Nvidia ties AI factory economics to tokens and power efficiency

Nvidia's Ian Buck said AI data center economics now hinge on token output per watt rather than GPU access alone. Each GPU generation improves efficiency roughly 10x, with Blackwell delivering a 30x gain in tokens per watt.

• Nvidia: each GPU generation boosts tokens per watt up to 10x
• Blackwell delivered a 30x improvement in tokens per watt
• Inference does not replace training as models are continually refined
• Groq 3 LPX inference accelerator works with the Vera Rubin platform