The Network Design Problem Nobody Talks About

Network design gets treated as a cost optimization problem. Feed your shipment data into a solver, constrain the number of distribution centers, minimize total landed cost. The model spits out an optimal network, you implement it over 18 months, and you declare success. Transportation costs are down 12%. Warehouse consolidation savings are real. The CFO is happy.

Then a container ship runs aground in the Suez Canal. Then a pandemic shuts down ports in a single week. Then a new administration announces tariffs on components sourced from your three most important manufacturing geographies. And the network you optimized is suddenly a liability.

The problem with cost-optimal network design is not that cost does not matter. It is that cost optimization in isolation produces networks that are fragile — efficient under the conditions that existed when the model was run, brittle when those conditions change.

What Fragility Looks Like in Practice

At Applied Materials, the tariff shocks that started in 2018 and accelerated through the early 2020s forced a fundamental rethink of supply network assumptions. Networks that had been designed around stable trade relationships suddenly faced new landed cost structures that made previously optimal flows economically irrational. Parts that made sense to source from one geography at one tariff rate made no sense at a rate 25 points higher.

The companies that handled this well were not the ones with the most cost-optimal networks. They were the ones with options. Dual-sourced components. Manufacturing capability in multiple geographies. Distribution infrastructure that could flex across trade lanes. The companies that had ruthlessly consolidated to single sources and single corridors for cost reasons had no room to maneuver.

COVID made the same case more dramatically. The companies that came through the 2020-2022 supply chain disruptions with acceptable customer service levels were, almost without exception, the ones that had built in redundancy that looked wasteful on paper before March 2020. Extra inventory. Backup suppliers. Relationships with alternative freight carriers. None of it showed up as a positive on the cost optimization model.

The Option Value of Spare Capacity

Here is the conceptual shift that changes how you think about network design: spare capacity is an option, not waste.

In financial options theory, an option gives you the right — but not the obligation — to take an action at a specified price. The option has value because the future is uncertain. If you knew exactly how the future would unfold, you would not need options — you would just position yourself optimally for that known future. Options are valuable precisely because they give you flexibility to respond to outcomes you cannot predict.

Extra capacity in a supply network works the same way. A distribution center running at 70% utilization looks inefficient in a normal year. When your primary DC in another region goes offline because of a flood, a port closure, or a political disruption, that 70%-utilized facility is the option that lets you reroute volume and maintain service. The question is not whether it is underutilized — it is whether the option value of that capacity is worth its carrying cost.

This reframing matters because it changes the calculation. Instead of asking "can we justify this extra capacity against expected demand?" you ask "what is the probability that we will need this capacity as a backstop, and what is the cost of not having it when we do?" When stockout costs are high — and in semiconductor equipment supply chains they are very high — the option value of backup capacity is substantial.

Quantifying Disruption Risk

The reason resilience does not show up in traditional network design models is that disruption risk is not in the objective function. You can fix this, but it requires committing to an economic model of disruption.

The framework I have found most useful has three components:

Disruption probability. For each node and lane in your network, estimate the annual probability of a significant disruption. "Significant" means an event that takes the node offline or severely constrains the lane for a period meaningful to your service obligations. This is hard to estimate but not impossible — you can draw on historical data, insurance industry risk models, and geopolitical risk assessments. Even rough estimates are better than implicit zero-probability assumptions.

Disruption duration and severity. A node going offline for three days is categorically different from one going offline for three months. Estimate not just probability but duration distribution. Some disruptions are brief — weather events, temporary customs issues. Others are structural — trade policy changes, regional infrastructure failures. The right network response differs dramatically.

Disruption cost. What does it actually cost when a node or lane fails? In a pure cost model, you add the expedite costs and unmet demand costs and routing inefficiencies. In a semiconductor spare parts context, you also include the customer cost of equipment downtime — which is where the numbers get large very quickly. Map each critical flow in your network to the customer impact of its failure.

With these three components, you can compute an expected annual disruption cost for your current network — the probability-weighted cost of the disruptions that are plausible given your network topology. You can then evaluate alternative network configurations against both their operating cost and their expected disruption cost. The total cost of network ownership is operating cost plus expected disruption cost. Optimizing only on operating cost is optimizing for the wrong objective.

The Practical Trade-Off

There is no free lunch here. Resilient networks cost more to operate than cost-optimal networks. Dual sourcing is more expensive than single sourcing. Geographic redundancy requires capital investment. Extra capacity has carrying costs. The question is how much more, and whether it is worth it.

In my experience, the incremental cost of building genuine resilience into a network is meaningful — often 5 to 15% of total network operating cost — but rarely as large as the first-pass cost optimization model implies. The reason is that many resilience measures also have operational benefits that the pure cost model misses.

Dual sourcing, for example, creates cost reduction opportunities through supplier competition. Geographic distribution of inventory reduces expedite freight costs when demand is geographically concentrated. Flexibility in routing reduces your dependence on any single carrier, which improves negotiating leverage. The resilience case is often more economically compelling than it appears when you account for these co-benefits.

Where to Start

If you are responsible for supply network design, here is a practical starting point:

Map your single points of failure. What nodes or lanes in your current network, if disrupted, would cause severe service failures? These are your highest-priority resilience investments. The analysis does not require a sophisticated model — it requires honest identification of dependencies.
Put a number on downtime. For your most critical customers and product lines, what is the cost of a supply failure? Getting this number, even roughly, transforms the resilience conversation from philosophical to economic.
Evaluate your network against scenarios, not just forecasts. Run your network design model against disruption scenarios — supplier failure, port closure, tariff shock — not just against expected demand. Networks that perform poorly in multiple scenarios are candidates for redesign regardless of how they look in the base case.
Think in options. When evaluating investments in redundant capacity or dual sourcing, calculate the option value explicitly. What is the probability this capacity will be needed? What is the cost when it is not there?

The companies that came through the last five years of supply chain disruption in the best shape were not the ones with the most optimized networks. They were the ones with the most robust networks — built with an understanding that the future will not look like the past, and that the cost of being wrong about that assumption is enormous. Design for that reality, not for the clean world the optimization model assumes.