SC26 week produced three announcements that look unrelated: a benchmark debut, a network upgrade and an energy study. Read together, they share one denominator. GPU gains are now being reported in tokens per watt, operators are reclaiming watts from the network, and researchers are tallying what the remaining watts cost the grid. The industry's scoreboards are quietly unifying on power.
The Benchmark Moved to Watts
In its MLPerf debut, NVIDIA swept every MLPerf Training v6.1 benchmark and, in its first MLPerf Endpoints submission, reported more than 10x generational gains in throughput per watt for agentic AI — on a provisional Kimi K3 result, Vera Rubin NVL72 delivered more than 10x the throughput per watt of GB300 NVL72 at the same response speed S1. Training was strong too: up to 2x more performance per GPU than GB300 NVL72, and about 97% scaling efficiency from 256 to 1,024 GPUs on DeepSeek-V3.
The interesting part is not the sweep — vendor wins on home turf are expected — but what MLPerf Endpoints chooses to measure: whether agents complete real coding and enterprise tasks correctly, and how much work lands per watt at a fixed interactivity target. That is essentially the metric buyers of AI capacity purchase. One caveat before quoting these numbers: the Kimi K3 comparison is provisional pending MLCommons verification, while the GB300 baseline results are verified.
The Network Wants Its Watts Back
Lambda's answer to power pressure comes from the switch. The cloud provider is running NVIDIA Photonics co-packaged optics to connect more than 10,000 GB300 GPUs, reporting roughly 4% lower network latency at rack scale and up to 10% at pod scale at unchanged throughput, alongside up to 5x better power efficiency and 15x higher reliability than conventional pluggable optics S2. The optics run cooler — a 36°C median versus 51°C — and light sources are field-replaceable modules rather than whole transceivers.
The engineering logic is straightforward: every watt spent moving bits is a watt unavailable for compute, and in a large training job a single interrupted link can stall thousands of GPUs. Co-packaging turns networking from a power and failure tax into reclaimed headroom — and increasingly, a procurement line item worth asking about in cluster RFPs.
The Meter Is Already Running
The constraint behind both moves is now quantified. A new Nature Sustainability study identified 403 US hyperscale data centers that likely consumed about 82 terawatt-hours between May 2024 and April 2025 — roughly 1.8% of all US electricity — producing 25.7 million metric tons of CO2, comparable to Denmark's annual footprint, with around 54% of that power coming from gas and coal S3.
The study's sharpest finding for builders is geographic: the same electricity in coal-reliant Wyoming generates roughly seven times more carbon per kilowatt-hour than in hydropower-rich Oregon. Per-generation efficiency gains are real, but siting and grid mix still dominate the carbon math. Tokens per watt is the right scoreboard — pair it with tokens per gram of CO2 when you pick the site.
Sources and editorial method
This AI-assisted report synthesizes the sources below. Automated source checking is not independent human verification; follow the original reporting for context and updates.
Common questions
What does tokens per watt measure in AI infrastructure?
The amount of agentic AI work a system completes per unit of electrical power; the new MLPerf Endpoints benchmark scores it at a fixed response speed.
How much electricity do US hyperscale data centers use?
A Nature Sustainability study estimates 403 US hyperscale facilities consumed about 82 TWh from May 2024 to April 2025, roughly 1.8% of US electricity.


