NVIDIA Cuts Supply Chain Planning Time by 144x With NVIDIA cuOpt and AI Agents

Over the past five years, NVIDIA’s product portfolio has evolved from discrete graphics processors to fully integrated, rack-scale AI factories, where a single customer deployment requires coordinating entire ecosystems: GPUs, switches, NICs, cables, and server platforms that must arrive and come together seamlessly. Rather than treating this growing complexity as a risk, NVIDIA saw an opportunity to build a more resilient, future-ready planning capability.

To stay ahead of shifting demand and potential disruptions, NVIDIA set out to modernize how it evaluates what-if scenarios across its end-to-end network. 

The goal: to move beyond manual, batch-style planning and adopt a more dynamic, high-frequency approach that could continuously assess alternatives as conditions change.

Events like pandemic-era factory shutdowns or shipping route changes underscored how critical it is to anticipate disruption, not just react to it. Teams recognized that sustaining growth at data center scale would require planning tools that could “see around corners”—enabling teams to rapidly simulate alternate scenarios and make confident decisions well before physical supply chain impacts are felt.

NVIDIA’s IT and Operations Software Engineering team launched AI Planner—a multi-year internal program to infuse NVIDIA’s own AI technologies into the full breadth of our supply chain planning. The goal was not incremental improvement but a fundamental shift: Compress critical planning cycles from hours to minutes, and empower supply chain planners to generate and explore optimization scenarios with minimal operations research expertise required.

At the heart of AI Planner is the NVIDIA cuOpt GPU-accelerated decision optimization solver. Benchmarked at up to 10x faster than leading CPU-based solvers on linear programming (LP) problems, cuOpt gave the team a path to solve the kinds of large-scale, complex optimization models that had previously been too slow for interactive planning.

The team built Max-Supply, a unified optimization model that represents NVIDIA’s global supply network as a single mixed-integer linear programming (MILP). Max-Supply models the entire wafer-to-server supply chain as a connected system, allowing planners to evaluate the downstream effects of any change—a factory constraint, a demand shift, a component shortage—across the full network simultaneously. NVIDIA cuOpt, running on accelerated on-premises infrastructure, solves these large models natively on GPU, making scenarios that previously required overnight runs possible in minutes.

Powerful solvers alone don’t help if planners can’t interact with them. To close the usability gap, the team layered a multi-agent reasoning system on top of cuOpt, built with state-of-the-art large language models and NVIDIA NIM™ microservices. The agent system serves as an intelligent interface between the planner and the optimization engine, handling the full loop:

• Understanding planner intent from natural-language questions
• Translating intent into data and model modifications
• Orchestrating optimization runs through cuOpt
• Interpreting results and explaining trade-offs back to planners in plain language

This design solves what the team calls the “usability paradox” of optimization tools: The models sophisticated enough to handle NVIDIA’s complexity were historically accessible only to specialists. With the Nemotron-powered agent layer, any supply chain planner can submit a natural-language query—'“What happens to our GPU availability if factory X goes down for two weeks?”—and receive an optimized scenario with an explanation, without writing a line of code or waiting for an OR analyst.

The AI planner program launched with three workstreams now in production: 

• Customer Allocation: Optimizing how supply is distributed across customers 
• AVC (Available versus Commit): Matching committed deliveries to available supply 
• Max-Supply: The full end-to-end network model

All three run on NVIDIA’s accelerated on-premises infrastructure and are integrated with enterprise systems, including SAP.

Speed and Scenario Agility

What-if scenarios that previously required a full day to run now complete in under 10 minutes—a reduction of more than 144x. This means that planning leaders can run dozens of scenarios within a single decision window rather than choosing one or two and waiting overnight for answers. 

Teams now stress-test their plans against demand upsides, supply constraints, route changes, and temporary factory pauses before those events materialize.

Productivity and Scale

Beyond speed, AI Planner delivered a reported 6x increase in planner productivity—meaning each planner can now evaluate and manage significantly more supply scenarios than before. This has enabled them to keep pace with growing business complexity and meet evolving business needs.

The platform has also united teams. Sales and operations teams that historically operated in separate planning workflows now collaborate around shared scenario models, improving alignment and reducing the time spent reconciling conflicting plans across groups.

Influence Beyond NVIDIA

NVIDIA’s internal deployment has become a reference architecture for the broader supply chain software ecosystem. The AI planner implementation is being shared with leading ISVs as a model for GPU-accelerated, agentic supply chain optimization. NVIDIA has also open sourced a cuOpt agent blueprint that enables developers to turn natural-language queries into optimization models, making this approach accessible to organizations beyond NVIDIA’s own IT organization.

“Optimization-driven systems have elevated decision speed and quality,” said Debora Shoquist, executive vice president of operations at NVIDIA. “Production planning scenarios that previously required several days now complete in minutes, enabling leaders to evaluate trade-offs and make informed decisions in real time.”

The AI Planner program is actively expanding. The team has several new workstreams in development that will extend AI-driven optimization across additional critical supply chain processes.

Longer term, the team’s vision is a supply chain that is fully proactive—one where AI agents continuously monitor signals across the demand-supply network, surface emerging risks, propose optimized responses, and allow planning leaders to act with confidence before disruptions cascade. The Franz Edelman recognition has also opened doors for broader collaboration with the operations research and AI communities, reinforcing NVIDIA’s position at the intersection of GPU acceleration and enterprise decision intelligence.