The Depreciation Conveyor & Vintage Economics

The standard commoditization story assumes that when hardware prices normalise, the cost advantage of whoever bought early erodes and everyone converges. Matsuoka's preprint (arXiv 2607.07207, 2026-07-08) models the opposite and gives the mechanism a name: a depreciation conveyor delivers newly amortized fleets to incumbents faster than hardware prices normalize, so the entrant-incumbent cost gap never closes.

Every figure on this page is modelled or projected, not measured. They are outputs of a single-author, non-peer-reviewed preprint by Satoshi Matsuoka (RIKEN R-CCS) — the director of Japan's national supercomputing centre, a correction to the earlier "independent author" note that matters because it surfaces an undisclosed structural interest (see the desk close-read). The full 22-page PDF has now been read by the hardware desk; this page is reconciled to that read. The numbers are recorded because the mechanism is a serious answer to the topic's central question — not because they are established.

The conveyor

Fleets bought in prior years finish depreciating and become effectively free capacity on the incumbent's books. If that stream of newly-amortized capacity arrives faster than the price of new hardware falls, the entrant — who must buy at today's price — is permanently behind. The modelled gap does not decay monotonically; it compresses and then re-widens:

The 2027 trough is the interesting part: it is the one window in which an entrant is close to parity. That the gap re-widens afterwards is the claim that makes this a structural argument rather than a temporary supply-shock observation.

Matsuoka names the entrants directly: Meta and xAI entering compute resale on fleets bought before the memory repricing — i.e. the conveyor is not hypothetical, it is already being monetised by parties whose vintage happens to be favourable.

⚠ First-order conflict: is hardware even where the cost story lives? (added 2026-07-23)

This entire page rests on a premise: that hardware vintage and memory pricing are the load-bearing variables in inference cost structure. A competing source, compiled the same day, cuts directly against that premise and this KB records the conflict without resolving it.

Tiered Super-Moore's Law (Mingdeng Du, arXiv 2603.28576, 2026-03-30) decomposes the ~600x LLM token-price decline via growth accounting and finds:

• Total-factor productivity / software-architectural innovation accounts for ~103.7% of the cost reduction.
• GPU hardware contributes ~–0.9%.
• It also dates a May-2024 structural break (Chow F=5.74, p=0.005) from technology-driven to competition-driven price decline, with market concentration falling (HHI 4,558 → 2,086).

If hardware contributes roughly nothing (indeed slightly negative) to the fall in prices, then the depreciation conveyor — a mechanism entirely about hardware vintage and memory cost — may govern supply/capacity without ever having been the source of falling prices. The two claims can be partly reconciled ("memory scarcity gates capacity while software drives price"), but as stated they locate the cost lever in different places: Matsuoka in the silicon-and-memory layer, Du in the software/architecture layer.

Provenance asymmetry to weigh: Du's decomposition is an estimated econometric result read from the abstract page only (26pp PDF unread; the >100% TFP residual and the Malmquist internals need the full text to judge). Matsuoka's is a modelled scenario set reconciled to a full desk close-read. Neither is peer-reviewed. Kept on the record, unadjudicated — see frontier.md → Conflicts, item 7 and the same conflict on the price-decline distribution.

Vintage economics — capacity is not fungible

The second structural claim: capacity bought in different years is a different asset, with different survivability under different pricing regimes. The abstract's "2026/2028-29 fatally exposed, 2027 robust" is directionally right but lossy — the desk close-read shows it is a 16-cell grid (4 vintages × 2 regimes × 2 HBM-price branches read off Fig. 4), and the abstract-derived rendering inverted the 2026 exposure. Breakeven premium-share required (plausible band 10–20%; above = does not clear):

The mechanism is the finding: coupled is a ratio peg (premium = 7× the mass floor); sticky is an absolute peg ($0.40/PB) — so they cross over as the incumbent floor collapses. Therefore:

• 2026 vintage's worst cell is STICKY (~31.3%), not coupled — it prefers coupled. Premium stickiness hurts the 2026 buyer, it does not help them. (This is the inversion the old rendering got backwards.)
• 2027 is robust in all four cells (7.5–11.1%) — the rational entry year in every future.
• 2028–29 clear under sticky (10.3–17.3%) but break under coupled+shortage (21.1% / 37.6%); 2029-coupled-shortage (37.6%) is the single most exposed cell — nearly 2× the 2028 figure the abstract's lumping hid. Always state the regime AND the branch.

The investment reading, stated as a question rather than a call: if vintage determines solvency, then "who owns AI compute" is a less useful question than "which year's compute do they own, and at what memory price."

The solvency corridor

The most consequential single sentence in the abstract for this topic: solvency of the announced buildout is confined to a corridor requiring roughly 2x annual token-demand growth for four years, with sticky premium pricing. Both conditions, jointly — not either.

That second condition is directly observable, and this KB already tracks it. "Sticky premium pricing" is price-rung persistence. When OpenAI held the frontier rung at $5/$30 and the strong rung at $2.50/$15 across the GPT-5.6 generation (2026-07-16, own-dataset primary), that was premium stickiness being maintained in public, on a specific date. The Token Price Index is therefore not just a pricing curiosity — it is an instrument on a named solvency variable. A frontier rung cut, when it comes, is the leading indicator that one of the two corridor conditions has broken.

A derived comparison, flagged as derived. This KB separately holds Goldman's forecast of 24x token-consumption growth to ~120 quadrillion tokens/month by 2030 (moderate evidence, logged 2026-07-17 — see frontier.md). Compounding 24x over four years implies ~2.21x/yr; over five years, ~1.89x/yr. The reported base year is ambiguous in what this KB holds, so the honest statement is: the most-cited bullish demand forecast implies roughly 1.9-2.2x/yr — i.e. it lands on Matsuoka's ~2x threshold, not comfortably above it. This arithmetic is the desk's, derived from two stated figures; neither source makes the comparison. And note the circularity risk: Matsuoka separately argues the token trackers those forecasts rest on overstate monetizable demand (see the cost-measurement problem), which would push the realised figure below the threshold.

Training-cost bifurcation

A supporting projection with direct bearing on who can afford to stay at the frontier:

• Luxury tier: $18-38B per frontier training run by 2030.
• Mass tier: previous-frontier parity via RL/distillation, falling toward $5M.

If both hold, the interesting number is the ratio — roughly three to four orders of magnitude between being at the frontier and being one generation behind it. That is the economic engine behind frontier-capable open-weight models (Matsuoka cites GLM-5.2), and it is the mechanism by which the price of yesterday's capability collapses, which is exactly what the price-decline distribution measures from the buy side.

The scenarios — the model does not resolve the question

Matsuoka assigns probabilities across five 2026-2030 outcomes:

Read against this topic's umbrella question, the two poles are tied at 25%. The first serious quantitative model to engage "commodity infrastructure or durable margin?" prices it as a coin flip between the two. That is a genuinely useful result — it says the question is not yet decidable on current evidence — and it should be reported as such rather than mined for whichever branch suits a thesis. These are iterated expert judgments, not computed outputs: the close-read shows they were revised across five rounds (Crash 15% → peak 30% → tempered to 25% on external-review grounds, §8), and Table 3's per-scenario outcome splits — including the 25/34/41 custom-silicon distribution below — are explicitly labelled elicited subjective probabilities, not model outputs.

Key Claims

• A depreciation conveyor keeps the entrant-incumbent cost gap open indefinitely: 3.2x (2026), 1.9x (2027), re-widening to 3-4x (2029-30). Evidence: weak-moderate — modelled, single-author preprint, input assumptions undisclosed in the abstract (Matsuoka)
• Buildout solvency requires ~2x annual token-demand growth for four years AND sticky premium pricing — jointly. Evidence: weak-moderate (modelled corridor) (Matsuoka)
• Vintage determines survivability, on a 16-cell grid: 2027 robust in all 4 cells (7.5–11.1%); 2026 broken by sticky (~31.3%, prefers coupled); 2028–29 broken by coupled+shortage (21.1% / 37.6%). Evidence: weak-moderate (modelled; read off Fig. 4 by the desk close-read — corrects the abstract's lossy 3-way lumping and its inverted 2026 exposure) (Matsuoka, close-read)
• Training bifurcates: $18-38B per frontier run by 2030 (luxury) vs previous-frontier parity falling toward $5M (mass). Evidence: weak-moderate (projection to 2030) (Matsuoka)
• A greenfield custom-silicon entrant removes the merchant margin but not the memory premium — 25% success / 34% mediocre / 41% loss. Evidence: weak — this is ELICITED SUBJECTIVE PROBABILITY (Table 3), NOT a model-output distribution; revised across 5 rounds (Matsuoka)
• Scenario probabilities place Commoditization Crash and Rotating Landlord Oligopoly in a dead heat at 25% each. Evidence: weak (iterated expert judgments, revised 5 times; not computed) (Matsuoka)
• Meta and xAI are named as entering compute resale on pre-repricing fleets. Evidence: weak (asserted in the abstract; not independently checked in this KB) (Matsuoka)
• China's LineShine LX2 — domestic HBM on a standard ISA — decouples its cost curve from the memory crisis. Evidence: weak (asserted; not independently checked) (Matsuoka)
• CONFLICT (unresolved): a competing decomposition attributes ~103.7% of the token-price decline to software/TFP and –0.9% to GPU hardware — cutting against this page's hardware-vintage-as-cost-lever premise. Memory may gate capacity while software drives price; the two sources locate the lever in different layers. Evidence: estimated econometric decomposition, abstract-page only, vs modelled scenario; both preprints; unadjudicated (Tiered Super-Moore's / Du)
• DERIVED (desk, not a source claim): Goldman's 24x-by-2030 token-demand forecast implies ~1.9-2.2x/yr depending on base year — landing on Matsuoka's ~2x solvency threshold rather than above it. Evidence: derived arithmetic over two moderate-evidence figures; base year ambiguous

Benchmarks & Data

All figures MODELLED/PROJECTED as of the 2026-07-08 preprint. None are measured costs.

Open Questions

• Which pricing regime kills the 2026 vintage, and which kills 2028-29? RESOLVED 2026-07-23 (desk close-read): 2026 is killed by sticky (~31.3%, and it prefers coupled); 2028–29 are killed by coupled + HBM-shortage (21.1% / 37.6%). See the 16-cell grid above.
• What amortisation schedule and hardware-price path generate the 3.2x → 1.9x → 3-4x path? Without the inputs, the conveyor is a plausible mechanism attached to unfalsifiable numbers.
• Does the conveyor survive a real accounting check? Hyperscaler useful-life assumptions for AI servers are disclosed in 10-Ks and have been revised upward in recent years; that is a primary-source test of the conveyor's core premise and this KB has not run it.
• Is "sticky premium pricing" holding right now? The Token Price Index is the instrument; it needs repeated release-event observations across Anthropic and Google, not just OpenAI, to answer.
• Do the scenario probabilities move if the demand critique is applied to the demand forecasts? Matsuoka argues trackers overstate demand and prices solvency on demand growth — worth checking whether the scenarios already internalise his own critique.

Related Concepts

• Token Pricing & the Inference-Margin Question — this concept is the first quantitative answer offered to that umbrella question (and it answers "25/25, undecided").
• The Cost-Measurement Problem — the other half of the same preprint ($/PB unit, demand-tracker critique) plus Patil's utilization work.
• Price-Rung Persistence — "sticky premium pricing" is one of the two named solvency conditions, and it is directly observable in the Token Price Index.
• The Price-Decline Distribution — training-cost bifurcation is a mechanism that drives the buy-side collapse in the price of yesterday's capability.
• HBM4 Memory Architecture — the memory premium the conveyor runs on; the hardware/ topic owns the supply-side read.
• The Price-Decline Distribution — carries the same software-vs-hardware conflict from the price-trend side; Du's tier half-lives live there.

Backlinks

Pages that reference this concept:

• Token Pricing & the Inference-Margin Question
• The Cost-Measurement Problem
• Price-Rung Persistence

Changelog

• 2026-07-23 (compile) — Recorded the software-vs-hardware first-order conflict: Tiered Super-Moore's Law (Du, 2603.28576) attributes ~103.7% of the price decline to software/TFP and –0.9% to GPU hardware, cutting against this page's hardware-vintage-as-cost-lever premise. Added a flagged conflict section + Key Claim; kept unadjudicated (memory may gate capacity while software drives price). No change to the conveyor mechanism itself.
• 2026-07-23 (reconcile) — Reconciled to the hardware desk's full-PDF close-read. Load-bearing corrections: author is RIKEN R-CCS (not independent); the vintage result is a 16-cell grid (the abstract's 3-way lumping was lossy and inverted the 2026 exposure — 2026 is killed by sticky, not coupled); the 25/34/41 custom-silicon split and the five scenario probabilities are elicited/iterated expert judgments, not model outputs. Resolved the "which regime kills which vintage" open question. Mechanism/architecture unchanged.
• 2026-07-22 — Created from the Matsuoka preprint (arXiv 2607.07207), ingested this compile. Records the depreciation-conveyor mechanism, vintage-breakeven exposure, the two-condition solvency corridor, training-cost bifurcation and the five-scenario probability set — all marked modelled. Connects "sticky premium pricing" to the Token Price Index as an instrument on a named solvency variable.