NVIDIA argued in a July 14, 2026 blog post that performance per watt is the ultimate metric for judging AI infrastructure efficiency. In the same post, NVIDIA stated that its Grace Blackwell GB300 NVL72 delivers up to 25 times the performance per watt of the previous-generation Hopper on DeepSeek V4 Pro, up to 20 times on GLM5.1, and up to 10 times on Kimi K2.6. The claim is that in an era where power has become the effective ceiling, the yardstick of competition is no longer a chip's peak speed but how much useful computation it extracts from the same electricity.

Framing Performance per Watt as a Metric That Cannot Be Gamed

NVIDIA framed performance per watt as a metric that cannot be gamed and can only be earned through real-world results. Catalog peak FLOPs or tokens per second can be inflated for specific conditions, but the actual work completed within a fixed power budget is a value the whole system produces together. NVIDIA called this metric the foundation of the "AI factory," and presented the GB300 NVL72 as reaching up to 25x performance per watt over Hopper on DeepSeek V4 Pro, up to 20x on GLM5.1, and up to 10x on Kimi K2.6. It also framed the generational shift as an expansion from an 8-GPU domain (Hopper) to a 72-GPU domain (Blackwell NVL72).

How to Read the 25x, 20x, and 10x Spread

The figures of 25x, 20x, and 10x across three models show plainly that the gain varies sharply by workload. The most notable point is not the size of the multiplier but that it differs from model to model. On the same hardware, a specific compute pattern like DeepSeek V4 Pro widens the gap to 25x, while Kimi K2.6 lands at 10x. This suggests that performance per watt is not a fixed property of the chip but a value that shifts with model architecture and inference method. Isolating the top "up to 25x" number and generalizing from it therefore risks overstating the benefit felt in a real deployment. Anyone evaluating adoption is safer re-measuring the multiplier on a workload close to the models they actually run.

The Point That Only 60 Percent of Power Becomes Useful Compute

NVIDIA stated that only about 60 percent of the electricity pulled from the grid is converted into useful AI computation. The remaining 40 percent or so scatters into cooling, power conversion, and idle time rather than compute itself, revealing that much of the room to raise performance per watt lies outside the chip. NVIDIA also presented DSX MaxLPS, which it said lets operators run up to 40 percent more GPUs within the same power budget. The implication is clear: the center of gravity in the performance race is shifting from building faster chips to converting secured power into computation without waste. Where power is a fixed cost and hard to secure in the first place, a few percentage points of conversion efficiency become throughput gained with no added facilities.

Why Software Alone Can Swing Results 5x in a Month

NVIDIA stated that on DeepSeek V4, software improvements alone raised performance per watt by up to 5x in a single month. Without changing hardware, optimizations in the software layer such as Dynamo, TensorRT LLM, NVFP4 quantization, and EAGLE3 speculative decoding sharply lifted throughput per unit of power. This shows performance per watt is not decided by hardware generation alone. On the same GPU, how the inference stack is assembled can widen effective efficiency several fold, signaling that software operations skill has become as important a variable in infrastructure investment as hardware choice. That said, a 5x gain in one month is likely the steep early phase of optimization, and whether improvement continues at the same pace is a separate question.

What It Means for Korean Data Centers and Practitioners

In Korea, where grid headroom is limited and competition for data-center sites and contracted power is fierce, performance per watt lands with particular weight. As the basis for adoption shifts from catalog peak compute to actual throughput per contracted watt, whoever handles more service on the same megawatt gains the edge in cost competition. The 60 percent conversion rate and the 5x software gain NVIDIA presented read, for domestic operators, as a signal that there is ample room to inspect cooling, power conversion, and the inference stack before replacing hardware. Still, these figures are all NVIDIA's own measurements carrying an "up to" qualifier. Whether a vendor's top numbers reproduce as-is in one's own environment requires independent verification.

Source: NVIDIA Blog — Why Performance per Watt Is the Ultimate Metric for AI Infrastructure Efficiency (2026-07-14)