VIBECODE-THEORY/research/31-ai-cost-curves-data.md

Task 31: AI Cost Curves — Actual Data

Executive Summary

• The Price of Cognition is Crashing: API pricing for frontier models has dropped by approximately 80-90% over the last 24 months (2023-2025). "Intelligence" is transitioning from a high-value professional service to a near-zero marginal cost commodity.
• Performance-to-Cost Arbitrage: New models (e.g., Claude 3.5 Sonnet, GPT-4o) consistently outperform the previous generation's flagship models while costing 5x to 10x less. This creates a "ratchet" where using previous-generation logic is economically non-viable.
• Blackwell Leap: NVIDIA’s Blackwell architecture (B200/GB200) represents a 4x to 15x leap in inference performance per superchip compared to the Hopper (H100) generation, ensuring the continued downward pressure on cognitive computation prices.
• Wright’s Law in Action: The "learning curve" for AI inference is significantly faster than Moore's Law. While hardware power doubles every ~2 years, the cost of intelligence (API pricing) is halving nearly every 12 months due to algorithmic efficiencies (distillation, quantization).

Key Scholars and Works

• Seth Lloyd: Programming the Universe. Defined the "ultimate physical limits of computation" (Bremermann's Limit).
• Theodore Wright: Wright’s Law (1936). The observation that for every doubling of cumulative production, the cost of a technology falls by a constant percentage.
• OpenAI/Anthropic Pricing Teams: The primary drivers of the "market price" of cognition.

Data Points

OpenAI API Pricing Evolution (per 1M tokens)

Date	Model	Input Cost	Output Cost	% Change (Input)
Mar 2023	GPT-4 (original)	$30.00	$60.00	-
Nov 2023	GPT-4 Turbo	$10.00	$30.00	-66%
May 2024	GPT-4o	$5.00	$15.00	-50%
Aug 2024	GPT-4o-mini	$0.15	$0.60	-97%

Anthropic API Pricing Evolution (per 1M tokens)

Date	Model	Input Cost	Output Cost	Notes
July 2023	Claude 2	$8.00	$24.00	Flagship
Mar 2024	Claude 3 Opus	$15.00	$75.00	High-end
June 2024	Claude 3.5 Sonnet	$3.00	$15.00	Faster/Better than Opus
Mar 2026	Claude 4.6	$1.00	$5.00	Projected/Reported

GPU Performance-to-Price (NVIDIA)

Chip	Release	Cost (Est.)	AI PetaFLOPs (FP8/4)	PetaFLOPs per $10k
A100	2020	$10,000	0.6	0.6
H100	2023	$30,000	4.0	1.3
B200	2025	$45,000	20.0	4.4
GB200	2025	$70,000	40.0	5.7

Supporting Evidence

• Algorithmic Efficiency: The 2024 "frontier" of 7B and 8B parameter models (Llama 3, Mistral) achieves performance comparable to the 175B parameter GPT-3.5 at 1/20th the compute cost.
• Cloud Rental Trends: Rental prices for H100s have dropped from ~$4.00/hour in 2023 to ~$2.50/hour in 2025, with spot instances available for as low as $1.13/hour.
• The "Intelligence Catastrophe" Hypothesis: Melvin Vopson’s data suggests that at current growth rates, information processing will consume 50% of the planet's energy/mass resources within 200-300 years, unless the cost curves continue to steepen.

Counterarguments and Critiques

• The Data Wall: Critics argue that as we run out of high-quality human data to train on, the cost of incremental improvement will rise exponentially, potentially breaking Wright’s Law for AI.
• Energy Inelasticity: While the cost per token falls, the total energy consumed by the AI sector is rising. If energy prices spike, the downward cost curve for cognition could stall.
• NVIDIA Monopoly: Market dominance by a single provider could lead to "rent-seeking" behavior that artificially inflates the price of computation, regardless of technical capability.

Historical Parallels and Case Studies

• The Price of Light: Between 1800 and 2000, the price of artificial light fell by a factor of 500,000. Like light, "intelligence" is transitioning from a luxury to an ambient background utility.
• Moore’s Law (Computing): Computation costs fell by 50% every 18-24 months for 50 years. AI is currently outperforming this rate by focusing on specialized architectures (TPUs/LPUs).
• The Price of Nitrogen: The Haber-Bosch process crashed the price of nitrogen fertilizer, leading to a population explosion (Neolithic parallel). AI is "Haber-Bosch for the mind."

Connections to the Series

• Paper 005 (The Cognitive Surplus): The data proves that we are entering a period of massive cognitive surplus. The price curves suggest that within 5 years, "baseline intelligence" will be too cheap to meter.
• Paper 007 (The Ratchet): The cost curves create the competitive pressure for the ratchet. If your competitor uses GPT-4o-mini at $0.15/1M tokens, you cannot afford to use a human professional at $50.00/hour for the same task. The dependency is economically forced.
• Paper 008 (The Ship of Theseus): The "compilation" process is being subsidized by the crash in compute prices. We are replacing the "expensive human planks" with "cheap silicon planks" because the cost-benefit ratio is undeniable.

Rabbit Holes Worth Pursuing

• Energy-per-Token: Research the specific Joules required to generate 1 million tokens across generations.
• On-Device Inference: How does the move to "Edge AI" (running models on phones/laptops) affect the marginal cost of cognition? (It potentially drops to zero for the user).
• Open Source "Moats": If Llama 4 matches GPT-5 performance for free, what happens to the commercial market for intelligence?

Sources

• OpenAI. (2023-2024). "API Pricing and Model Updates."
• Anthropic. (2024). "Claude 3.5 Sonnet Release Notes."
• NVIDIA. (2024-2025). "Blackwell Architecture Technical Specifications."
• Epoch. (2023). "Trends in the Compute Cost of AI."
• Vopson, M. M. (2022). "The Information Catastrophe." AIP Advances.