Nvidia Blackwell Cuts AI Token Costs 35-Fold as Providers Ready Price Slashes
Updated
Updated · Business Insider · Jun 12
Nvidia Blackwell Cuts AI Token Costs 35-Fold as Providers Ready Price Slashes
3 articles · Updated · Business Insider · Jun 12
Summary
AI token prices are poised to fall sharply later in 2026 as new models run on Nvidia Blackwell systems, with some model providers already cutting prices and others signaling cheaper offerings.
SemiAnalysis found Nvidia's GB300 NVL72 generated 6,000 tokens per second per GPU versus 90 on Hopper HGX 200, while cost per 1 million tokens dropped to 12 cents from $4.20.
Power efficiency also jumped: Blackwell produced 2.8 million tokens per second per megawatt versus 54,000 for Hopper, helping offset rising electricity costs in AI data centers.
Silicon Data's token spending index peaked near 2.06 in late May and fell to 1.75 by June 10, a move its CEO said could already signal broad token price declines.
Blackwell deployments, delayed by water-cooling and other data-center requirements, are now ramping in volume and should operate at scale in the second half, expanding cheap token supply.