Data center performance has traditionally been evaluated through power and infrastructure metrics such as Power Usage Effectiveness, compute cost, or measures like FLOPs per dollar. These benchmarks focus on how efficiently facilities consume energy and deploy hardware, rather than how effectively they convert that capacity into usable artificial intelligence output.
NVIDIA has recently urged the industry to rethink this approach for AI-specific data centers, advocating a shift away from power- and infrastructure-only views toward measures that capture how well facilities deliver AI work. At the center of this push is the cost per token metric, which NVIDIA has positioned as the single metric needed to evaluate AI data center output.
Cushman & Wakefield takes a more cautious stance. In a recent DC Insights Brief, the firm agreed that tokens are an important tool but argued they should supplement, not replace, established power-based metrics. The analysts noted that many AI-focused data centers are already using tokens as their primary performance and billing measure, moving beyond megawatts or GPU hours as the main yardstick.
Tokens, which first surfaced in industry discussions in late 2025, represent units of data processed by AI models, such as words, pixels, or audio snippets. Proponents say traditional indicators like PUE remain useful for tracking energy efficiency but do not show how that energy translates into productive AI output. By contrast, cost per token ties the type and volume of AI data processed directly to pricing and usage, offering a more output-oriented view of performance.
Cushman & Wakefield’s analysts supported the concept, describing cost per token as a more accurate reflection of AI output than price per kilowatt, which they called detached from actual AI performance. This helps explain why newer AI platforms are gravitating toward token-level metering and billing structures.
However, the Insight Brief also highlighted important caveats. Because similar power consumption can yield very different token output across systems, tokens can provide a less consistent basis for benchmarking performance between different data center configurations. The analysts also warned of potential cost spikes under token-based billing, especially where usage is not tightly managed and where overall system efficiency is weaker.
For now, Cushman & Wakefield recommends that hyperscalers and occupiers continue to rely on power usage metrics alongside cost-per-token measures. The firm suggests operators can differentiate themselves by maximizing token efficiency per megawatt, focusing on how much AI output they generate from a given power envelope rather than only scaling capacity. Achieving this, the analysts added, requires deeper integration across compute, networking, and orchestration layers, along with disciplined utilization to control both performance and cost.
