Navigating uncertainty in Amazon's middle-mile network

Amazon's middle-mile network—the system of fulfillment centers, sort centers, and trucks that moves items closer to customers before last-mile delivery—faces constant uncertainty. While dramatic disruptions like snowstorms or power outages grab attention, Amazon's engineers emphasize that day-to-day variations in demand and travel times erode efficiency more subtly. To tackle this, they built optimization tools that solve a mixed-integer programming problem combining continuous decisions (e.g., shipment volumes) and binary choices (e.g., open/close shipping lanes). Even a simplified version with only 300 binary variables yields more combinations than atoms in the universe; Amazon's real network involves millions of such decisions. Their key insight: don't optimize for a single forecast. Instead, the tools use Monte Carlo methods to generate hundreds of plausible scenarios—varying demand, road conditions, processing times—and then stress-test each network design across them. Graph attention networks help identify which consolidation points and routing options provide the most flexibility under all scenarios. The result is a network design with built-in optionality, not a brittle plan optimized for one prediction.

This risk-aware approach promises measurable gains: even accounting only for demand variability, Amazon sees potential savings of 0.5% in network efficiency. While small in percentage, that translates to significant cost and carbon reductions across a massive operation. The tools also enable faster, more reliable delivery promises—customers can trust their packages will arrive on time even when a viral product spikes demand or a highway closes. By moving from deterministic optimization to probabilistic network design, Amazon's engineers are pushing logistics science forward, proving that building for uncertainty can be both computationally feasible and commercially essential.