Memory Scarcity, Open Models, and the Restructuring of the AI Industry, 2026-2030

Read discipline — reconciled, not re-read. The prior version of this file was built from the arXiv abstract only. The hardware desk has since read the full 22-page PDF and produced a close-read (kb/hardware/raw/memory-scarcity-2607.07207-closeread.md, 2026-07-20). This file is now reconciled to that read. The material abstract-only errors it corrects — author affiliation, the "modelled" mislabel on the custom-silicon distribution, and the vintage-grid structure — are recorded in the CORRECTIONS block below. Everything the close-read verified takes precedence over anything the abstract implied.

CORRECTIONS applied 2026-07-23 (from the desk close-read)

1. Author is RIKEN R-CCS, not independent. This is material: it is a paper by the director of Japan's national supercomputing centre, and the close-read surfaces an undisclosed structural interest running through the paper's AI4SIS "inelastic floor" argument (§7), its favourable read of the Fugaku-lineage LineShine LX2 architecture (§10.1), and a policy conclusion addressed explicitly to Japan (§11). Weight §7/§10.1/§11 accordingly.
2. The 25% / 34% / 41% custom-silicon outcome distribution is ELICITED SUBJECTIVE PROBABILITY, not model output. Table 3 is explicitly labelled "elicited subjective probabilities — judgmental assessments conditioned on each scenario's definitions, not model outputs." The prior version tagged it "modelled." It is calibrated opinion.
3. The scenario probabilities are iterated expert judgments, revised across five rounds (Crash 15% → peak 30% → tempered to 25% on external-review grounds; §8) — not computed.
4. The vintage-breakeven result is a 16-cell grid (4 vintages × 2 pricing regimes × 2 HBM-price branches), read off a bar chart (Fig. 4), not a 3-row / coupled-only structure. See the corrected vintage read below — in particular the 2026-vintage exposure was INVERTED in the abstract-derived rendering.

Abstract

"We analyze how four forces restructure the AI industry over 2026-2030: the DRAM/HBM price surge, frontier-capable open-weight models (GLM-5.2), rapid inference-efficiency gains (near-Shannon-limit KV-cache compression, lightweight local runtimes), and the entry of Meta and xAI into compute resale on fleets bought before the memory repricing. Formulating inference economics in dollars per petabyte of bandwidth delivered ($/PB) -- model-agnostic for bandwidth-bound decode -- we show the entrant-incumbent cost gap never closes: a depreciation conveyor delivers newly amortized fleets to incumbents faster than hardware prices normalize (3.2x in 2026, 1.9x in 2027, re-widening to 3-4x by 2029-30). Training bifurcates into a luxury tier ($18-38B per frontier run by 2030) and a mass tier (previous-frontier parity via RL/distillation falling toward $5M). Solvency of the announced buildout is confined to a corridor requiring roughly 2x annual token-demand growth for four years with sticky premium pricing; a measurement critique shows public token trackers overstate monetizable demand, and all pre-Q2-2026 projections predate the industry's shift from token maximization to token minimization. A vintage-breakeven analysis finds 2026 and 2028-29 capacity each fatally exposed to one pricing regime, with only the 2027 vintage robust. A greenfield custom-silicon entrant removes the merchant margin but not the memory premium (central outcome: 25% success/34% mediocre/41% loss, improvable via staged go/no-go gates). China's LineShine LX2 -- domestic HBM on a standard ISA -- decouples its cost curve from the memory crisis. Scenario probabilities: Rotating Landlord Oligopoly 25%, Commoditization Crash 25%, Jevons Absorption 20%, System-Layer Re-differentiation 18%, Geopolitical Bifurcation 12%. Solvency now depends on monetized bandwidth demand, premium stickiness, and vintage ownership."

Key Contributions (inference-economics lens)

• A different cost unit. Inference economics in dollars per petabyte of bandwidth delivered ($/PB) rather than per token, argued model-agnostic for bandwidth-bound decode. The close-read confirms the closed form (§2 Eq. 1) with explicit parameters (Table 5): 4-yr straight-line depreciation, 100% uptime assumed, PUE 1.3, $0.08/kWh, $0.25/GPU-hr ops, MBU 0.50–0.62, 1.5× system multiplier.
• The depreciation conveyor. Already-amortized fleets arrive faster than hardware prices normalise, so the entrant-incumbent cost gap never closes. The close-read calls this the paper's strongest contribution (a structural argument, not a forecast).
• A solvency corridor, not a forecast. ~2x annual token-demand growth for four years, with sticky premium pricing — both conditions. The desk's threshold band is 1.6–2.4× (the abstract's "~2x").
• A measurement critique of demand, and (from the close-read, absent in the abstract) two operational refinements the desk rates as the most actionable lines in the paper: denominate the corridor in dollars, not tokens (neutral-router blended realization ~$1/Mtok vs $25–30 flagship), with crash signature "tokens growing while dollars stall"; and a §9.2 reallocation channel — metered-cloud demand can decelerate even while total inference grows, via on-prem migration and the non-US sovereignty channel.
• Vintage economics — corrected below; a 16-cell grid, not "2026/2028-29 vs 2027."
• Scenario probabilities — iterated expert judgments (5 revision rounds), NOT computed.

Vintage economics — CORRECTED (16 cells, from the close-read's Fig. 4)

The abstract's "2026 and 2028-29 fatally exposed to one regime, 2027 robust" is directionally right but the abstract-derived rendering inverted the 2026 exposure. The two regimes cross over because coupled is a ratio peg (premium = 7× the mass floor) and sticky is an absolute peg ($0.40/PB) — so as the incumbent floor collapses over time, which regime is friendlier flips.

Breakeven premium-share required per cell (plausible band = 10–20%; above it does not clear):

• 2026 vintage's worst cell is STICKY (~31.3%), not coupled — it prefers coupled (~24.5%), because in 2026 7× a not-yet-collapsed mass price beats a fixed $0.40. The prior "2026 hurt by premium stickiness" reading is the correct one; any implication that 2026 is helped by stickiness is wrong.
• 2027 is robust in all four cells (7.5–11.1%) — the rational entry year in every future.
• 2028–29 survive under sticky (10.3–17.3%) but are broken by coupled+shortage (21.1% and 37.6%). 2029 coupled-shortage (37.6%) is the single most exposed cell in the paper — nearly 2× the 2028 figure the abstract's lumping hides.
• Always state the regime AND the branch — there are 16 cells, not 6.

Results (all figures MODELLED / PROJECTED unless noted, as of the 2026-07-08 preprint)

Cross-link the close-read makes (paper-1 ↔ paper-2)

The close-read's strongest internal critique (WHERE THIS IS WEAK §1): the paper's own cited evidence [25] — Patil, arXiv:2606.11690 — shows effective cost ranging $0.21–$15.25/M output tokens on identical H100 hardware (a ~73× realized-cost spread from utilization alone). That operational spread swamps the 2–4× vintage effect this paper builds its edifice on. So §11's claim that "vintage timing dominates operating skill" is close to backwards on the paper's own cited evidence. Now that this desk has fully read Patil and anchored its dollar level to $6.98/GPU-hr Azure H100 NVL on-demand (see beyond-per-token-pricing-...md), that 73× spread is firmly established, not floating — which sharpens the critique.

Named entities

• GLM-5.2 — frontier-capable open-weight model.
• Meta, xAI — entrants into compute resale on fleets bought before the memory repricing. Close-read adds: the resale is itself a demand signal — "insiders selling forward is what a marked-down internal demand forecast looks like" (§9.1).
• LineShine LX2 (China) — domestic HBM on a standard ISA, claimed to decouple its cost curve from the memory crisis; the close-read flags this reading as aligned with the author's own Fugaku/A64FX architectural lineage.

Limitations (from the close-read)

• Single-author, non-peer-reviewed preprint (econ.GN).
• The 10–20% plausible-premium-share band that turns every vintage number into a verdict is asserted in §6 with no citation; move it to 25–35% and the headline "unsustainable" conclusions substantially dissolve.
• The 2–3×/yr revised demand figure is asserted, not computed — five biases named, four unquantified, one number out the other end, and that number is what places the central case at the solvency boundary.
• Eq. 1 assumes 100% uptime and a single MBU scalar per platform; every $/PB is a best-case floor.
• AI-authorship reflexivity: models/figures/drafts developed with Anthropic's Claude (Fable 5); ChatGPT 5.5 review comments adopted into §§8–9 and Tables 2–3 — i.e. into the very measurement critique and scenario probabilities, which assess the commercial prospects of the companies whose models co-developed the paper.

Source: Memory Scarcity, Open Models, and the Restructuring of the AI Industry, 2026–2030 by Satoshi Matsuoka (RIKEN R-CCS), arXiv:2607.07207 [econ.GN], submitted 2026-07-08. This inference-economics file reconciled 2026-07-23 to the hardware desk's full-PDF close-read: kb/hardware/raw/memory-scarcity-2607.07207-closeread.md.