Nvidia Says AI Tokens Turn Profitable as GB300 Cuts Cost per Token 60%
Updated
Updated · Benzinga · May 26
Nvidia Says AI Tokens Turn Profitable as GB300 Cuts Cost per Token 60%
3 articles · Updated · Benzinga · May 26
Jensen Huang told Nvidia’s earnings call that AI-generated tokens have shifted from a cost center to a profitable product, calling the change a turning point for the industry.
Agentic AI is driving that shift, Huang said, because models can now perform productive work that customers will pay for rather than simply generate expensive output.
Colette Kress said Nvidia’s GB300 platform lowers cost per token by 60% versus systems available six months ago, improving economics for model developers and cloud providers.
That combination of lower inference costs and rising demand supports Nvidia’s view that AI capacity expansion is accelerating and could keep spending on data centers and chips growing toward trillions of dollars annually.
If AI labs lose billions on inference, who truly profits from the 'token economy' besides the chipmakers?
As AI agents move to desktops, is this the end of the cloud's dominance in artificial intelligence?
With AI causing an energy crisis, can new infrastructure be built fast enough to prevent widespread power shortages?
Nvidia GB300 NVL72 and the Rise of Cost Per Token as AI’s Defining Metric
Overview
Nvidia's announcement of the GB300 NVL72 platform at COMPUTEX 2026 marks a turning point for artificial intelligence, introducing a new era of profitable AI. Building on the success of previous Grace Blackwell systems and NVLink 72 scale-up switches, the Vera Rubin platform—now in full-scale production—promises to dramatically reshape AI economics. Supported by a robust global supply chain, this breakthrough enables AI to shift from a cost center to a profit driver, setting the stage for more accessible and scalable AI-powered services across the industry.