The Physics of Supply Planning

Supply chain planning is often described as judgment.

A planner looks at demand, inventory, purchase orders, lead times, supplier promises, warehouse constraints, forecast error, budget limits, and whatever crisis just appeared in the inbox. Then the planner decides what to buy, move, reserve, expedite, allocate, delay, or escalate.

From the outside, that can look like experience. Pattern recognition. Spreadsheet craft. Human intuition over messy data.

But that is not the deepest truth.

Supply chain planning has physics.

Not physics as decoration. Not physics as a metaphor borrowed to make software sound scientific. Physics in the operational sense: repeatable quantitative relationships, conserved quantities, delays, buffers, constraints, feedback loops, nonlinear tradeoffs, and causal chains that determine what is possible.

Inventory is flow through time.

Lead time is a delay.

Capacity is a limit.

Variability must be buffered.

A promise consumes future supply.

A plan that violates constraints is not ambitious. It is false.

That is what I mean by the physics of supply chain planning.

The phrase is worth taking seriously because there is a real body of work behind it. Factory Physics, by Wallace Hopp and Mark Spearman, is built around a rule-based, data-driven approach to operations planning and control, and its coverage spans EOQ, ROP, MRP, variability, queueing, flow laws, batching laws, scheduling, supply-chain management, and capacity management.

The phrase also draws a boundary. Planning is not forecasting. It is not visibility. It is not a dashboard. Planning is the discipline of converting uncertain future demand and qualified future supply into feasible, governed, time-phased actions and commitments under the laws of inventory, flow, variability, lead time, capacity, constraint, and authority.

A good planning system makes those laws explicit.

A weak planning system hides them.

Planning is where the future becomes governable

The simplest definition I know is this:

Future demand − usable future supply = required action

That looks almost too simple. But every word matters.

Not historical demand. Future demand.

Not visible supply. Usable future supply.

Not a recommendation in the abstract. Required action.

The equation becomes real only when evaluated across the dimensions that make supply chains hard:

item × location × time × constraint × priority × commitment state

Do we need this item in this location? By what date? For which customer, facility, production line, patient population, contract, program, or mission? Is the demand firm, forecasted, dependent, safety-stock-driven, or policy-driven? Is the supply on hand, inbound, planned, reserved, quarantined, expired, substitutable, late, low-confidence, or already promised to someone else? Is the supplier approved? Does the order violate MOQ, budget, production capacity, transport capacity, shelf life, quality release, working capital, or time fence?

Only after those questions are answered does subtraction become planning.

Forecasting estimates future demand. Visibility observes state. A dashboard displays signals. Planning decides what must happen.

A forecast can be excellent and still not be a plan. A visibility layer can show inventory and still not know whether that inventory is usable. A dashboard can point to a stockout and still not know whether to buy, move, reserve, expedite, substitute, allocate, or revise a promise.

Planning is the layer where the future becomes an object. It can be evaluated, reserved, revised, traded, governed, and explained.

The compact planning backbone looks like this:

Plan = f(
  state,
  demand,
  supply,
  policies,
  constraints,
  lead_times,
  commitments,
  engine_version
)

The stronger planning invariant adds the pieces that make a plan governable in real organizations:

decision context =
  accepted state
  governed demand
  candidate and usable supply
  policies and constraints
  lead times and calendars
  commitments and protected claims
  objective and risk profile
  horizon and effective dates
  uncertainty controls
  overrides and accepted judgment
  engine / model / solver / configuration version

official plan =
  published recommendation
  + named authority path
  + commitment transition, if any

If the same decision context and calculation method are supplied, the plan should be reproducible to the declared standard: exact for deterministic rules, within tolerance for optimization, tied to seeds or scenario sets for stochastic methods, and forensically explainable where full recomputation is not practical.

That is not a philosophical preference. It is the basis of trust. If a plan cannot be replayed, it cannot be audited. If it cannot be audited, it cannot be governed. If it cannot be governed, it should not be allowed to make or change consequential commitments.

Inventory is time made visible

Little's Law is one of the cleanest examples of operational physics:

L = λW

In operating language:

Inventory = Throughput × Flow Time

Little's Law was published by John D. C. Little in 1961 as a proof for the queueing formula L = λW, and it became central because of its practical importance in queueing, operations research, and operations management.

If a system ships 100 units per week and the average unit spends four weeks in the system, the system will carry about 400 units in process:

100 × 4 = 400

This is not a KPI preference. It is a relationship.

If someone says, "Cut inventory in half," the physics asks: which variable are you changing? Will throughput fall? Will flow time shrink? Will demand be rejected? Will another buffer absorb the shock?

Inventory is not just stuff. It is time made visible. Some inventory is waste. Some inventory hides poor process control. But some inventory is compensating for long lead times, unreliable suppliers, forecast uncertainty, capacity constraints, or bad data.

The buffer function has to live somewhere. If you remove inventory without reducing variability, shortening lead time, adding capacity, improving reliability, changing service expectations, or accepting more shortages, the system will pay elsewhere: in expedites, missed commitments, longer queues, lower service, planner heroics, or hidden inventory.

Physics does not care what the quarterly target was.

Variability must be buffered

The most practical idea in Factory Physics is that systems with variability must be buffered. In the presence of variability, Factory Physics frames the available buffers as inventory, time, and capacity.

That should be printed on the wall of every planning room.

If demand varies, supply slips, production yields fluctuate, suppliers miss dates, transport lanes are unreliable, or inventory records are wrong, the system must absorb that variability somehow.

You can hold more inventory. You can reserve extra capacity. You can quote longer lead times. You can accept lower service. You can expedite. You can ration.

You cannot pretend the variability is gone.

Safety stock is the most familiar expression of this principle:

Safety stock = z × σ × √L

Suppose average lead time is 9 days, demand standard deviation is 20 units per day, and the target service factor is 1.65, roughly associated with a 95% cycle service level under common assumptions.

Safety stock = 1.65 × 20 × √9
Safety stock = 99 units

Now increase the service expectation. Using the same demand variability and lead time:

90% service ≈ 77 units
95% service ≈ 99 units
98% service ≈ 123 units
99% service ≈ 140 units

This is the inventory-service tradeoff. Higher service means more buffer, better reliability, more capacity, more time, or more expensive recovery options. The physics is not telling the business what to value. It is telling the business that tradeoffs cannot be wished away.

Utilization is not free

In many operating meetings, high utilization sounds like virtue. Keep the factory loaded. Keep people busy. Keep assets productive.

But in a variable system, high utilization creates waiting time. As utilization approaches capacity, queues and delays rise sharply.

The Kingman or VUT approximation captures the relationship:

E(Wq) ≈ [ρ / (1 − ρ)] × [(ca² + cs²) / 2] × τ

Kingman's formula is an approximation for mean waiting time in a G/G/1 queue and is commonly described as depending on utilization, variability, and service time.

Take a simple case where arrival variability and process variability are both moderate and average process time is one day.

At 80% utilization:

waiting time ≈ 0.8 / 0.2 × 1 × 1 = 4 days

At 90% utilization:

waiting time ≈ 0.9 / 0.1 × 1 × 1 = 9 days

Nothing magical happened between 80% and 90%. The system did not become lazy. The math changed.

This is why a plan that assumes every resource can run at 100% forever is not efficient. It is brittle. It has no capacity buffer left. Any variability becomes queue, delay, expedite, or failure.

Capacity is not just a number in a master-data table. It is a buffer against variability.

Lead time turns math into action

A shortage detected after the release date is not a planning signal. It is an emergency.

The minimum lead-time rule is:

release date = need date − lead time

If a product is needed on Day 90 and replenishment lead time is 60 days, normal action had to start on Day 30. If the shortage is discovered on Day 70, the normal planning window is gone. The remaining options are recovery options: expedite, substitute, borrow, ration, split, defer, or fail.

A real planning system must distinguish need date, ship date, arrival date, usable date, release date, commitment date, and frozen date.

The usable date may be later than the arrival date because of receiving, quality release, customs clearance, inspection, kitting, labeling, quarantine, cold-chain verification, or putaway. The frozen date may prevent a mathematically valid change from being operationally allowed.

This is why time fences matter. The same recommendation can be smart in the planning zone and reckless in the frozen zone. Planning software that ignores time fences creates bad advice with good math.

Usable supply is not visible supply

The most dangerous word in supply chain planning may be "available."

Available to whom? Available where? Available when? Available under which policy? Available after which commitments?

A warehouse may show 1,000 units on hand. That does not mean 1,000 units are available to cover the demand in front of you.

Some may already be reserved. Some may be allocated to higher-priority demand. Some may be blocked for quality release. Some may expire before use. Some may be in the wrong location. Some may require a transfer that arrives too late. Some may be visible in the ERP but physically wrong.

This is why gross-to-net is not simple subtraction. It is permission.

Oracle's MRP documentation describes net-requirements calculation as evaluating the master schedule, bills of material, scheduled receipts, on-hand inventory balances, lead times, and order modifiers, then creating recommendations to release or reschedule orders.

That is already far more than:

demand − inventory

The hard planning question is not: do we have enough total supply?

The hard question is: which supply is allowed to cover which demand, by which date, under which commitments, constraints, priorities, and policies?

That is pegging.

A plan needs to explain the causal chain between demand and supply. If 500 units of demand are covered, the system should be able to say what supply covers them. If 200 units are short, it should explain why the apparently visible supply did not count: late, reserved, low confidence, wrong location, wrong status, wrong policy, wrong time fence, not approved, not usable.

This is where planning systems either earn trust or lose it.

Constraints bend the plan

A naive system says, "Order more."

A real planning system asks, "Can we actually do that?"

Constraints are not edge cases. They are part of the physics: minimum order quantities, order multiples, supplier capacity, production capacity, transport capacity, storage capacity, budget, working capital, approved-source rules, route availability, calendars, shelf life, batch size, quality release, cold-chain requirements, regulatory holds, and substitution eligibility.

These constraints turn smooth math into discrete, governed action.

Economic Order Quantity gives a classic example:

EOQ = √(2DS / H)

If annual demand is 10,000 units, order cost is $50, and annual holding cost is $10 per unit:

EOQ = √((2 × 10,000 × 50) / 10)
EOQ ≈ 316 units

EOQ is one of the oldest classical inventory models; Ford Whitman Harris first presented it in 1913.

The formula is useful. It tells us where ordering cost and holding cost balance under simple assumptions. But the real plan may not be allowed to order 316 units. The supplier MOQ may be 500. The order multiple may be 120. Now the feasible order is not 316. It is 600.

The formulas reveal the tradeoff. The constraints determine the feasible action.

A plan that violates hard constraints is not aggressive. It is fiction.

Promises consume the future

A commitment ledger is not administrative overhead. It is how a planning system prevents the same future supply from being promised twice.

The system must distinguish uncommitted, proposed, planned, approved, firmed, released, reserved, allocated, promised, consumed, cancelled, and superseded.

Those states are not labels for a workflow diagram. They determine what future supply still exists.

Available-to-promise asks:

What existing usable, uncommitted supply can still be promised?

Capable-to-promise asks:

What can become available if we make, buy, move, or expedite within lead-time and constraint limits?

Microsoft's supply-chain documentation describes CTP as accounting for material availability, capacity availability, and lead times for the bill of material and route. It also defines ATP as the quantity that is available and promiseable to a customer on a specific date, including uncommitted inventory, lead times, planned receipts, and issues.

ATP protects the business from false promises. CTP extends the question into feasible future action.

When planning and commitment are separated, organizations drift into duplicate promises. Sales thinks supply is available. Operations knows it is reserved. Procurement sees an inbound order. Finance has not released funds. A country team assumes allocation is approved. A central team treats it as proposed. The spreadsheet says covered. Reality says otherwise.

The future has already been consumed, but the system does not know it.

That is not a reporting problem. That is a physics problem with a governance failure.

Feedback amplifies error or compounds trust

Supply chains are feedback systems with delays.

Demand changes. Supply slips. Inventory is corrected. A supplier misses. A customer cancels. A planner overrides. A shipment arrives late. A forecast is revised. An allocation decision creates scarcity somewhere else.

The plan begins decaying the moment reality moves.

That is why the planning loop matters:

plan → execute → observe → update state → detect variance → replan → preserve evidence

The bullwhip effect is the canonical example. Lee, Padmanabhan, and Whang describe the bullwhip effect as information distortion where orders can have higher variance than sales, and distortion increases upstream in the supply chain. They identify demand signal processing, rationing games, order batching, and price variations as four major causes.

This is not because planners are foolish. It is because feedback with delay is hard.

Planning must therefore do more than calculate a single answer. It must preserve the evidence trail, compare plan to actuals, expose variance, and distinguish signal from noise. A system that learns from overrides, late receipts, forecast bias, lead-time drift, and repeated exceptions can build trust over time. A system that discards that evidence forces every planning cycle to start from institutional amnesia.

Forecasting Physics and Planning Physics are not the same

Forecasting and planning are often blurred because both deal with the future. But they are different disciplines.

Forecasting Physics asks:

Given observations, history, drivers, uncertainty, and policy,
what demand should we expect?

Planning Physics asks:

Given accepted demand, qualified supply, constraints, commitments,
objectives, uncertainty, and authority,
what action is feasible, required, and governable?

Forecasting is probabilistic and evidence-weighted. Planning is governed and reproducible once the decision context is declared, even when the method includes optimization, scenarios, heuristics, or AI-assisted estimates.

A forecast may say demand is likely to be 1,000 units, with uncertainty between 800 and 1,300. Planning must decide what to do with that uncertainty. It may hold safety stock, place an order, split supply, reserve capacity, delay a commitment, create an exception, or ask a human to approve a tradeoff.

Forecasts estimate.

Plans commit.

Forecasting should feed planning only through governed materialization: accepted forecast output becomes canonical demand with quantity, timing, confidence, source, and policy context. Otherwise, weak forecast truth becomes unsafe planning truth.

Why black-box planning is dangerous

This is where AI must be handled carefully.

Large language models can be useful in planning systems. They can summarize evidence, explain exceptions, draft narratives, compare scenarios, identify likely drivers, and make the planning record easier to understand.

But they should not become silent authority. Learned models may generate validated estimates or recommendations when governed, but a generative system should not silently decide which supply covers which demand, choose policy, or advance commitment state without evidence, validation, and named authority.

In high-consequence supply chains, the question is not whether an AI-generated answer sounds plausible. The question is whether the number can be traced.

What input created it? What rule was applied? What engine version computed it? What assumptions were used? What constraints were checked? What supply was pegged? What commitment state changed? Who approved it? What changed since the prior run?

The world is uncertain. Demand is uncertain. Supply is uncertain. Lead time is uncertain. Records are uncertain. Supplier promises are uncertain.

But once those uncertainties are represented as inputs, confidence levels, policies, constraints, scenarios, and tolerances, the planning decision must be reproducible or forensically explainable.

The system can say: given this uncertain demand, this usable supply, this lead time, this capacity, this service policy, this commitment ledger, and this approval rule, here is the required action.

That is verifiable.

That is governable.

That is the opposite of a black box.

Physics-grade planning

The phrase "the physics of supply chain planning" is useful because it says something stronger than "best practice."

Best practices change by industry, company, and fashion. Physics names the underlying structure.

It says there are relationships you cannot negotiate away.

If lead time is longer than the time remaining, normal replenishment cannot arrive on time. If service level rises under uncertainty, required buffer rises. If utilization approaches capacity in a variable system, queues and delays grow. If demand information is distorted upstream, order variability can amplify. If supply is already committed, it cannot safely be promised again. If an order violates MOQ, capacity, shelf life, route, budget, or approval constraints, it is not feasible. If the same declared context and calculation method cannot reproduce or explain the same plan, the system cannot be audited.

Supply chain planning is not a guessing process. It is the discipline of converting uncertain future demand and qualified future supply into feasible, governed, time-phased actions under the laws of inventory, flow, variability, lead time, capacity, constraint, and commitment.

Those laws are the physics of planning.

A planning system earns trust when it makes that physics explicit: every number computed, every action traceable, every commitment governed, every exception tied back to the causal chain.

Many planning environments scatter the physics. Spreadsheets distribute it across cells, tabs, macros, copies, and human memory. Dashboards observe fragments of it. Forecasting tools estimate one important input to it. ERP systems execute parts of it. AI copilots can explain pieces of it.

But real planning requires the whole machine: accepted state, governed demand, qualified supply, policies, constraints, lead times, calendars, commitments, objectives, uncertainty controls, engine/version provenance, publication authority, and feedback.

The goal is not to remove human judgment. The goal is to put judgment where it belongs.

Humans should decide policy, priority, acceptable risk, escalation, tradeoffs, and commitments.

The system should compute the consequences.

That is the planning standard high-consequence supply chains deserve.

Not vibes.

Not black boxes.

Not heroic spreadsheet archaeology.

Physics-grade planning: reproducible, governed, traceable, and honest about the constraints of the real world.