04 / 05
Chapter Four · The New Moore's Law

Intelligence,
commoditized.

The price of frontier-class intelligence halves every three months. Moore's Law halved the cost of computing every two years.

Moore's Law started as an engineering observation: transistors on a chip doubling every 24 months. Its lasting consequence was economic. Computing power got cheaper by half every two years, and that compounding is what put computers on desks, then in pockets, then inside everything. The AI equivalent is on the same axis, intelligence getting cheaper, and it is moving roughly 10× faster. A query that cost a dollar eighteen months ago costs a cent today. The model that needed a data center fits on a phone.
Doubling time · log scale · lower is faster

The price of intelligence collapsed.

Frontier-class intelligence has fallen roughly 200× since March 2023. Moore's Law would have cut it in half.

$50$10$1$0.102023202420252026Price per million tokens (log)Actual PaceMoore's Law pace

§4.1Three forces, compounding.

  1. 1.

    Better chips.

    Each new GPU generation does more work per dollar than the one before it. NVIDIA's Blackwell delivers roughly 3× the performance-per-dollar of Hopper. Rubin, shipping now, delivers another 5 to 10× on top. Every model running on newer hardware gets cheaper to train and to serve.

  2. 2.

    Smaller models.

    Fewer parameters mean fewer calculations per query, which means less hardware, less electricity, and less time. The model size needed to clear a given capability threshold has fallen 142× in two years, and the cost savings pass through almost directly to price.

  3. 3.

    Better serving.

    Once a model is trained, the software around it keeps getting smarter. Techniques like quantization, speculative decoding, and continuous batching were barely deployed three years ago. Today they let the same model on the same chips serve 5 to 10× more queries per GPU, dropping the cost of each query in proportion.

Primary sources: Xiao, Cai, Zhao et al., “Densing Law of LLMs,” Nature Machine Intelligence (Nov 2025), DOI 10.1038/s42256-025-01137-0. Stanford HAI AI Index 2026, technical performance chapter. Ho et al., “Algorithmic Progress in Language Models,” arXiv:2403.05812 (Epoch AI, 2024). OpenAI, Anthropic, and Google published API pricing histories. Phi-3 Technical Report (Microsoft, 2024, arXiv:2404.14219). Sam Altman, “Three Observations” (Feb 9 2025). Inference cost figures refer to retail API pricing for a given capability tier, aggregated across major providers; underlying provider compute costs are estimated to be falling at comparable or faster rates.