Close read — Matsuoka, arXiv:2607.07207, §6 and §9.1/§9.3

What text this is built from

Read: the full PDF, all 22 pages, downloaded from https://arxiv.org/pdf/2607.07207 (HTTP 200, 812,523 bytes, PDF 1.7, arXiv GenPDF (tex2pdf:8def8d8)) and read directly page-by-page. Nothing here is derived from the abstract, from search snippets, or from a fetch-and-summarize pass.

Retrieval routes, in the order tried:

1. https://arxiv.org/abs/2607.07207 — worked, confirmed metadata.
2. https://arxiv.org/html/2607.07207v1 — worked, but returns a summarized rendering, not raw text. Not relied on.
3. https://arxiv.org/pdf/2607.07207 — worked, direct download. This is the source for everything below.

Nothing was inaccessible. All 22 pages, all 9 figures, all 5 tables, Appendix A, the Acknowledgments/AI-disclosure, and all 27 references were read.

Correction to the record: arXiv's abstract page says 21 pages; the actual PDF is 22 pages (references [15]–[27] sit on p.22). Minor, but the parent ingest repeats the 21-page figure.

Correction to the record (material): the parent ingest lists the author as "Independent (no institutional affiliation listed)." The title page states RIKEN Center for Computational Science (R-CCS), [email protected]. Matsuoka directs Japan's national supercomputing centre (Fugaku). This materially changes how the paper should be weighted — see WHERE THIS IS WEAK §4.

(A) The vintage-breakeven analysis — §6, Figure 4 (p.8)

What the model actually is

§6 is one paragraph of text plus one bar chart. There is no table. The parent ingest describes a "Table-style breakdown"; that is not in the paper. Every number below was read off Figure 4's bars (precision ≈ ±0.5pp), cross-checked against the four numbers the §6 prose states explicitly.

The breakeven question is narrow and specific: for each purchase-year cohort, what share of the tokens it serves must earn premium pricing for that fleet to break even? It is a break-even mix requirement, not an IRR, not a payback period, not an NPV.

Inputs (all from Appendix A / Table 5, p.20)

The breakeven inherits the whole cost stack of Eq. 1 (§2, p.4):

Cost_PB = [ (P_acc · k)/(8760·T) + (W/1000)·PUE·c_e + o ] / (B · η · 3.6)   [$/PB]

• P_acc accelerator price · k system multiplier = 1.5× (chassis, networking, DC shell)
• T = 4-year straight-line depreciation; 8760 = hours/yr (implies 100% uptime assumed)
• W watts · PUE 1.3 · c_e $0.08/kWh
• o = $0.25/GPU-hr allocated operations (staff, network, maintenance)
• B theoretical HBM bandwidth (TB/s) · η MBU = 0.50 Hopper / 0.55 Blackwell / 0.60–0.62 Rubin
• B · η · 3.6 = petabytes moved per hour

Vintage prices, the actual driver of the whole result (Table 5):

The breakeven condition

A vintage breaks even when its blended realized revenue per PB covers its full cost (Eq. 1 with the amortization term). Revenue is a mix of two prices, and the premium share s is solved for:

s · P_premium + (1 − s) · P_mass  =  Cost_PB(full, that vintage)

P_mass is defined as the incumbent floor + 30% margin (Table 5). This is the hinge: the mass price is pinned to the depreciated incumbents' marginal cost, which falls every year as older fleets finish amortizing ($0.054/PB in 2026 → $0.022/PB by 2029, Fig. 1). So a new vintage is always priced against a floor set by someone else's sunk capital.

The two pricing regimes

• Coupled — routing arbitrage drags premium down with the mass floor: P_premium = 7 × P_mass. A ratio peg.
• Sticky — premium holds at $0.40/PB in absolute terms, defended by enterprise contracts, compliance lock-in, switching costs.

This is the mechanism the parent ingest lost. Because coupled is a ratio and sticky is an absolute, they cross over:

• In 2026 the floor is $0.054 → mass ≈ $0.070 → coupled premium ≈ $0.49/PB, which is above the $0.40 sticky level. Coupled is the friendlier regime for a 2026 buyer.
• By 2029 the floor is $0.022 → mass ≈ $0.029 → coupled premium ≈ $0.20/PB, now half the sticky level. Coupled is lethal for a late buyer.

Derived results — Figure 4 (p.8), all 16 cells

Shaded band on the chart = 10–20% plausible realized premium share. Above the band = does not clear.

Internal-consistency check that validates the chart reading: the 2026 row is identical across HBM branches, exactly as it must be — that hardware was already purchased, so the future HBM branch cannot alter its capital cost. The model is self-consistent here.

What each vintage's exposure actually depends on

Not on operating skill, and not on the HBM branch alone. It depends on the ratio of the vintage's purchase price to the incumbent floor prevailing during its service life — i.e. on how much already-depreciated capacity is competing with it. The curve is U-shaped and regime-dependent:

• 2026 vintage — paid peak memory prices with no compensating price umbrella. Worst cell is sticky at ~31%. Counterintuitively it prefers coupled pricing, because in 2026 7× a not-yet-collapsed mass price beats a fixed $0.40.
• 2027 vintage — robust in every one of its four cells (7.5–11.1%), entirely inside the plausible band. The paper calls it the rational entry year in all futures. Reason: it buys after the worst of the memory spike but before the floor has fully collapsed.
• 2028–29 vintages — survive comfortably under sticky (10.3–17.3%) but are broken by coupled-plus-shortage (21.1% and 37.6%). The 2029 cohort is the single most exposed cell in the paper.

The paper's own conclusion (§6, p.7): routing success breaks the Stargate-class 2028–29 commitments; routing failure breaks today's peak-price 2026 buyers. Someone holding peak-vintage capacity is structurally underwater in every regime — the pricing regime only selects who.

Assumptions vs. derived results — the split

ASSUMED (inputs, not findings):

• The 7× coupled ratio and the $0.40/PB sticky level. Both stipulated; neither is estimated from data.
• The 10–20% plausible premium share band. This is the benchmark that turns every number above into a verdict, and it is asserted in §6 with no citation. The only empirical anchor appears 3 sections later (§9, p.10): Anthropic taking ~42% of OpenRouter revenue on ~11% of token share. One platform, one month.
• Mass price = floor + 30% margin.
• 4-year life, 100% uptime, $0.08/kWh, PUE 1.3, MBU 0.50–0.62, 1.5× system multiplier.
• The two HBM price branches.

DERIVED (genuine model output):

• The 16 breakeven percentages.
• The U-shape and — the real contribution — the regime crossover, which falls out of pegging one regime to a ratio and the other to an absolute. This is not assumed; it is a consequence.
• The identification of 2027 as robust in all four of its cells.

Jump-to refs: model = §2 Eq. 1, p.4. Parameters = Table 5 / Appendix A, p.20. Breakeven = §6, p.7 (prose) and Fig. 4, p.8 (all values). Floor trajectory = Fig. 1, p.5.

(B) The demand-measurement critique — §9.1 (p.11–12) and §9.3 (p.13)

The starting point (§9, p.9–10)

Public trackers the paper assembles in Fig. 7 (p.12):

• Google [12]: 9.7T tokens/month (May 2024) → 480T (I/O 2025) → 1.3 quadrillion (Oct 2025) → 3.2 quadrillion (May 2026). Stated 7× YoY; most recent seven months annualize nearer 4.7×.
• Microsoft Foundry [13]: ~30% QoQ in FY26 Q3 ≈ 2.9× annualized.
• OpenRouter [11]: ~5T tokens/week (spring 2025) → >31T (late May 2026) ≈ 4–6× YoY. Agentic traffic overtook human traffic ~1 Feb 2026, consuming ~15× the tokens per request.
• China aggregate [14]: ~30T/day (mid-2025) → ~180T/day (Feb 2026) — 6× in eight months.
• Goldman Sachs [8]: 24× 2026→2030 ≈ 2.2×/yr compound.

Face-value read (§9, p.11): gross 4–7×/yr, net of 30–45%/yr efficiency ⇒ delivered-bandwidth demand ~2.5–5×/yr — above the 1.6–2.4× solvency threshold.

The five biases — as the paper names them (§9.1, p.11)

Verbatim labels, in the paper's own order:

1. Supply-injected volume — the largest single tracker aggregates inference the platform operator pushes onto users at its own discretion (AI summaries in search results, productivity surfaces), plus multimodal processing whose token counts are enormous by construction. Measures product strategy, not willingness to pay.
2. Substitution — router-platform growth includes workloads migrating onto the router from direct APIs, so platform growth exceeds market growth by the migration rate.
3. Subsidy — a meaningful share of reported volume, particularly in the Chinese price war, is quantity demanded at below-cost prices and would not survive normalization.
4. Quality composition — the ~15× agentic token multiplier is substantially redundant context re-reading: the lowest-value tokens in the economy and the first to be eliminated as token optimization becomes enterprise discipline. Budget-rationing behaviour is read as the leading edge of that optimization, not a peripheral counter-signal, since 2027 budgets will be set with full knowledge of the 2026 burn.
5. Base effects and disclosure selection — two mechanisms under one label. (a) The agentic era began in earnest only Feb 2026, so mid-2026 growth rates annualize the steepest point of an S-curve. (b) The visible sample is biased upward because operators publish token metrics when, and only when, they are spectacular.

Revised estimate (§9.1, p.12): quality-adjusted underlying demand growth is plausibly 2–3×/yr rather than 4–7×/yr — at or modestly above the 1.6–2.4× threshold band, with a visibly decelerating second derivative.

How well is each bias evidenced?

The structural problem, and it is the important finding of this close read: the paper lists five biases and then states a revised number. It never decomposes the adjustment. There is no accounting of how much each bias removes, no error bars, no sensitivity. The move from "4–7×" to "2–3×" — which is the position-relevant number in the entire paper, because it lands the central case barely above a threshold the same paper derives at 1.6–2.4× — is a judgment presented immediately after a qualitative list, in a way that reads as though the list produced it. It did not. Four of the five biases contribute no quantity at all.

Two further corrections in §9.1 (p.12) the parent ingest omits entirely

• Denominate the corridor in dollars, not tokens. Solvency depends on PB moved × achievable dollars per PB. The neutral router's blended realization is ~$1 per million tokens against flagship closed pricing near $25–30. So bandwidth demand can grow while revenue per unit of capacity stalls. The paper's stated crash signature: tokens growing while dollars stall. For a desk this is the most operational line in the paper and it is not in the parent ingest.
• Meta/xAI resale is a demand signal, not just a supply event. Capacity offered to the market is capacity the largest internal AI operators concluded they did not need. In the author's words, "insiders selling forward is what a marked-down internal demand forecast looks like" (Matsuoka, §9.1). Also absent from the parent ingest.

§9.2 — omitted from the parent ingest, and load-bearing

The reallocation channel: metered cloud token demand — the only demand that services hyperscaler/neocloud revenue and therefore capex solvency — can decelerate even while total inference grows, because inference migrates onto customer-owned hardware (DGX Spark-class boxes [23], Mac-mini clusters, rack-scale private inference) where it is invisible to and unmonetizable by the providers financing the buildout. Plus a sovereignty channel: non-US enterprises (Japan explicitly [22], "the same posture recurs across much of Europe and Asia") enter directly at the on-premises open-weight stage, skipping the metered cloud-API stage entirely. Chinese open-weight models already carry ~60% of neutral-router token traffic.

This is effectively a sixth bias, and arguably the most consequential, since it breaks the link between "AI usage grows" and "the buildout gets paid" without requiring demand to fall at all. Paper's framing: demand relocation reads as demand destruction on the balance sheets that matter.

§9.3 — the projection-vintage problem (p.13)

A correction not to the measurements but to their dates. Every optimistic projection in the survey is a product of the token-maximization regime: Goldman's 24× was published early May 2026; tracker growth rates run through May; agentic multipliers were measured at the height of everything-in-context engineering. If the doctrinal inversion to token minimization occurred across the client industry in Q2 2026 — which budget-rationing behaviour and collapsing average tokens-per-request indicate — then all trailing-CAGR extrapolations measure a regime that no longer exists, and extrapolating across the break is a category error.

Consequences the author accepts:

• Pre-break projections — including this report's own quality-adjusted 2–3× estimate — are upper bounds.
• The go-forward prior must be re-anchored on post-break data, which at writing spans only weeks.
• The downside path (bandwidth demand peaking ~2028) is promoted from tail risk to co-equal stress case.
• Tracking cadence should move monthly, not quarterly, for at least two quarters.

One honest decomposition the author supplies: token minimization is an efficiency shock, not a demand shock. It enters through the efficiency parameter (pushing toward the 45%/yr stress bound, lifting the solvency threshold toward 2.4×+) while the number of tasks can keep growing independently. Whether task growth offsets tokens-per-task compression is directly observable weekly in tokens-per-task and dollars-per-task. Countervailing datum: neutral-router volumes were still accelerating through late June 2026.

WHERE THIS IS WEAK

1. A 73× realized-cost spread swamps a 2–4× vintage effect. The paper cites [25] showing effective cost on identical H100 hardware ranging $0.21 to $15.25 per million output tokens depending on utilization and offered load. The entire vintage-breakeven edifice discriminates between cohorts separated by 2–4×. If operational execution moves realized economics by up to 73×, then vintage timing is a second-order term dressed as a first-order one — and §11's claim that "vintage timing dominates operating skill" is close to backwards on the paper's own cited evidence. The author flags the spread as a scope qualification (§2, p.4) but never propagates it into §6.

2. The 10–20% plausible premium band is asserted and carries the whole verdict. Every "unsustainable" call in §6 is a comparison against this band. Move it to 25–35% and the 2026 vintage clears, 2028 clears everywhere, and only the 2029 coupled-shortage cell fails — the paper's headline conclusion substantially dissolves. It is sourced to nothing in §6; the nearest anchor (Anthropic ~42% of revenue on ~11% of tokens, §9 p.10) is a single platform in a single month, and is a revenue-share datum being used to bound a token-share parameter.

3. The 2–3×/yr revision is asserted, not computed. See (B) above. Five biases, four unquantified, one number out the other end — and that number is what places the central case at the solvency boundary. This is the paper's most consequential figure and its least supported. Note also the convenient arithmetic: Goldman's independent 2.2×/yr sits almost exactly on the paper's independently-derived ~2× threshold, a coincidence the author remarks on (p.10) but does not treat as a reason to suspect the threshold construction.

4. Undisclosed structural interest — the paper's meticulous AI disclosure has no counterpart for the author's own position. Matsuoka directs RIKEN R-CCS. The paper: (a) identifies AI4SIS — mission-funded scientific and industrial-R&D compute, i.e. precisely his institution's demand class — as the corridor's "inelastic floor" and "rationing-proof" base load, and as the strongest structural argument for retaining material probability on the benign Jevons scenario (§7, p.8); (b) reads LineShine LX2 as vindicating the "matrix-enhanced general-purpose CPU (SME/SVE) fed by on-package HBM-plus-DDR, in direct lineage from Fugaku's A64FX" — his own institution's architecture — and cites his own [17][18] Dongarra–Hoefler–Matsuoka series as the frame (§10.1, p.14–15); (c) closes §11 with a policy recommendation explicitly addressed to Japan, arguing sovereign capability needs efficient open-weight inference on bandwidth-first architectures rather than frontier-scale training. None of this is disqualifying — he is among the best-placed people alive to hold these views — but a paper this careful about disclosing that Claude helped draw the figures does not mention that its architectural and policy conclusions align with its author's institutional bet. Weight §7's AI4SIS floor and §10.1's LineShine reading accordingly.

5. AI-authorship reflexivity runs deeper than the parent ingest recorded. Acknowledgments (p.20): models, figures and drafts developed with Anthropic's Claude (Fable 5), and review comments from OpenAI's ChatGPT 5.5 that were adopted into Sections 8–9 and Tables 2–3 — that is, into the measurement critique, the scenario probabilities, the pricing-regime framing, and the financing-web classification. The two sections this close read exists to examine are among those shaped by review from a model built by a company whose commercial prospects those same sections assess. The author does flag this reflexivity himself, to his credit. Then note the loop closes twice: reference [27], the export-control episode that is the evidentiary basis for the 12% Bifurcation scenario, is Anthropic's statement about restrictions applied to Claude Fable 5 — the model that co-developed the paper.

6. The scenario probabilities are iterated judgments, and the paper says so. §8 (p.9): across five revision rounds the Crash scenario went 15% → peak 30% → tempered to 25% "on external-review grounds." Table 3's per-scenario outcome splits are explicitly labelled "elicited subjective probabilities — judgmental assessments conditioned on each scenario's definitions, not model outputs." So the headline 25/34/41 custom-silicon distribution is a weighted average of expert priors presented in a table, not a derived result. Treat as calibrated opinion.

7. Uptime and utilization are assumed away in Eq. 1. The 8760·T denominator implies 100% uptime, and MBU is a single scalar per platform. Real fleets have maintenance windows, stranded capacity, and diurnal load. Every $/PB figure in the paper is therefore a best-case floor, and — importantly — the entrant is likelier to be under-utilized than the incumbent with an established book, which means the true entrant/incumbent gap is probably wider than 3.2×. This one cuts in the paper's favour, but it is unmodelled either way.

8. Single-author, non-peer-reviewed preprint, econ.GN, 12 days old at time of reading. Matsuoka's HPC authority is genuine and directly relevant to the bandwidth-bound framing — which is the paper's best idea. His authority does not extend to enterprise SaaS pricing dynamics, subsidy economics, or credit analysis of the circular-financing web, and §10.1's financing material leans on press aggregations [19] whose contract-status heterogeneity the author himself flags in Table 2.

WHAT THIS DOES AND DOESN'T LICENCE US TO CLAIM

Supported — we can say this and cite the paper:

• $/PB (dollars per petabyte of memory bandwidth delivered) is a defensible model-agnostic unit for saturated, bandwidth-bound decode, given in closed form with explicit parameters (§2 Eq. 1; Table 5).
• Under the paper's parameters, a 2026 GB300-class new build costs ~$0.174/PB against a depreciated-H100 incumbent floor of ~$0.054/PB — 3.2× — narrowing to 1.9–2.0× in 2027 and re-widening to ~3× (base) or >4× (shortage) by 2029–30 (§3, Fig. 1).
• The depreciation conveyor: the incumbent cost advantage rotates among incumbents by vintage and never transfers to entrants within the horizon. This is a clean structural argument, not a forecast, and it is the paper's strongest contribution.
• 2027-vintage capacity is robust in all four regime/branch cells (7.5–11.1% breakeven premium share); 2026 and 2028–29 vintages are each fatally exposed to one regime — 2026 to sticky (~31%), 2028–29 to coupled-plus-shortage (21–38%) (§6, Fig. 4).
• Public token trackers have at least five identifiable upward biases, of which the quality-composition bias (redundant agentic context re-reading; ~15× multiplier; ~60% of enterprises throttling spend) is genuinely evidenced across three independent sources [11][24][26].
• Metered cloud demand can decelerate while total inference grows, via on-premises migration and the non-US sovereignty channel (§9.2). Independent of any view on aggregate demand.
• Pre-Q2-2026 demand projections predate a doctrinal shift and should be treated as upper bounds; only weeks of post-break data existed at writing (§9.3).
• The right thing to watch is realized token-demand growth net of efficiency, denominated in dollars as well as tokens, against a 1.6–2.4× threshold band. Crash signature: tokens growing while dollars stall (§9.1, §11).

NOT supported — over-reach if we claim it:

• "Demand growth is 2–3×/yr." That is the author's judgment after a qualitative bias list, undecomposed, and self-labelled an upper bound in §9.3. Cite it as his estimate, never as a measurement.
• "The 2026 vintage is underwater." It requires ~24.5–31.3% premium share against a 10–20% band the paper asserts without sourcing. The conclusion is a function of an unsourced parameter. Report both numbers or neither.
• "2028–29 capacity requires 21–38% premium share." Only in the coupled + shortage cells. Under sticky pricing those same vintages need 10.3–17.3% and clear comfortably. Always state the regime and the branch — there are 16 cells, not 6.
• "Vintage timing dominates operating skill." The paper says this (§11) but its own cited evidence [25] — a 73× realized-cost spread on identical hardware — argues the opposite. Do not repeat it as fact.
• "There is a 41% chance a custom-silicon entrant loses money." Table 3 is explicitly elicited subjective probability, not model output, revised across five rounds. It is a calibrated opinion.
• "The paper shows Samsung/SK Hynix/Micron are doing X." It contains no supplier-level analysis whatsoever — no per-supplier capex, allocation, margin, or strategy. Memory scarcity is a single exogenous input. It cannot corroborate any memory-maker-specific claim. (The parent ingest states this correctly; preserving it.)
• "Scenario probabilities are computed." They are iterated expert judgments, revised five times, tempered by external review (§8).
• "An independent analyst finds..." Single-author preprint, 12 days old, no peer review, developed with substantial assistance from two AI systems built by companies the paper values, by an author whose institution's architectural and policy bet the paper's conclusions favour. Frame it as an argument to test, never as a source to cite for a fact.

Was the deep read worth it? — verdict

Yes, materially. The primary text changes four things versus the summary-derived ingest:

1. A factual error is corrected — the author is RIKEN R-CCS, not independent. This is the difference between an anonymous preprint and a paper by the director of Japan's national supercomputing centre, and it surfaces an undisclosed structural interest running through §7, §10.1 and §11.
2. The §6 result is structurally different from the ingest's rendering — 16 cells, not 6; four vintages, not three; a bar chart, not a table. The ingest's "2028-29" lumping hides that 2029 (37.6%) is nearly twice as exposed as 2028 (21.1%).
3. The ingest inverts the 2026 vintage's regime exposure — it reports ~31% under coupled; the paper puts 2026 at ~24.5% coupled and ~31.3% sticky. The ingest's own table contradicts its own conclusion sentence. The underlying mechanism — coupled is a ratio peg, sticky is an absolute peg, so they cross over as the floor collapses — is the actual finding, and it flips the position read: 2026-vintage owners are hurt by premium stickiness, not helped by it.
4. The most position-relevant material was missing entirely — the 2–3×/yr revised demand figure, the dollar-denomination correction (~$1/Mtok blended vs $25–30 flagship), the "tokens growing while dollars stall" crash signature, the §9.2 reallocation channel, and the Meta/xAI-resale-as-demand-signal read. All of these are more actionable for a desk than anything the ingest carried from §9.

What the read did not change: the headline architecture (depreciation conveyor, 2027 as the robust vintage, the five biases, the scenario set, the LineShine decoupling) was directionally right in the ingest. The summary got the shape; it got the load-bearing details wrong in ways that would have produced a wrong position read.

Close read of: Memory Scarcity, Open Models, and the Restructuring of the AI Industry, 2026–2030, Satoshi Matsuoka (RIKEN R-CCS), arXiv:2607.07207 [econ.GN], submitted 8 July 2026. Full PDF, 22 pages, read 2026-07-20. Parent ingest: kb/hardware/raw/memory-scarcity-ai-industry-restructuring-2026-2030.md.