By DataCenterKnowledge
Publication Date: 2026-04-16 13:45:00
As generative AI workloads reshape data center economics, Nvidia is arguing that traditional metrics for evaluating infrastructure – including FLOPS per dollar and raw compute cost – no longer reflect how AI systems deliver business value.
In a blog post, the company said data centers are evolving from systems that process data into what it describes as “AI token factories,” where the primary output is tokens generated during inference.
That shift, Nvidia said, requires a corresponding change in how operators measure total cost of ownership.
From Compute Metrics to Output Economics
The company draws a distinction between three commonly cited metrics:
-
Compute cost – what customers pay for infrastructure.
-
FLOPS per dollar – theoretical compute efficiency.
-
Cost per token – the total cost to generate usable AI output.
Nvidia asserts that the first two are input metrics, while cost per token reflects actual business outcomes.
“Optimizing for inputs while the business runs on output is a fundamental mismatch,” the company said in the post.
The framing aligns with a broader industry shift toward inference-heavy workloads, where performance is increasingly measured by throughput, latency, and efficiency at scale rather than peak compute.
The Denominator Problem
At the center of Nvidia’s argument is a simple equation: cost per token depends not just on infrastructure cost, but on how many tokens a…