Satoshi Matsuoka
21 pages
The cost advantage of AI infrastructure incumbents over new entrants is structural and self-renewing: the gap never closes within 2026–2030 because depreciation continuously delivers cheap legacy capacity to whoever bought last cycle's hardware, while the 2028–2029 build vintage is the single most financially exposed tranche given current demand and pricing assumptions.
Quantitative scenario analysis using a closed-form cost model denominated in dollars per petabyte of memory bandwidth delivered ($/PB), applied to hardware vintages from H100 through a hypothetical 2029 refresh, under two HBM pricing branches (normalization in 2028 vs. shortage persisting to 2030), two pricing regimes (sticky vs. coupled), and a solvency corridor defined by aggregate token-demand growth and inference-efficiency trend. Greenfield entrant outcome probabilities (25% success / 34% mediocre / 41% loss) are derived from scenario distributions rather than empirical cohort data.
Incumbent-to-entrant cost gap is 3.2x in 2026, narrows to 1.9x in 2027 (best entry window), then re-widens to 3x (base branch) or above 4x (shortage branch) by 2029–2030. Frontier training cost rises from roughly $1.5B in 2026 to $18B–$38B by 2030; good-enough replication via RL and distillation on open bases falls toward $5M, a divergence of 3,600x–7,400x. Aggregate solvency requires approximately 2x annual token-demand growth for four consecutive years; the corridor threshold shifts between 1.6x and 2.4x across efficiency-trend bounds. The 2027 vintage is described as robust across all pricing regimes; the 2028–2029 vintage is identified as the most exposed.
Authors acknowledge the model is calibrated to a regime break (Q2 2026 shift from token maximization to token minimization) that is not yet confirmed by sufficient post-break time series, so all projections should be treated as upper bounds until two quarters of post-break data are available. The $/PB unit is valid only for saturated bandwidth-bound decode and does not capture prefill/decode ratio, context-length distribution, batchability, latency, or utilization effects. Outcome probabilities for greenfield entrants are scenario-derived, not empirically estimated from a cohort. The paper identifies five systematic upward biases in public token-demand trackers but cannot fully correct for them. Demand reallocation to on-premises and sovereign infrastructure is acknowledged but not precisely quantified.
A skeptic would note that the core conclusions rest on a single-author scenario model with no empirical validation of the $/PB framework against real operator economics, no peer review, and outcome probabilities that are essentially calibrated expert judgment rather than statistical estimates. The regime break on which the paper's most bearish conclusions depend has not yet been confirmed in the data by the paper's own admission, making the "Commoditization Crash as co-modal" finding speculative. Hardware price trajectories, especially HBM normalization timing, dominate the model, and the paper's two branches bracket a wide range without resolving which is more likely. The GLM-5.2 benchmark parity claim is stated as fact without independent replication.
If the depreciation-conveyor and solvency-corridor findings are even roughly correct, the financial risk in AI infrastructure over the next 12–24 months is concentrated in capacity ordered at peak 2025–2026 memory prices and scheduled for delivery in 2028–2029, exactly the window currently being committed. Investors and operators evaluating infrastructure financing should treat realized token-demand growth net of efficiency gains, measured in dollars not tokens, as the primary solvency indicator. The training-cost divergence finding, if sustained, implies that the mass-inference market structurally migrates to open-weight models served on depreciated fleets, compressing API margins for closed-model providers outside a narrow premium segment. These dynamics could materially affect hyperscaler capex return profiles and neocloud valuations within the 12–24 month horizon, though confirmation depends on demand and pricing data not yet available.
Market & Valuation: Infrastructure investors and hyperscaler equity analysts need to assess whether the 2028–2029 capex vintage is impaired; the paper's solvency corridor provides a concrete falsifiable threshold for that judgment.
Technical Capability: The finding that KV-cache compression is near its information-theoretic limit and efficiency gains may slow from 30%/yr to 15%/yr is directly relevant to teams betting on continued inference-efficiency as a cost-reduction lever.
Geopolitics & Security: The paper treats Chinese full-stack decoupling and the CLOUD Act jurisdictional argument as a coupled transmission mechanism for the crash scenario, with sovereign infrastructure as a third balance-sheet class outside market solvency logic.
Regulatory & Legal: The CLOUD Act analysis implies that data sovereignty cannot be achieved through open weights alone, which has procurement and policy implications for governments building national AI capacity.