When AI gets smarter, do companies use less of it? New research from the University of Chicago and Cursor suggests the opposite — and the implications for the economics of AI investment are profound.
A preprint by Suproteem Sarkar and Luke Melas-Kyriazi documents what they call a Jevons-like effect in AI demand. Studying firms on the Cursor coding platform following the release of Claude Opus 4.5 and GPT-5.2 in late 2025 — widely recognized as step-change improvements in model capability — they find that workers sent 44% more agent messages per week in the months that followed. Efficiency gains did not reduce consumption. They supercharged it.
This pattern echoes a well-established phenomenon from energy economics. When steam engines became more coal-efficient in the nineteenth century, total coal consumption rose sharply, because cheaper power unlocked entirely new categories of application — railways, textile mills, steamships. The same logic appears to be playing out in AI. Lower effective cost per task does not shrink the market; it expands the frontier of tasks worth attempting.
What makes this research particularly instructive for executives and investors is the granularity of the findings. The usage surge was not uniform. Smaller, newer, and privately held firms responded more aggressively than their larger, more established counterparts — a signal that organizational flexibility is itself a competitive asset in an AI-accelerating environment. Firms capable of restructuring workflows quickly capture disproportionate value from capability improvements.
The nature of the tasks being delegated to AI also evolved. The initial post-release surge concentrated on lower-complexity messages, suggesting workers first reached for familiar applications. Over subsequent months, however, usage migrated toward higher-complexity tasks — architecture decisions, deployment work, cross-system coordination — domains where human bottlenecks have historically been most expensive.
For model providers and infrastructure investors, this dynamic materially changes the return calculus. With global AI infrastructure spending projected to approach one trillion dollars by 2028, the central question has been whether downstream usage would justify that capital. This evidence suggests it will, provided capability improvements continue. Latent demand appears deep, and organizational adaptation appears faster than most forecasts have assumed.
The more consequential risk may not be oversupply of AI capability, but undersupply of organizational readiness to absorb it.
Source: Raw/trigger-returns-to-intelligence.md