NVIDIA Pushes 'Cost Per Token' as the New Metric for AI Inference, Citing 35x Savings on Blackwell

NVIDIA is repositioning the economics of running AI, arguing that as companies move from pilots to production the decisive metric is no longer peak chip specifications but cost per token — how many useful tokens a system can deliver per dollar, per watt and within latency targets. In a technical push detailed on its corporate blog, the company said its full-stack inference software, co-designed with hardware on the Blackwell platform, is continually driving that cost down.