Navigating uncertainty in Amazon's middle-mile network

Amazon's middle mile — the maze of fulfillment centers and sort hubs that gets your package close enough for a same-day driver to grab it — is a logistics puzzle with millions of moving parts. The company just published details on how it's tackling that puzzle, and the headline isn't the dramatic stuff like blizzards shutting down highways or a sort center losing power. It's the unglamorous, everyday variance in demand and travel times that quietly bleeds efficiency out of the system, one shipment at a time.

The scale here is genuinely absurd. Amazon's engineers note that with just 300 yes-or-no routing decisions, you already have more possible combinations than atoms in the observable universe. Now multiply that by millions of real decisions across hundreds of facilities, each running on tight operating windows, and you understand why even solving this problem under perfect, no-surprises conditions takes hours of computing time. Add uncertainty, and the standard academic playbook — model a bunch of scenarios, then build a robust solution across all of them — simply doesn't scale to Amazon's size.

So Amazon's approach isn't about bulletproofing the network against every conceivable disaster. It's about optionality: building in enough alternate routes and slack that day-to-day chaos gets absorbed rather than triggering a crisis. To make the underlying math tractable, Amazon's network-design tool narrows the search by focusing on sensible consolidation points — sort centers where shipments from different origins can share a truck — and it handles scheduling with a clever trick: plan at a coarse 24-hour resolution but attach precise 15-minute timing windows to each candidate route, so trucks don't get scheduled to leave before their cargo shows up.

The more interesting piece is how Amazon models uncertainty itself. Instead of assuming demand is simply proportional to a facility's size and proximity, the company built a graph attention network that represents the middle mile as two layered graphs — one mapping physical sites and routes, another mapping origin-destination pairs directly. That second layer catches something the first one misses: two nearby fulfillment centers serving the same city don't each get the full demand, they split it. Miss that nuance and you'll systematically overestimate how much a facility actually needs to ship.

Run enough Monte Carlo scenarios through that model — including ones the network has never actually experienced, like a brand-new facility opening nearby — and Amazon can stress-test its network designs before committing to them. The payoff, per Amazon's own numbers, sounds tiny: accounting for demand variability alone yields roughly 0.5% in savings. But at Amazon's volume, half a percent is a lot of packages arriving on time instead of late, which is presumably the entire point of publishing this in the first place.