← AI in the supply chainC4 · LESSON 4 OF 7

C4, Lesson 4: Where to Hold Stock: Networks, Regions and the Cost of Being Close

Tags: Sourced / General / Our view / Illustrative. About 13 minutes.

1. The story: Amazon cuts its network into eight regions, 2021 to 2023

In a July 24, 2023 article on its science site, Amazon described how its US fulfillment network, which had grown quickly during the pandemic, risked becoming "overcomplicated and unwieldy." A long-term planning science team had warned in late 2020. The issue, Amazon said, was that each fulfillment center (FC) served not only its own area but customers all over the country: "We would fulfill customer orders from the FCs near them until we couldn't anymore, and then, okay, it's coming from wherever we still have capacity." [Sourced: Amazon Science, July 24, 2023; Amazon's own account] Amazon's illustration: to ship 10,000 products to 100 distant locations from 10 FCs, if every FC sent a truck of 10 items to every location, that is 1,000 long-haul trucks; if the 100 locations are split into 10 regions, each served by one FC, the same volume needs 100 trucks carrying 100 packages each. [Sourced: Amazon Science; a simplified illustration] By mid-2021 Amazon committed to the idea. The team spent about a year using network-optimization tools to model options, and chose eight regions, the most it could have "without sacrificing speed or requiring excessive inter-regional movement of inventory." Each region had to carry the breadth of selection customers expect. Amazon said the share of orders filled entirely from FCs within a region had risen to 76% and was expected to climb. [Sourced: Amazon Science; company claims] On its October 26, 2023 earnings call, CEO Andy Jassy said regionalization had "more impact than we optimistically expected," bringing higher in-region stock, shorter routes and faster delivery. [Sourced: Retail Dive, Oct 31, 2023, reporting the Oct 26 call; Amazon-reported, no independent assessment exists] The key trade: put stock closer to customers and trips get shorter, but you hold stock in more places. [Our view]

Amazon kept the approach after 2023. In January 2024 an Amazon executive said the company would keep refining its regionalization model, and a January 2026 paper by Amazon authors describes the eight-region design. These are still Amazon's own statements. [Sourced: Doug Herrington, aboutamazon.com, Jan 30, 2024; INFORMS Journal on Applied Analytics, 2026, Amazon-authored; company claims]

2. The one idea

Where you hold inventory is a choice between freight cost and speed (closer is better) and inventory cost (fewer locations pool risk better). There is a best number of locations for your business, and it depends on your freight rates, your demand variability and how fast customers expect delivery. [General]

3. The kitchen-table version

You sell the $12 bottle across the country. One garage holds everything: cheap to run, but every parcel travels far. Four garages, one in each region: parcels travel less, but each garage needs its own cushion of stock, so you hold more in total. [General]

4. Two forces

  • Freight falls with distance: more locations, shorter trips and fuller trucks (Amazon's truck illustration).
  • Safety stock rises with the number of locations: if you split the same demand across n locations, the total safety stock needed rises roughly with the square root of n. This is the standard "risk pooling" result. [General] AI and optimization tools help by modeling millions of combinations of locations, flows and inventory (Amazon says its team handled "millions of variables and constraints"). They do not remove the trade-off. [Sourced: Amazon Science; Our view]

5. Worked example: how many locations?

All numbers [Illustrative], computed by script. A business ships 100,000 units a year across 300 items. Freight costs $0.00045 per unit per mile. Each item's national weekly demand has a standard deviation of 40, lead time is 2 weeks, 95% service, item cost $7, carrying cost 20% (see Lesson 3). Average ship distance falls as locations rise.

LocationsAvg milesFreightSafety stock (per item)Holding costTotal
11,200$54,00093$39,200$93,200
2700$31,500132$55,400$86,900
4400$18,000187$78,400$96,400
8250$11,250264$110,900$122,100
In this made-up case, two locations is cheapest, four is about equal to one, and eight costs most. Freight savings flatten out while holding costs keep rising. Change the inputs (higher freight rates, higher service targets, slower items) and the answer shifts. If customers will pay for or choose faster delivery, that value belongs in the table too. [Our view]
Amazon's result (eight regions) fits its scale: tens of millions of items and huge volume per region make more locations worth it. A small business is closer to the table above. [Our view]

6. What AI adds, and what it does not

  • Adds: network design models, order placement ("which FC should fill this order"), inventory positioning forecasts by region.
  • Does not: negotiate freight, fix a bad forecast, or choose your service promise.
  • Amazon's Jassy said the company kept "fine-tuning the placement algorithms to enable even more in-region fulfillment and to further increase consolidation into fewer shipments." [Sourced: Retail Dive, Oct 31, 2023]

7. What it costs and what it breaks

  • More stock: each extra location holds its own cushion.
  • Stranded inventory: stock in the wrong region sits while another sells out.
  • Fixed costs: leases, staff and systems per site.
  • Complexity: more flows to plan (Amazon's own warning about an "unwieldy" network). [Sourced: Amazon Science]
  • Rebalancing: you need rules for moving stock between regions.

8. Monday checklist

  1. List where you hold stock and where customers are.
  2. Compute average miles shipped and freight per unit.
  3. Compute total safety stock across locations.
  4. Add one more or one fewer location on paper and recompute (Section 5).
  5. Decide how fast you promise delivery and what that is worth.

9. Try it yourself (no code, 30 minutes)

Rebuild Section 5 in a spreadsheet with your freight rate and units. Find the number of locations with the lowest total cost.

10. Self-check

  1. How many regions did Amazon choose, and why not more? Eight; more would sacrifice speed or require too much inter-regional movement of inventory.
  2. What happens to total safety stock when you split stock across more locations? It rises, roughly with the square root of the number of locations.
  3. In the example, which option was cheapest? Two locations.
  4. Name a cost Amazon's own article points to. An overcomplicated, unwieldy network.
  5. Whose numbers are Amazon's 76% and Jassy's "optimistically expected"? Amazon's own claims.

11. Where this is going

Planning tools will place stock more finely, and delivery promises will keep tightening. The trade between closeness and stock will stay. [Our view]

Sources

Review flags

  • The main source is Amazon's own science article (read in part: the first half). All Amazon numbers (76%, eight regions, the truck illustration) are company statements, not independent measurements.
  • Every Amazon figure in this lesson is Amazon-reported. No independent assessment was found (the 2026 INFORMS paper is Amazon-authored). RESOLVED (Oct 7, 2026) as an accepted limitation, now stated in the lesson: Amazon's later statements were read. Doug Herrington (Amazon, Jan 30, 2024) said Amazon regionalized its U.S. network, shipped nearly 600 million more items from in-region fulfillment centers in Q4 2023 than a year earlier, and would "refine our regionalization model" and tweak it over time. The INFORMS Journal on Applied Analytics paper (Jan 2026, Amazon authors) says the U.S. network was partitioned into eight interconnected, largely self-sufficient regions after 1.5 years of operations research work. Nothing found contradicts the eight-region description; nothing independent confirms the results. Later changes beyond these were not searched.
  • The risk-pooling rule is the standard square-root result; in practice demand across regions is not perfectly independent.
  • Example inputs (freight rate, miles, volumes) are illustrative; the "two locations cheapest" result depends on them.

Claim ledger (C4 Lesson 4)

#ClaimSourceVerifiedStatusRe-check
1Network grew fast in the pandemic; warning in late 2020; "overcomplicated and unwieldy"Amazon Science Jul 24, 20232026-10-06Read first half; company accountnone
2Truck illustration 1,000 vs 100 trucksSame2026-10-06Readnone
3Eight regions; why not more; 76% in-regionSame2026-10-06Read; company claim2027-01
4Jassy: "more impact than we optimistically expected"; placement algorithm quoteRetail Dive (article Oct 31; call Oct 26, 2023); Amazon-reported2026-10-06Readnone
5Network-cost arithmeticOur computation2026-10-06Illustrative, script-checkednone