Are Tokens the Must-Have Measure for Data Center Output?

Common tools used to measure data center output are power-based (Power Usage Effectiveness), infrastructure-specific (Compute Cost), or efficiency-based (FLOPs per Dollar).

Recently, NVIDIA made the case for rethinking how we measure artificial intelligence data center performance, moving from a power- or infrastructure-focused mindset to one that assesses how well AI-specific data centers actually get the job done.

Enter the cost per token metric.

NVIDIA argued that cost per token is the only metric needed to analyze AI output. However, Cushman & Wakefield isn’t so sure. A recent DC Insights Brief explained that tokens should be considered a complement to, rather than a replacement for, power use metrics.

According to a recent Cushman & Wakefield Insight Brief, most AI-focused data centers are using tokens as their primary measure, rather than megawatts or GPU hours. However, the insight also points out that token-based billing can have its issues.

What They Are

Tokens are a recent topic of discussion among data center and AI experts, first appearing on the internet in late 2025. Breaking it down, data center tokens use data units processed by AI models (think words, pixels or audio snippets) to measure performance and efficiency.

The argument for token use is that metrics like PUE are great for measuring energy efficiency and usage. However, they don’t measure how that energy translates into productive AI output.

The Cushman & Wakefield analysts agreed with the token concept, saying that the cost per token, which measures the type and amount of data output, can be a more efficient way to quantify AI data center usage.

“Cost per token reflects more accurately for AI output, while price per kilowatt is detached from it,” the brief said, explaining why newer AI platforms are shifting toward token-level metering and billing.

The Caveats

As Nvidia and others push the “tokens for all” concept, the Cushman & Wakefield analysts are more hesitant, suggesting that “shifting fully to cost per token is not straightforward for all data centers.”

One reason is a lack of consistency. Similar power use leads to different levels of token output, providing a “less consistent basis for benchmarking across different setups and infrastructure,” the brief commented.

There are also potential cost spikes, especially if usage is poorly managed. “Cost per token aligns pricing with consumption, but depends on system efficiency,” the Cushman & Wakefield analysts explained.

Cushman & Wakefield recommends that hyperscalers and occupiers continue using power usage metrics alongside newer cost-per-token metrics to maintain a competitive advantage.

“Operators can differentiate on token efficiency per megawatt, maximizing output, rather than simply scaling capacity,” the analysts pointed out. Going this route “requires higher integration across compute, networking and orchestration, and higher utilization discipline,” they added.