How Amazon Delivery Actually Works: The Logistics Machine, Explained
The box on your porch is the last step of a machine built from eight regional networks, robotic warehouses, a private trucking layer, and thousands of small delivery companies wearing Amazon's colors. Here's the whole run, from stow to doorstep.
Evergreen explainer. Figures are checked against the primary sources listed at the end. Corrections policy.
Somewhere between the tap of a "Buy Now" button and the thump on your porch, a package moves through one of the largest privately built logistics systems ever assembled — and almost none of it looks like what people picture. There is no warehouse worker strolling long aisles with your printed order in hand, hunting shelf by shelf. The brown-van driver probably doesn't work for the company on the box. And the most important decisions about your delivery were made days or weeks before you ordered, when an algorithm decided which buildings should hold that phone case in the first place.
This is a systems walkthrough of the machine — what the industry calls first mile, middle mile, and last mile — using the operator's own published numbers plus federal retail data. No verdicts here on the live controversies around the company (labor practices, market power, and municipal tax deals are all real debates); today's assignment is the conveyor belt itself.
Why the machine exists: the prize keeps growing
The economic backdrop first. E-commerce accounted for 17.1% of U.S. retail sales in the second quarter of 2026, per the Census Bureau's quarterly report — about $340 billion in a single quarter, still growing at double-digit annual rates long after the pandemic surge faded. Every percentage point of that shift converts store trips into parcels, which is why the U.S. now moves tens of millions of packages a day across all carriers, and why the retailer with the most parcels decided to become its own carrier. Amazon's delivery arm now handles more U.S. parcel volume than the legacy carriers it once relied on — a vertical-integration story two decades in the making.
First mile: inventory placement is the real product
The journey begins before you order. When a manufacturer or third-party seller ships stock into the network, algorithms decide how to split it among fulfillment centers — the vast robotic buildings where goods live. Two design choices here do most of the work people attribute to "fast shipping."
Random stow. Items aren't shelved by category; they're stowed wherever space exists, with every placement barcode-tracked. A given phone case might sit in forty locations across the country. Chaotic to human eyes, the scheme maximizes storage density and means a picker (or a robot carrying a shelf pod to a picker) is always close to something on any order.
Regionalization. The bigger idea arrived in 2023, when the company reorganized its U.S. network into eight largely self-sufficient regions, each stocking the full popular catalog. Before the change, by Amazon's own account, 62% of orders were fulfilled entirely from within what would become the customer's region; after it, 76% and climbing. The point is subtraction: an order that never leaves its region skips cross-country linehauls, extra handoffs, and days in transit. Speed was bought with geography, not faster trucks.
The regionalization dividend
Share of U.S. orders fulfilled entirely within the customer's region — %
Source: Amazon Science, fulfillment network regionalization (2023)
When you finally click "buy," the system runs a rapid auction among buildings holding your item: which fulfillment center can hit the promised date at the lowest total cost, considering inventory, outbound truck schedules, and everything else in your cart? The winner gets the order; a picker and a fleet of floor robots assemble it; a machine builds a right-sized box; and your package — now a barcode with an itinerary — heads to the dock.
Middle mile: a private freight railroad made of trucks
Between the fulfillment center and your neighborhood sits the layer nobody sees: sortation centers and a scheduled trucking network running between them, overwhelmingly on the interstates — the same 90/10-funded pavement whose origin story we've told, now functioning as the fixed rail of American e-commerce. Packages from several fulfillment centers converge on a sortation center, get sorted by destination zip clusters, and roll out again toward delivery stations — smaller depots placed inside metro areas.
Regionalization simplified this layer enormously: with most volume staying inside one region, the number of long trucking lanes to manage dropped sharply, per the company's engineers. Fewer lanes means denser, more predictable truckloads — and trucking, like broadband and every other fixed-cost network, rewards density above all.
Alongside the linehaul trucks sits an air cargo fleet for the long, thin routes, and — for the fastest promises — a newer class of smaller same-day facilities that combine picking and delivery dispatch under one roof for a curated set of high-velocity items in big metros.
Last mile: the vans that aren't Amazon's
Now the part you actually see, and the part most people get wrong. The blue-gray van with the smile logo is almost certainly operated by a Delivery Service Partner — an independent small business that contracts exclusively with Amazon, leases branded vans, and employs its own drivers on routes dispatched from a delivery station. Per the company's own program figures, about 4,400 DSP businesses employ some 390,000 drivers across 19 countries, together delivering more than 20 million packages a day; the program has generated $58 billion in revenue for those firms since 2018.
Why structure it this way? Mechanically, the DSP model converts a monolithic labor and fleet problem into hundreds of franchised ones: each DSP handles hiring, scheduling, and vehicle upkeep for 20–40 routes, while Amazon controls the algorithmic core — route assignments, package loads, navigation, and performance metrics flow from Amazon's systems to DSP drivers' handhelds. Critics argue this keeps control while outsourcing employment obligations; the company frames it as small-business creation. Both descriptions point at the same architecture. A gig-style layer, Amazon Flex, adds ordinary people driving their own cars for overflow and same-day surges — a shock absorber around the scheduled system.
The last mile is also where the densest engineering lives, because it's the most expensive mile: an industry rule of thumb puts it at roughly half the total cost of delivery. Route algorithms sequence 150-plus stops against traffic and delivery windows; station software sequences the loading of the van so the next package is always near the door; photo-on-delivery closes the loop. Every optimization compounds across millions of daily stops.
One package's run through the machine
The standard journey from seller to doorstep
Diagram: The Explainer Desk. Source: Amazon's published network descriptions. Amazon Science, 2023
The machine in reverse: returns
The conveyor also runs backward, and the reverse direction is deliberately built not to mirror the forward one. A returned item can't just slot back on a shelf — it must be inspected, regraded, repackaged or routed to liquidation, and none of that is worth much for a $12 phone case. So the system triages: drop-off partnerships (no box, no label, a barcode at a kiosk) collect returns cheaply in bulk; high-value items flow back through fulfillment centers; low-value items often go to liquidation marketplaces in bulk lots — and sometimes, when reverse logistics cost more than the goods, customers are told to keep the item and refunded anyway, a calculation that startles people until they price a single package's round trip. Returns are the tax e-commerce pays for replacing the fitting room, and minimizing that tax is its own quiet discipline of the machine.
And every November, the whole apparatus stress-tests itself: peak season roughly doubles daily volume within weeks, absorbed by temporary hires in the hundreds of thousands, pre-positioned holiday inventory, and the Flex layer swelling on demand — capacity elasticity as a designed feature, not an emergency.
The economics under the conveyor belt
Why build all this instead of just paying the parcel carriers? Three mechanical reasons, each visible in the design.
Density pays for itself. Delivery cost per package falls as stops per mile rise. Once your own volume in a zip code exceeds a threshold, an in-house van beats a per-package carrier rate — and every new Prime subscriber deepens the density, which lowers cost, which funds faster promises, which attracts more subscribers. The flywheel is a density engine.
Speed is a placement problem. Since same-day delivery can't outrun geography, the only way to promise it profitably is to pre-position inventory close to buyers — which is what regionalization, same-day facilities, and demand-forecast stowing all do. The forecasting layer runs at the level of individual products and zip clusters: the system estimates how many of each item a region will want next week and stows accordingly, before anyone has ordered. When it guesses right, your "same-day" item was waiting eight miles away all along. The fast promise is made by the map, then merely kept by the van.
Control of the whole chain compounds. Owning first through last mile lets one scheduler trade off truck departures, route density, and delivery promises against each other in real time — optimizations unavailable when each leg belongs to a different firm, and the same end-to-end logic that shows up wherever perishable capacity meets variable demand.
None of this machinery is secret; most of it is now the template. Rival retailers ship from stores, carriers rebuild their networks around e-commerce parcels, and the same regional-density logic spreads through the industry. Which is the evergreen takeaway: the porch thump isn't a delivery story so much as an inventory-geography story. The package was, in a probabilistic sense, already in your neighborhood before you wanted it. The click just told it whose porch — and the machine had been rehearsing the answer for weeks.
Primary Sources
Reads the primary documents — agency data, GAO reports, court opinions — and explains what they actually say.
No invented credentials: the sourcing is the credential.