy claim, 86.7% versus 55.5%, comes from a self-defined internal benchmark that is not reproducible externally.