Tiered Super-Moore's Law: Price Evolution, Production Frontiers, and Market Competition in LLM Inference Services
First systematic empirical analysis of LLM token pricing across 3,237+ models (2020-2026); ~600x price decline; 'Tiered Super-Moore' hypothesis (economy 1.10yr / mid 1.55yr price half-life vs 2yr Moore benchmark; flagship/reasoning resists via ~31.5x premium); cost decline ~103.7% software/architecture-driven, ~-0.9% hardware.
Abstract
"This paper provides the first systematic economic analysis of token pricing in the large language model (LLM) inference market. Assembling a novel dataset integrating OpenRouter API data (318 models), Epoch AI records (3,237 models), and 62 cross-validated milestone observations spanning 2020-2026, we document an approximately 600-fold decline in token prices and propose the 'Tiered Super-Moore' hypothesis."
Key Contributions
- Builds the largest known cross-validated dataset on LLM token pricing: 318 models (OpenRouter), 3,237 models (Epoch AI), 62 milestone observations, 2020-2026.
- Documents an ~600-fold decline in token prices across the study window.
- Proposes the "Tiered Super-Moore" hypothesis: price declines are not uniform across the model stack — cheaper tiers fall faster than Moore's Law, premium tiers resist.
- Identifies a structural break in market dynamics around May 2024 (F = 5.74, p = 0.005), separating a technology-driven price-decline regime from a competition-driven one.
- Decomposes the cost decline by driver: software/architectural innovation ≈ 103.7% of the reduction, GPU hardware ≈ -0.9% (i.e., hardware improvements alone did not meaningfully drive price down over this period — algorithmic/serving-stack efficiency did essentially all the work).
Methodology
- Data Envelopment Analysis (DEA) to establish production-frontier efficiency across models/providers.
- Chow structural-break testing to locate and validate the regime change in pricing dynamics.
- Cost decomposition analysis separating hardware vs. software/architecture contribution to price decline.
Results
| Segment | Price half-life | vs. Moore's Law (2-yr benchmark) |
|---|---|---|
| Economy tier | 1.10 years | ~1.8x faster |
| Mid tier | 1.55 years | ~1.3x faster |
| Flagship / reasoning tier | Resists the trend | Reasoning commands a ~31.5x premium over non-reasoning price |
- Structural break: May 2024, F = 5.74, p = 0.005 — before this point pricing behaves like a technology curve; after, it behaves like a competitive market (entrants, undercutting, tiered product lines).
- Cost-decline attribution: ~103.7% software/architecture, ~-0.9% GPU hardware (the negative figure implies hardware cost/perf was a mild drag or noise term relative to the software-driven decline, not a meaningful independent contributor).
Limitations
- Author and institutional affiliation not resolved from the abstract-level fetch; full PDF body not ingested in this pass (abstract + reported findings only — flagged as summary-derived).
- "Reasoning premium" (31.5x) is a cross-sectional average; the paper does not (per available summary) break this down by lab or model generation.
- DEA-based frontier estimates are sensitive to model/provider list composition (OpenRouter + Epoch AI); coverage of closed, enterprise-only pricing (e.g., custom enterprise contracts) is likely incomplete.
Why this matters for the KB
Directly instantiates the AI LENS's "Compute economics" required chapter and the content-focus.md spine (inference economics / tokenomics — "who captures the value; the real constraints"). The tiered-decay finding is a testable frame for MenFem's Token Price Index work in kb/inference-economics/ and for reading GPT-5.6's three-tier (Sol/Terra/Luna) launch this same window as tier-differentiated pricing behavior, not a single "the market" price.
Source: Tiered Super-Moore's Law (arXiv:2603.28576), Mingdeng Du, submitted 2026-03-30.