Breaking

NVIDIA Says GB300 NVL72 Delivers Up to 25x Better Performance Per Watt Than Hopper on Leading AI Models

July 14, 2026 at 12:00 EDT

Infra & Chips
Foundation Models

NVIDIA is positioning performance per watt as the defining competitive metric for AI data centers, arguing that its GB300 NVL72 system delivers up to 25 times more work per watt than the prior Hopper generation on some of today's most demanding models. The company framed the metric as the foundational yardstick for any "AI factory" running under tight power constraints.

Performance Per Watt · NVIDIA

The new yardstick for AI factories is work per watt — not raw compute

NVIDIA's GB300 NVL72 rack claims up to 25× more work per watt than the Hopper generation on today's most demanding models — reframing AI infrastructure around energy efficiency in a power-constrained era.

25×

perf/watt vs Hopper on DeepSeek V4 Pro

20×

perf/watt on GLM 5.1

10×

perf/watt on Kimi K2.6

GB300 NVL72 vs Hopper — per-model gains

Performance per watt (Hopper baseline = 1×), by workload

1×

Hopper
baseline

10×

Kimi
K2.6

20×

GLM
5.1

25×

DeepSeek
V4 Pro

Inside the rack

GPUs linked over NVLink

NVFP4

low-precision compute

1.1–1.25

target PUE, liquid-cooled

135–150 kW

per-rack power draw

Why it resonates

With data centers hitting hard electricity limits, tokens-per-watt maps directly to operating cost and margin. Lower energy per token is broadly welcomed as the right frame for a power-constrained era.

The caveats

Independent verification of the headline figures is limited and gains swing with model and workload. Racks drawing 135–150 kW demand power delivery and cooling many facilities can't yet handle.

The efficiency trajectory

Hopper

baseline

→

Blackwell

GB300 NVL72

→

Vera Rubin

up to +10× more

~1,000,000× cumulative inference throughput per megawatt claimed across these three generations.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.