AI Tokens Get Cheaper, Challenging Trillion-Dollar Boom Thesis

The prevailing economic thesis underpinning the trillion-dollar artificial intelligence boom is facing a significant challenge as the cost of AI tokens, a key metric for compute usage, is rapidly decreasing. This trend suggests that demand for AI compute may not be as infinitely elastic as previously assumed, potentially impacting the high valuations of AI companies. New data published on Wednesday by Ramp, a corporate spending platform, indicates that American businesses are now paying approximately 41% less per million tokens compared to their peak in March. The effective price has fallen from $1.15 to 68 cents per million tokens. This decline in token cost is occurring despite ongoing advancements in AI model capabilities and increasing adoption. Further evidence of this shift comes from the changing usage patterns of AI models. The proportion of AI compute dedicated to frontier models, which are the most advanced and computationally intensive, has dropped from about 53% in early August to 45% by September. This indicates a move towards less resource-intensive models or a broader distribution of usage across different tiers of AI capabilities. The impact of this trend is particularly noticeable among the largest AI spenders. The top 1% of companies, which account for approximately 80% of the enterprise revenue for leading AI labs like OpenAI and Anthropic, reduced their per-employee AI spending by nearly 10% in August. This suggests that even the heaviest users are finding ways to optimize their AI expenditure or are experiencing diminishing returns from further scaling of their current AI deployments. Ara Khazarian, the chief economist at Ramp and head of the Ramp Index, described this situation as a "crack in the AI thesis" in comments to Fortune. He explained that instead of driving a surge in demand for the most advanced models, tokens are increasingly being priced like commodities. This commoditization implies that AI compute is becoming more standardized and less differentiated, a characteristic typically associated with lower valuations compared to high-growth, unique technology markets. The traditional AI economic model relies on the premise that increasing demand for AI compute, driven by more complex models and wider adoption, would allow AI labs to capture surplus value. This surplus was intended to cover their substantial investments and debts. However, the falling token prices and shifting usage patterns question the sustainability of this model and the justification for current market valuations, which often project future earnings based on decades of anticipated growth and premium pricing for AI services. The situation implies that the AI market may be maturing faster than anticipated, moving from a speculative growth phase towards a more cost-sensitive and competitive landscape where efficiency and price become more critical factors for sustained success.