
**Source:** [X/Twitter post by @demian_ai](https://x.com/demian_ai/status/2052395660108759174)
**Author:** Dylan (@demian_ai) - GTM for Nebius Token Factory
**Date:** 2026-05-07
**Tags:** #AI #inference #Jevons-paradox #compute #infrastructure

---

## Key Thesis

Inference got a hundred times cheaper this year. The compute bill went up anyway. This is the Jevons paradox playing out at trillion-dollar scale.

## Summary

- 12 months ago, 1M tokens of frontier-class reasoning cost ~$60. Today, equivalent quality costs ~$0.50.
- Price/token of o1-level intelligence dropped ~128x in a year. GPT-4-level output dropped ~100x since original GPT-4.
- Despite this, total compute bills at every hyperscaler are going up, not down.
- Anthropic signed multi-year capacity deals with XAI and Amazon. Microsoft Azure capex for 2026 starts with an eight. Nvidia paid ~$20B to acquire Groq.

## Why Costs Rose Despite Cheaper Tokens

- Reasoning models burn ~10x the output tokens of non-reasoning models (thinking out loud).
- Agentic workflows chain ~20x the requests of a single-shot completion (loops, tools, retries).
- A deep-research query costs more compute than 10 original GPT-4 queries combined.
- Net effect: tokens got 100x cheaper, but products now consume 10,000x more tokens = 100x larger total bill.

## The Jevons Paradox (1865)

Jevons noticed that making coal-burning more efficient did not reduce coal consumption -- it increased it, because efficiency unlocked previously uneconomic uses. The same dynamic is playing out with AI compute, faster than any analogous historical cycle.

## Infrastructure Implications

**Hyperscalers:** 2026 capex is a step on a curve, not a peak. Inference is always-on 24/7 (unlike bursty training).

**AI software companies on closed APIs:** Gross margins worsen as customers get more value, because usage = compute cost. Vertical integration is the winning strategy.

**KV Cache as the silent monster:** Running memory state of long conversations/agent loops scales with context length and agent steps. A long agent session can hold tens of GB of state per user, per session.

**CPU-to-GPU ratio shifting:** Training is 1:8. Basic chat inference is 1:4. Agentic inference is 1:1, sometimes CPU-heavy.

## Key Quote

> "The right framing is that AI got dramatically cheaper, dramatically more capable, and dramatically more useful, and the cost of running it at the new equilibrium of demand is much higher than the cost at the old equilibrium of demand, because the new equilibrium is enormous."

## Bottom Line

Cheaper tokens. More tokens. Same coal as 1865.
