AI industry faces the returns question as costs outpace performance gains

On September 20, TechCrunch reported, in an article relayed the same day by monitoring service Mezha.net, that the race for ever-larger AI models is now meeting a harder question than technical feasibility alone: can their growing price tags deliver equally powerful returns? The article documents growing doubt, among investors and labs alike, about the sustainability of the current spending trajectory.

Three structural drivers are cited for rising costs: the enormous computing power and significant electricity consumption required to train the most advanced systems, the high cost of dedicated computing infrastructure, and a persistent shortage of advanced semiconductors. Combined, these mean the cost of developing, training and maintaining a model is growing faster than the revenue it generates — an imbalance the article frames as the central question of the moment, distinct from the already well-covered debate over the sheer scale of capital committed.

For industrial leadership evaluating AI investment — in aerospace, energy, oil & gas or luxury — this shift in the debate, from 'how much to spend' to 'what return to expect', has an immediate practical consequence: economic due diligence on an AI project becomes as important as technical due diligence. Bounded, measurably-returning use cases (predictive maintenance, document automation, quality control) are structurally less exposed to this doubt than bets on still-emerging generative capabilities.

Cardan-AI analysis: this anticipated slowdown in the race for frontier models connects to an established economic framework on declining research productivity, which we develop in the companion analysis.