AI costs outpace Moore's Law, threatening industry profits

Summary

Epoch AI data shows that the cost of a fixed level of AI capability has dropped by approximately 47% per quarter (a 13x reduction annually). This decline outpaces historical price drops in DNA sequencing, lithium batteries, and computing power, creating deflationary pressure on the revenue models of AI labs.

Why it matters

If the 'cost of thought' continues to fall at this rate, the hundreds of billions of dollars in infrastructure investments supporting the capital expenditures of OpenAI, Anthropic, and hyperscalers will face a fundamental return on investment (ROI) problem. Products trained at peak costs become commodities by the time they are priced. This dynamic is reshaping not only the AI economy but also the broader technology investment thesis that has driven capital markets since 2023.

Highlights

• The cost per question on GPQA Diamond fell from 30 cents (o3, January 2025) to 0.04 cents (GPT-5.6 Luna, ~mid-2026) — a 725x drop in 18 months
• Silicon Data’s token index reached 97 cents per million tokens on September 1, 2026; Chinese open-source models and strategic price cuts are accelerating deflation
• OpenAI still lacks a proven viable business model nearly four years after ChatGPT’s launch and has postponed its IPO

Critical take

The article relies almost entirely on Epoch AI’s 'cost of thought' metric. This metric structurally measures the cost required to achieve a fixed benchmark score. As models improve, this number mechanically decreases, meaning the dramatic 13x annual figure partially reflects the measurement methodology rather than a unique economic law. Additionally, while the text frames deflation as a pure threat to labs, it fails to adequately examine who actually captures the value—application developers and end users. It also provides no counter-evidence or alternative data sources to test its headline claim.