Simulation as a Standing Capability
In comments published July 15, Walmart's senior vice president of supply chain technology, Indira Uppuluri, described a network that no longer waits for disruption to react. The company runs digital twins of its supply chain to simulate responses to facility closures, transportation delays, and demand shifts. Instead of modeling these scenarios after they hit, Walmart rehearses them in software, so the physical network already has a tested playbook when a port stalls or a distribution center goes offline.
The distinction matters at Walmart's scale. When you move product across thousands of stores and a global sourcing base, the cost of a bad reaction to a shock compounds fast. Uppuluri's framing is that the twin does the reasoning ahead of time. As she put it, the systems "come up with actions that we can take." That phrasing moves the technology from dashboard to decision support, which is the threshold most supply chain analytics never crosses.
A Stack Built From Several Model Types
Walmart is not betting on a single model class. Uppuluri describes machine learning for predictive analytics, large language models, open-source models, agentic AI agents for holistic resource optimization, and custom tools built by internal data science teams. Each does a different job. Predictive models forecast, language models interpret and interface, and agents coordinate actions across the network rather than optimizing one node in isolation.
We read the open-source inclusion as a cost and control signal. A retailer running inference at Walmart's volume has strong reasons to avoid paying frontier-model prices for every routine decision, and to keep sensitive supply data inside its own environment. Mixing proprietary and open models, with custom tooling on top, is the pragmatic architecture for an operator that treats AI as core infrastructure rather than a purchased feature. It also reduces dependence on any one vendor's roadmap.
The Objective Function: Assortment, Speed, Cost
Uppuluri names the trade-off the technology exists to manage. "Assortment, speed, and cost are the three [factors] that we are trying to balance and optimize in through the supply chain," she said. That is a clean statement of the retail supply chain problem. Carry more assortment and you raise complexity and holding cost. Push speed and you raise transportation and fulfillment cost. Cut cost and you risk stockouts or slower delivery. The three pull against each other constantly.
The value of agentic optimization is holding those three variables in view at once, across many nodes, faster than a human team can. Uppuluri's teams use the models to manage supply chain nodes, optimize fulfillment engines, and guide transportation logistics. We would note that a stated objective function is itself a discipline. It forces the AI toward a defined goal and gives the organization a way to judge whether a recommendation actually improved the balance or just shifted cost from one line to another.
Skilling the Workforce That Operates the Agents
Technology this central needs people who can operate it. Walmart is certifying staff on AI through OpenAI and Google, delivered via its internal Squiggly platform. That is a meaningful detail. A supply chain organization that hands agentic tools to teams without building fluency ends up with expensive software nobody trusts. Formal certification signals that Walmart intends the workforce to interrogate and act on model output, not defer to it blindly.
For enterprise leaders, this is the part that is easy to underfund. The model spend gets board attention while the change-management and skilling budget gets trimmed. Walmart pairing named vendor certifications with an internal delivery platform suggests it learned that lesson. The agents recommend actions, and the humans still decide, which only works if those humans understand what the agent is optimizing and where its reasoning tends to fail.
Why Now: Tariffs and Geopolitics
The timing is not incidental. In 2026, supply chains face tariff pressures and geopolitical challenges that make predictive modeling more valuable to retailers. When trade costs move on policy and routes shift on conflict, the ability to simulate outcomes and reallocate quickly turns into direct margin protection. A digital twin that can price the impact of a new tariff line or a closed corridor before it lands is worth real money to a company sourcing at Walmart's scale.
This reframes the investment case for supply chain AI. The pitch is no longer efficiency in a stable world. It is resilience in an unstable one, where the network absorbs shocks that competitors take on the chin. We think that framing will resonate with any retailer whose 2026 planning assumptions were upended by trade policy, and it explains why Walmart is willing to run this much modeling infrastructure as a permanent capability.
The Benchmark Everyone Else Now Measures Against
Walmart's scale makes its choices a reference point for the industry. When the largest retailer treats digital twins and agentic agents as standing infrastructure rather than pilots, it resets expectations for what a modern supply chain looks like. Smaller operators cannot match the spend, but they can copy the shape: define the objective function, blend model types, instrument for action, and train the people who use it.
The honest caveat is that we are hearing this from Walmart's own executive, and the hard proof lives in inventory turns, fill rates, and cost per unit that the company does not fully disclose. What is verifiable is the architecture and the intent. Walmart has committed to running its supply chain as a continuously simulated, agent-optimized system balancing assortment, speed, and cost. For competitors, the question is no longer whether to build this capability, but how fast they can close the gap.
