The tech world’s favorite defense for infinite AI scaling—Jevons paradox—is facing intense scrutiny from engineer and analyst Bookworm Eng. In a viral thread that gained traction on Hacker News, the author argues that the sheer volume of compute infrastructure coming online by 2027 will likely outpace demand absorption, leading to a significant overbuilding phase rather than the efficiency gains proponents predict.

The Power Supply vs. Revenue Gap

Central to this argument is data from SemiAnalysis, which projects 50GW of net new AI power availability by the end of 2027. When combined with existing capacity, the total cumulative AI power reaches 91.6GW. The analyst calculates that to monetize this capacity at DeepSeek V4.1 Flash inference rates—already validated by vLLM and SGLang teams—the market would need to generate $4.5 trillion in annual revenue. This assumes 50% of all compute is dedicated to inference, a high bar for any emerging technology stack.

Why Jevons Might Fail Here

Jevons paradox suggests that increased efficiency leads to increased consumption. However, the author posits that specific economic conditions must be met to avoid a glut: token prices would need to collapse by 20x, or new application classes like real-time voice and robotics would need to absorb the slack. With major labs like Anthropic and OpenAI currently focused on high-margin enterprise and consumer chat rather than these new verticals, the probability of full capacity absorption in the next 15 months is deemed low.

The Training Cost Reality Check

The analysis also scrutinizes the training side of the equation. It notes that Anthropic has spent under $16 billion on training and research since its inception, a sum sufficient to produce their current model lineages. The idea that the entire industry could sustain $459 billion in annual training costs by 2027 is viewed as implausible. Instead, the author suggests that excess capacity will likely be diverted to free services like Google Search AI features and Meta’s Muse, effectively subsidizing the overbuild through ad-supported models rather than direct API revenue.

Key Takeaways

  • $4.5T Revenue Requirement: To fully utilize projected 2027 AI power capacity at efficient inference rates, the industry would need $4.5T in annual revenue.
  • Price Collapse Risk: Absorbing the supply might require token prices to drop 20x, forcing labs to cut margins from 70-90% to below 30%.
  • Stranded Assets: With power continuing to come online in 2028-2030 and chips having long lifespans, valuations built on perpetual $1T+ spending are at risk.
  • Hyperscaler Pressure: To survive, hyperscalers will likely push for custom ASICs to bypass the 'Nvidia tax' and lower per-GW rental costs.

The Bottom Line

The reliance on Jevons paradox to justify current AI infrastructure spending ignores the hard math of revenue absorption. Unless token prices collapse drastically or entirely new consumer verticals emerge immediately, the industry is heading toward a massive overbuild where compute capacity far outstrips paying demand.