Falling token prices + sticky-high GPU prices is classic Jevons paradox playing out in real time.

Inference is getting dramatically cheaper (efficiency, competition, Blackwell-scale gains). That doesn’t reduce compute demand - it explodes it. Cheaper tokens unlock agentic workflows, more coding, more automation and more users. Total token volume is growing far faster than unit costs are falling, so the scarce resource (GPUs, HBM, power, data centres) stays tight.

Implications for AI infrastructure stocks:

- GPU / accelerator makers (NVDA and peers) still benefit. High utilisation + sustained scarcity supports pricing and volume.
- Memory (HBM/GDDR), networking and power infrastructure see durable demand as inference scales.
- Hyperscalers and neoclouds face a margin squeeze if token price declines outrun volume growth, but capacity constraints and rising absolute spend so far argue demand is still winning.
- The shift to inference-heavy workloads favours specialised silicon and high-throughput systems over pure training clusters.

Bottom line: AI is proliferating faster than efficiency is reducing the need for hardware. Unit prices collapse, total spend and infrastructure demand keep rising. That’s bullish for the physical layer as long as volume growth remains the dominant force. alphatarget.com