In a recent research report, Dario Perkins, an economist at research and consulting firm TS Lombard, pointed out that the AI industry is facing a divergence between continuously declining token sales prices and persistently high GPU rental prices. Jevons data shows that since 2026, the token price expenditure index has declined significantly, while the H100 GPU hourly rental rate has remained resilient, reflecting both intensified price competition in model services and a tight supply of computing resources.
Perkins noted that the unit token computing cost for model developers depends on the hourly GPU rental rate and the hourly token production. As model and chip efficiency improves, unit production costs are declining rapidly, but the widespread adoption of open-weighted models is also depressing token prices.
Perkins believes that the rate of decline in token prices and production costs will affect the profitability of model developers.
Meanwhile, data center operators rely on GPU rental revenue to cover fixed costs such as equipment procurement, debt, and long-term take-or-pay contracts. If future increases in computing power supply alleviate the current shortage, GPU rental prices may also face downward pressure, further adjusting the profit distribution across the AI industry chain.
