On October 26, 2026, a16z said OpenRouter token demand had climbed about 27x since last November, while Nvidia GPU rentals were still running above a standard six-year depreciation baseline.
In a16z’s State of AI, weekly OpenRouter traffic was cited at roughly 165 trillion tokens, up from about 6 trillion, with open-weight models now handling about 72% of tokens versus roughly 28% last November.
a16z also showed rental premiums above that baseline across several GPUs, including H200 at +143%. As XenoSpectrum’s analysis notes, that H200 figure is not the same as a year-to-date price jump. It reflects the gap between current rental pricing and a straight-line depreciation model anchored to early observed rates.
Why it matters: teams chasing lower inference costs may still run into expensive accelerator supply. Falling per-token prices and rising model usage do not guarantee cheaper serving if GPU rentals stay elevated.