Depreciation and amortisation of knowledge in supply chains

An asset that wears without a bill

Knowledge behaves like capital. Buy a patent, a mapped route, a trained analyst's judgement, and you do not expense the whole cost the day you acquire it — you spread that cost across the years it earns. Physical assets depreciate; intangible ones amortise. The distinction sounds like bookkeeping trivia until you notice that most organisations track the schedule for a forklift and ignore it entirely for a belief. A supplier's financial health, assessed eighteen months ago, is still sitting in a planning system as though solvency does not decay. It does. Economists who have tried to measure this put industrial research capital's annual decay near 15 per cent, faster for anything software-adjacent. Standard accounting almost never records the charge. The loss accrues silently, then arrives as a surprise that was never a surprise.

Supply chains make this concrete faster than most domains, because the underlying facts move on their own schedules regardless of who is watching. A tariff classification is stable until a customs authority reinterprets it. A supplier's capacity is accurate until a fire, a strike, or a debt covenant changes it. A transit time is a good estimate until a canal silts up. None of these facts announce their own expiry. The knowledge asset — "Supplier X can deliver in eleven days" — amortises whether or not the planning system updates its books.

Three postures, one thermodynamic fact

Depreciation is not primarily an accounting choice. It is what happens when a description stops tracking the thing it describes. A model of the world, held fixed, drifts toward irrelevance because the world keeps moving and the model does not — that is a statement about entropy before it is a statement about ledgers. Holding a correspondence steady costs energy: someone, or something, has to keep observing. Given that, there are only three coherent postures toward the cost. Never revalue. Revalue while looking. Revalue always. There is no fourth, because "always" has no successor — you cannot observe more continuously than continuously.

This is the axis that produces the lineage. A Large Language Model is knowledge purchased once, at a cutoff, then amortised in the dark; its errors are the depreciation charge, and the person who pays it is whoever trusted the answer, not the system that gave it. A Large World Model revalues continuously but only inside its sensed scene — accurate for what is in front of it, silent about everything outside the frame and everything already past. A Large Universe Model is the posture in which revaluation is not an event but an operating expense: every belief carries an age and a source, streams keep arriving, and the charge is booked as it is incurred rather than discovered as failure. This is not a claim that Large Universe Models exist as a deployed system. It is a claim about where the ladder of intake terminates, because "always" cannot be exceeded — what improves past that point is the price of observation and the quality of provenance, not the category of intake itself.

Supply chains are a useful test of that claim precisely because the domain refuses to sit still long enough for a frozen corpus to be adequate, and refuses to fit inside any one sensor's field of view long enough for a bounded scene to be adequate either.

Where the frozen plan fails

The characteristic failure in this domain has a specific shape: a plan survives on an assumption that a filing, buried somewhere in the stream, had already invalidated. A supply planner builds a sourcing plan assuming a supplier's export licence is current. It lapsed six weeks earlier, recorded in a customs filing nobody in the planning chain reads, because nobody's role is defined as reading it. The plan does not fail because the world changed unpredictably. It fails because a belief the plan depended on had already amortised to zero and nothing charged that loss against the plan before the shipment sat at a border.

This is depreciation exactly in Griliches's sense, just relocated from a research lab's patent stock to a logistics network's belief stock. The planner's model of "this supplier is compliant" is an asset bought at the moment of onboarding and never revalued, because revaluation was never made someone's job. The expense — a stranded container, a contractual penalty, a customer stockout — is the depreciation charge finally coming due, unpaid for months and then paid all at once, exactly the lumpy write-off pattern that afflicts a corpus retrained only periodically.

A bounded-scene system does better but not enough. Port telemetry, live and continuously ingested, will tell you the vessel is delayed the moment the vessel is delayed. That is a real gain over a static plan: the scene in view — berths, gantries, known vessels — is revalued in something close to real time. But the export filing was never in that scene. Tariff notices, supplier debt filings, a change in a subcontractor's ownership structure sit outside whatever the sensors are pointed at. The bounded scene fixes the part of the problem that looks like weather and leaves untouched the part that looks like paperwork.

The posture the failure actually calls for is the third one: every stream — manifests, port telemetry, supplier filings, tariff notices — treated as a running belief with an age and a source, so that "supplier is compliant" carries a timestamp and a provenance trail rather than being a fact asserted once and assumed durable. The plan does not need to be rebuilt from scratch weekly. It needs each load-bearing assumption tagged with how old it is and how it would be falsified, so that a stale filing shows up as a stale filing rather than as a shipment held at a border three weeks later.

The objection that concedes real ground

Half of what a sourcing plan depends on barely moves — geography, physics, contract terms fixed for a year. Re-checking all of it continuously is waste. Watch the fast-moving ten per cent and leave the rest alone.

This is correct, and the argument needs it to be correct. Li and Hall's decay estimates run from near zero for mathematics to roughly 40 per cent annually for computing-adjacent capital; supply chain knowledge spans a similarly wide range in miniature. A port's draught limit is stable for years. A single supplier's cash position can turn in a quarter. Continuous intake is not a demand to re-observe everything at the same rate. It is a demand that the system know which of its beliefs is a draught limit and which is a solvency estimate, and schedule its attention accordingly. A frozen plan cannot make that distinction at all — it has no field recording how old any given assumption is, so a fixed geographic fact and a six-week-stale compliance filing sit in the plan with identical, false, confidence. Rate-matched revaluation only becomes possible once the intake is open enough to carry an age on every belief; sampling policy becomes a decision the system makes, not a wall the system runs into.

Booking every belief's provenance and age is itself a system that needs maintaining. Schemas for supplier filings go stale. Calibrations on port telemetry drift. You have not eliminated the depreciation charge — you have added a second one, for the machinery that tracks the first.

Also correct, and probably the sharper objection for this domain, because supply chain data is notoriously heterogeneous — a tariff notice arrives as a PDF, a manifest as an EDI message, a supplier filing as a regulatory disclosure in a different jurisdiction's format. Reconciling those into a single provenance record is real, ongoing work, and that work decays too: mapping logic goes stale when a customs authority changes its filing schema, and nobody notices until reconciliation starts failing silently. This is a genuine cost, not a rounding error.

That is the whole of what the third posture buys here: not the abolition of decay, which is thermodynamically impossible, but the conversion of an invisible impairment into a visible reconciliation backlog — a number a supply planner can look at before the shipment moves, rather than a number discovered after it has stopped moving.