Layer 1 · The numbers that decide
Tokens per watt
The metric that actually matters in 2026: useful model output per unit of power. It joins the chip layer to the energy layer, which is why both are on this site.
The number
- What it is
- The metric that actually matters in 2026: useful model output per unit of power. It joins the chip layer to the energy layer, which is why both are on this site.
- Typical values
- Vendor-published only, and never on a common benchmark
Verified 2026-09-09.
The catch
There is no neutral, standardised tokens-per-watt benchmark. Every published figure comes from a company selling one side of it. Treat all of them as marketing until an independent body publishes a method.
What this is powering
A GB200 NVL72 rack draws 120–132 kW and a GB300 pushes 135–200 kW, against a 2026 average rack of about 27 kW. The silicon decision and the power decision are the same decision, made eighteen months apart.
Layer 2 — Silicon · GB200 NVL72 · The 150 W inference option
Others in the numbers that decide
Sources
Want this whole layer as one table?
The power and cooling options on these pages as a single CSV — lead time, capex, carbon, maturity, density supported, PUE, the catch and what each one decides — ready to sort, filter or load into a notebook.
Get the energy table — free for research and evaluation
Four short lines and the download starts at once (CC-BY-NC-4.0, built from these pages the moment you ask). What you type is used only to answer the request and is never sold. Prefer Hugging Face? Request it there.
Using it inside a product, a client report or anything you sell? That needs a commercial licence.