NVIDIA and Palantir Combine cuOpt, Foundry, and Post-Trained Nemotron to Automate Supply Chain Allocations

NVIDIA and Palantir Combine cuOpt, Foundry, and Post-Trained Nemotron to Automate Supply Chain Allocations NVIDIA has partnered with Palantir to optimize its complex hardware supply chain—spanning from wafer fabrication to data center deployment—by integrating Palantir Foundry, NVIDIA cuOpt, and a post-trained Nemotron 3.5 Lightning model to codify human allocation decisions. NVIDIA platforms like Grace Blackwell NVL72 require thousands of parts and dynamic allocation across contract manufacturers to compress 'time-to-rack' and Time of Ownership (TOO). Quantitative allocation is solved as a mixed-integer linear program via NVIDIA cuOpt, pulling structured constraints and binding bottlenecks directly from Palantir Foundry's Ontology. Unstructured factors such as geopolitical updates, weather, and partner emails are captured alongside planner rationales in Foundry to train a specialized 30B-parameter Nemotron 3.5 Lightning model using LoRA. On development benchmarks, the post-trained Nemotron model achieved 86.7% allocation-decision accuracy, outperforming the larger Nemotron 3 Ultra (55.5%) and base Lightning model (17.5%). All model fine-tuning and inference run inside a sovereign compute boundary, completing LoRA training in minutes on two NVIDIA B200 GPUs. Managing Complex Silicon Supply Chains NVIDIA measures its supply chain efficiency from wafer-out to first token, breaking this down into 'time-to-rack' (fabrication to data center arrival) and 'time-to-token' (commissioning power, cooling, networking, and software). Assembling platforms like the Grace Blackwell NVL72 involves millions of components across thousands of global suppliers, with a single compute tray requiring two Grace CPUs, four Blackwell GPUs, and thirty-two HBM3e memory stacks. Because component availability shifts weekly, manufacturers cannot build until all parts arrive from NVIDIA direct stock, consigned stock, or external suppliers. To minimize Time of Ownership (TOO)—the duration material sits at a site before departing—planners rework critical material allocations across quarters and weeks. Bridging Quantitative Solvers and Foundry's Ontology The NVIDIA supply chain operations team built a Digital Supply Chain Intelligence command center on Palantir Foundry. Foundry's Ontology links materials, manufacturing sites, commits, capacity, and qualitative signals into an object-based operational layer. To calculate how much constrained material to assign to each site, NVIDIA cuOpt uses this Ontology to run GPU-accelerated mixed-integer linear programming that minimizes TOO while identifying binding constraints such as site throughput or component shortfalls. Rapid solve times let planners evaluate tradeoffs and run counterfactual scenarios, such as capacity fluctuations or memory reductions. Capturing Institutional Knowledge Beyond Math Back-testing showed that historical mathematical models could not anticipate qualitative inputs handled intuitively by experienced planners, such as partner emails, weather forecasts, debrief calls, and geopolitical events. To address this, Foundry was configured to capture decisions, expected results, context, and actual outcomes. This established an explicit operational dataset used to post-train an open-weight LLM, Nemotron 3.5 Lightning. Chosen for its 30-billion-parameter mixture-of-experts design (roughly 3 billion active parameters per forward pass), the model operates as a task-specific agent within a sovereign enterprise boundary. Post-Training Workflow and Benchmark Results The model lifecycle relies on NeMo Anonymizer for data sanitization, NeMo Data Designer for synthetic operational edge cases, and NeMo AutoModel for LoRA adapter fine-tuning on two NVIDIA B200 GPUs. Historical point-in-time replays gate each deployment managed via Palantir Autopilot. On allocation benchmarks, the post-trained Nemotron 3.5 Lightning model achieved 86.7% accuracy, 58.6% balanced accuracy, and a 57.5% macro-F1 score, significantly beating the larger Nemotron 3 Ultra (55.5% accuracy) and base Lightning (17.5% accuracy). While the specialized model improved allocation decisions, general production-risk forecasting remained challenging. Continuous Feedback and Closed-Loop Learning Deployed models output allocation recommendations alongside supporting rationales and risk indicators for planner review. Every planner acceptance, override, and eventual production outcome is recorded back into the Foundry Ontology. This continuous record provides training data for subsequent supervised runs and planned reinforcement learning using preference pairs. This closed-loop flywheel preserves institutional memory, simplifies planner workloads across product lines, and creates an exportable blueprint for other complex supply chain operations.

NVIDIA与Palantir合作将供应链专家经验模型化以优化物料分配 面对GB200等计算平台极度复杂的全球供应链，NVIDIA与Palantir合作构建数字供应链智能指挥中心。方案结合Palantir Foundry数据本体、cuOpt数学规划求解器与后训练的Nemotron 3.5 Lightning大模型，将人类专家的隐性决策知识编码为可在私有环境中闭环迭代的模型系统，有效压缩生产周期。 复杂硬件生产受动态变化的物料约束限制，cuOpt虽能通过混合整数线性规划优化所有权时间（TOO），但无法吸纳天气、地缘政治及合作伙伴沟通等非结构化现实信号。 Palantir Foundry构建了包含材料、产能及质化信号的统一本体层，用于记录历史决策、推理过程及实际交付结果，为模型训练提供上下文支撑。 采用300亿参数（每次前向传播约30亿活跃参数）的Nemotron 3.5 Lightning MoE模型，通过NeMo组件结合LoRA进行监督微调，在2张B200 GPU上数分钟内即可完成训练。 在受限分配基准测试中，后训练的模型决策准确率达到86.7%，显著优于大一个数量级的通用模型Nemotron 3 Ultra（55.5%）和基模（17.5%）。 该架构兼具主权AI特性，数据与权重完全留在企业计算边界内，但模型专业化主要提升了具体分配决策，并未彻底解决未来的生产风险预测难题。 定量运筹优化的优势与边界 NVIDIA的GB200 NVL72单机架计算托盘包含2颗Grace CPU、4颗Blackwell GPU及32个HBM3e堆栈，物料供应按周动态波动，组件齐套才能开工。为压缩从晶圆出厂到机架装配的时间并降低物料在工厂的留存时间（TOO），NVIDIA利用cuOpt库将分配建模为混合整数线性规划问题，不仅输出每周物料调配方案，还揭示产能等制约瓶颈。然而，回测表明纯数学优化无法覆盖计划员所掌握的天气预报、地缘政治事件、邮件沟通及长期行业直觉等定性信息。 利用Foundry本体捕获专家决策脉络 为弥合定量模型与现实决策的差距，NVIDIA团队在Palantir Foundry中构建了统一的数据本体层。该本体采用对象和关系链接取代传统行列表格，将物料分配、代工厂产能、客户承诺与非结构化运营信号相连接。系统记录计划员的每次分配决策、背后逻辑、预期结果及后续实际产出，使隐性机构经验转化为显式且可追溯的训练数据，奠定了以真实决策反馈训练AI的基础。 Nemotron 3.5 Lightning的轻量化后训练 团队选用了适合智能体工作流执行层的Nemotron 3.5 Lightning模型。该模型具备30B总参数与约3B激活参数的MoE架构，具备高效推理能力。利用NeMo Anonymizer脱敏数据、NeMo Data Designer扩充破坏性场景样本，并通过NeMo AutoModel执行LoRA微调，保持基础权重冻结。通过Palantir Autopilot管理全生命周期，微调在2张NVIDIA B200 GPU上数分钟即可跑完，支持在私有计算边界内高频更新。 决策准确率提升与局限性验证 在同一分配基准评估中，后训练后的Nemotron 3.5 Lightning决策准确率达86.7%，相比基模（17.5%）提升69.2个百分点，比参数量大一个数量级的Nemotron 3 Ultra（55.5%）高出31.2个百分点。在平衡准确率与宏观F1分数等平衡稀有决策的指标上，该模型亦分别达到58.6%和57.5%，均优于Ultra的42.0%和39.5%。不过，尽管微调提升了受限领域的分配判断质量，未来生产风险预测的准确性依然存在瓶颈，模型未能在所有预测任务上完全突破。 闭环飞轮与主权AI范式 部署后，模型提供包含风险与推理的分配建议，并由人类计划员最终把关。所有接受、修改、驳回操作及后续实际产出均回写至数据本体，累积用于后续基于强化学习的偏好对齐训练，模型绝不在生产中直接自我重新训练。这一端到端流程确保了企业敏感供应链数据与模型推断不外流，展示了受限业务场景下利用专业小模型构建闭环主权AI的实践路径。

NVIDIA has one of the largest and most complex supply chains in the world, and its performance is measured from wafer-out to first token. The interval is in two parts. Time-to-rack runs from silicon leaving the fab to an assembled system arriving on a data center floor. Time-to-token covers everything thereafter: power, cooling, networking, and the software stack that makes the infrastructure productive on day one. NVIDIA Grace Blackwell NVL72 platforms draw on millions of parts and thousands of suppliers spread across the globe, and the final system is built by dozens of OEMs and ODMs. Just one compute tray–one of eighteen in a single rack–requires two NVIDIA Grace CPUs, four NVIDIA Blackwell GPUs, and thirty-two HBM3e stacks. The supply chain we created for Vera Rubin is twice as large as Grace Blackwell. CPUs, GPUs, and memory are all critical components, and the availability of each changes from week to week, so the part holding up a build one week may be freely available the next. Each carries its own bill of materials, its own suppliers, and its own lead times. Multiply that by every sub-assembly in the rack, and the result begins to look less like a supply chain and more like a daunting combinatorics problem. Contract manufacturers cannot start assembly until every component has arrived from one of three pools: parts from NVIDIA directly, parts that NVIDIA stocks on consignment, and parts from suppliers. Ideally, components all arrive at once, but when they don’t, whatever arrived early waits for everything that’s late. NVIDIA runs the clock from the moment a manufacturing site receives material to the moment it leaves as part of a sub-assembly or product, and that metric is known as Time of Ownership (TOO). With highly dynamic availability, NVIDIA must decide what and how much material to allocate to each manufacturing site. This is known as the critical material allocation problem, and it is manually reworked every week. The allocation runs through the current quarter and the next with the nearest weeks already committed, so each week’s new data mostly changes what happens further out. This post is primarily concerned with time-to-rack, and compressing it requires four things: Real-time visibility to pinpoint critical operational bottlenecks at any moment Redundancy where a single failure may otherwise halt production Reliability, so upstream production commits hold Codified human expertise, so the judgment behind a complex allocation decision becomes persistent knowledge that compounds over time While the initial three requirements provide the necessary operational baseline, it is the codification of human expertise where the most significant transformation occurs. Figure 1. The GB200 NVL72 compute tray requires two Grace CPUs, four Blackwell GPUs, and thirty-two HBM3e stacks, all of which are critical components with dynamic availability Building a supply chain command center in Palantir Foundry The NVIDIA supply chain operations team worked with Palantir to create a unified view of every input to a material allocation decision. The NVIDIA team calls this their Digital Supply Chain Intelligence command center, and it surfaces risks, blockers, and other signals that inform those decisions but may previously have remained buried across disjoint data sources. Video 1. Sovereign AI for the compute supply chain Under the hood, Palantir Foundry provides the operating context. The Ontology connects materials, manufacturing sites, commits, capacity, allocations, production outputs, and unstructured, qualitative signals into one governed data layer. Composed of objects and links rather than rows and tables, it forms a complete representation of operational reality. This representation enables allocation planners to simulate and analyze many different scenarios, giving them much broader access to the decision space and laying the groundwork for an AI flywheel that compounds new knowledge and improves performance over time. Figure 2. Unified operating picture for NVIDIA supply chain in Palantir Foundry Solving the quantitative side with NVIDIA cuOpt Allocating material across multiple manufacturing sites starts as a quantitative problem, and the formulation starts with the decision variables: How much of each constrained material goes to which sites and when over the coming period?Around them sits everything that bounds the answer. This includes every manufacturer capable of building a given Blackwell sub-assembly and the throughput each site can absorb once material lands. It also includes the dependency graph of every required part mapped backward through the chain, so the solver knows that a compute tray is blocked by its scarcest input rather than its average one. That binding constraint is not fixed. It shifts between: GPU, CPU, and memory from one week to the next Inbound timing across all three supply routes Commitments already made to customers, which determine what a shortfall at any one site actually costs Thousands of variables and constraints resolve into a single weekly allocation. NVIDIA cuOpt, an open-source library for GPU-accelerated decision optimization, solves this. It draws its inputs from the Ontology and writes the result back as an allocation decision. Allocation is posed as a mixed-integer linear program, and the objective minimizes Time of Ownership (TOO). cuOpt returns more than the allocation itself. It also reports which constraints are binding, so a planner can see that Taiwan capacity, not memory supply, is what held this week’s number down. Because the solve is fast, planners can also explore the space around the answer. What happens with ten percent less memory this period? What if a new manufacturing site comes online? Planners stop asking the solver for an answer and start asking it what the tradeoffs are. Where the math stops Quantitative optimization doesn’t tell the whole story. NVIDIA and Palantir back-tested historical allocation decisions against what actually happened, and it revealed a human factor that cuOpt wasn’t capturing. Planners were working from information the solver could not see: emails exchanged with partners that week, severe weather in the forecast for a key region or an ongoing geopolitical event, the transcript from the last supplier debrief call, and years of accumulated expertise. Those inputs feed an instinct about how to allocate material for the coming period, and that instinct is what makes the human experts better than the math. With that in mind, NVIDIA and Palantir built this workflow around those human experts. It captures the allocation decision, the rationale behind it, the expected result, and the actual outcome. Institutional knowledge becomes explicit, reviewable decision logic, and because that data lives in the Ontology, it serves as the foundation for training an LLM on expert judgment. Codifying decision intelligence We then post-trained an open-weight LLM to apply that same reasoning and make a recommendation. After evaluating NVIDIA Nemotron open models, we chose Nemotron 3.5 Lightning because it is purpose-built for the execution layer of an agentic workflow. It performs specialized tasks as part of a system of models that also includes larger variants for orchestration and general tasks. Its mixture-of-experts architecture lends itself to highly efficient inference, and while the model is lightweight at 30 billion parameters and roughly 3 billion active per forward pass, it’s still large enough to learn a focused policy. This footprint makes the post-training loop practical because smaller models learn faster and need far less compute both to train and to deploy than their larger counterparts. Because Nemotron is open, it can be post-trained inside your own compute boundary. Any organization can run the same flywheel on their own operational data without exposing it externally. The model learns from the signals planners actually use: how much constrained material was offered to a manufacturing site, what the manufacturer committed to produce, what it ultimately produced, and the qualitative operational evidence available at the time of the decision. The objective is to codify an allocation policy that can assess production risk, recommend an allocation range, identify the factors behind that recommendation, and explain its reasoning to the supply-chain team. The same record doubles as the evaluation harness. We replay each decision with only what was knowable that day, hide the outcome, and compare the model’s recommendation against both the planner’s call and what actually happened. The primary question that this evaluation answers is: If this model had been running last month, would it have made the right allocation call? From Ontology data to a specialized model The training process starts with operational history in the Palantir Ontology: Anonymization: NeMo Anonymizer removes personally identifiable information and obfuscates sensitive fields before training. Synthetic data generation: NeMo Data Designer expands and balances the examples, so the model sees not only routine weeks but also allocation increases, capacity constraints, and disruption scenarios. Supervised fine-tuning: NeMo AutoModel trains a small set of LoRA adapter parameters while the base weights stay frozen, which cuts training time, memory requirements, and checkpoint size. Evaluation: The point-in-time backtest replays the same historical decisions through both the base and fine-tuned models, isolating what post-training added. Palantir Autopilot manages the lifecycle end to end, launching each job from Ontology data, monitoring the deployed custom Nemotron model, and keeping lineage intact from data to model version to recommendation. Once deployed, the model reads the current operational context and returns a recommendation with its rationale and the risks attached. A planner reviews it and makes the final call. Figure 3. Allocation decisions and model training share one governed data layer. Planner decisions become training data, and a point-in-time backtest gates every deployment Closing the loop Every acceptance, edit, override, and production outcome is written back into Ontology, accumulating until there is enough representative data to justify another governed training run. In the future that feedback will be used for reinforcement learning. Accepted and overridden recommendations would form preference pairs with rewards covering allocation correctness, policy compliance, and evidence grounding. The model never retrains itself in production. The result compounds in two ways: Planners spend less time reconstructing routine decisions, so they cover more sites and products; and allocation expertise also becomes institutional knowledge, shortening onboarding and propagating key learnings across the organization. What post-training delivered The NVIDIA supply chain operations team worked with Palantir to specify what the model receives, what it may recommend, and how those recommendations are scored. That workflow became both the application and the allocation decision-intelligence benchmark. We compared three models on the same task and the same evaluation data: Base Nemotron 3.5 Lightning (BF16) Nemotron 3 Ultra (NVFP4) Our post-trained Nemotron 3.5 Lightning (BF16) On the development benchmark, the post-trained Lightning model reached 86.7% allocation-decision accuracy . Ultra reached 55.5% and base Lightning 17.5%. That puts the post-trained model 31.2 percentage points ahead of Ultra and 69.2 ahead of its own base model. Figure 4. Accuracy scores of post-trained Nemotron Lightning compared against Nemotron Ultra It also leads on the two metrics that weight decision types equally rather than examples. Balanced accuracy averages per-class recall, so rare calls count as much as common ones: 58.6% against Ultra’s 42.0%. Macro-F1 averages per-class F1, folding in precision so a model cannot inflate recall by over-predicting a rare class: 57.5% against 39.5%. Both matter because constrained supply means planners cut allocations far more often than they raise them, so plain accuracy alone would flatter a majority-class guesser. The lesson is specific but important: On a bounded allocation task, a specialized 30B model can outperform a general-purpose model that is more than an order of magnitude larger. This doesn’t mean the smaller model is more capable overall. Its gains are concentrated in the domain it was post-trained on. Future production risk forecasting remained difficult despite fine tuning. Specialization improved the decision task but failed to solve every prediction problem attached to it. The LoRA run finished on 2x NVIDIA B200 GPUs in minutes, light enough to repeat as feedback accumulates. This is sovereign AI in practice: proprietary supply-chain data, model weights, and inference all remain inside a single governed environment. This AI stack can be deployed on -premises or in the cloud enabling organizations to run AI where their data, systems, and operational requirements demand it. A supply chain that learns The supply chain workflow discussed in this post isn’t unique to semiconductors. Any operation where critical capacity is allocated by experienced people working from fragmented signals can run the same flywheel and tailor it to new domains. This requires three essential things: A governed operational layer Decision capture that includes both rationale and outcome An open model that can be post-trained inside your own secure compute boundary Operational data trains the model, the model improves the decision, and the decision becomes new operational data for the next governed training round. That’s how NVIDIA and Palantir compress the time from wafer-out to first token, and that’s how the world’s most reliable AI infrastructure supply chain learns faster than it grows. Get started on customizing efficient, open-weight models for your own agentic workflows. Stay up to date on NVIDIA Nemotron by subscribing to NVIDIA news and following NVIDIA AI on LinkedIn , X , YouTube , and the Nemotron channel on Discord . Access open Nemotron Models on Hugging Face and a collection of NIM microservices and Developer Examples on build.nvidia.com .