Blue Yonder and NVIDIA build AI model factory for supply chains

Blue Yonder has developed a Model Training Factory with NVIDIA to accelerate specialised AI agents for autonomous supply chain operations.

The system is built on NVIDIA Nemotron and uses NVIDIA NeMo tools to fine-tune, test, and orchestrate supply chain models. Blue Yonder is combining that AI stack with its supply chain decisioning expertise, operational data structures, and workflow knowledge to create agents designed for high-value supply chain tasks.

The agents are intended to support decisions across warehouse management, supply and demand planning, transport, merchandising, and network operations. The first models are expected to focus on warehouse management workflows, including allocation shorts, inventory exceptions, due-time urgency, and inventory across yard and receiving trailers.

Blue Yonder is using a hybrid model strategy. Frontier models will be used where broad reasoning is required, while smaller, domain-specific models will be trained for precision, speed, and cost control inside defined workflows. Models are trained on synthetic data rather than customer data, with evaluation criteria applied before deployment and during improvement cycles.

Large general-purpose models can be powerful, but supply chain operations need systems that can run repeatedly, cheaply, and within strict process limits. Warehouse and transport systems need answers under time pressure, often many times a day, with each decision affecting labour, inventory, vehicles, and customer commitments. A model that is impressive in a demonstration but too costly or unpredictable for continuous production use will not survive inside a live distribution network.

Warehouse management gives the model factory a practical starting point because the environment is rich with repeatable but time-sensitive decisions. Late inbound vehicles, missing inventory, labour imbalance, short allocations, urgent orders, and dispatch cut-offs all force supervisors to replan quickly. Human teams usually make those calls with partial visibility and limited time. A trained agent can evaluate more trade-offs in seconds, provided it operates inside clear operational guardrails.

The announcement also reflects a shift from AI as an assistant to AI as an execution layer. Earlier supply chain AI often focused on forecasting, visibility, anomaly detection, or conversational search. The current wave is reaching deeper into live workflows, where software can identify a problem, assess options, recommend or execute an action, and learn from the outcome.

That direction is already visible across the market. Infios has embedded AI agents into execution workflows, while Oracle has pushed supply chain software into agentic execution. Blue Yonder’s model factory adds another dimension by focusing on repeatable development of specialised agents rather than isolated AI features.

The phrase “model factory” is important because supply chains do not need one generic agent. They need many narrow agents that understand different tasks, constraints, and exception patterns. A warehouse allocation-short decision is not the same as a transport tendering decision or a merchandising allocation decision. Each requires different data, latency, approval paths, and risk tolerance.

Deployment discipline will determine whether the approach scales. Autonomous supply chain operations need models that are explainable enough to trust, cheap enough to run, accurate enough to reduce rework, and governed tightly enough to avoid uncontrolled decisions. In warehouses especially, a bad decision does not stay on a screen; it can move labour, inventory, equipment, and vehicles in the wrong direction.

Blue Yonder expects the first models to move into customer production through its Cognitive Solutions later this year. If the system performs in live operations, the model factory could give supply chain AI a more industrial structure: repeatable, domain-trained, governed, and tied directly to operational outcomes.