Silicon Data's LLM Token Spending Index fell to $0.97 per million tokens on August 31, a drop of more than half from its peak of approximately $2.05 at the end of May. This index is weighted by actual usage; the decline stems from both model price reductions and companies shifting to cheaper open-weighted and domestically developed models, and does not represent a general 50% price drop for the same model.
The price war initially compresses the unit revenue and pricing power of leading model vendors, but it doesn't necessarily negate the return on AI investment: lower prices may also stimulate increased usage by agents and enterprise applications. The dividing line is whether token demand can grow faster than the price decline;
If not, model vendor revenue and computing power utilization will both be under pressure, and the market will need to reassess the return cycle of the trillions of dollars invested in AI infrastructure.
