The electrical grid is designed to support loads that are relatively steady, such as lighting, household appliances, and industrial machines that operate at constant power. But data centers today, especially those running AI workloads, have changed the equation. Data centers consume a significant percentage of power plant and transformer capacity. Traditionally, diverse activities in the centers could average out consumption. Training large-scale AI models, however, causes sudden fluctuations in how much power is needed and poses unique challenges for grid operators:
If power demand suddenly ramps up, it can take one minute to 90 minutes for generation resources to respond because of physical limitations in their ramp rates.
Repeating power transients could cause resonance and stress equipment.
If the data center suddenly reduces its power consumption, the energy production systems find themselves with excess energy and no outlet.
These sudden changes can be felt by other grid customers as spikes or sags in supplied voltage.
In this blog, we'll detail how NVIDIA addresses this challenge through a new power supply unit (PSU) with energy storage in the GB300 NVL72. It can smooth power spikes from AI workloads and reduce peak grid demand by up to 30%. And it's also coming to GB200 NVL72 systems.
We will describe the different solutions for training workloads at the start, for running at full load, and for the end of the training run. Then we'll share measured results using this new power smoothing solution.
The impact of synchronized workloads In AI training, thousands of GPUs operate in lockstep and perform the same computation on different data. This synchronization results in power fluctuations at the grid level. Unlike traditional data center workloads, where uncorrelated tasks smooth out the load, AI workloads cause abrupt transitions between idle and high-power states, as shown in Figure 1.
data-src=https://developer-blogs.nvidia.com/wp-content/uploads/2025/07/image3-4-png.webp alt=A chart showing how thousands of GPUs processing an AI workload consume power simultaneously and synchronously shift between high and lower power states. Time is depicted on the x-axis and power on the y-axis. The line chart shows the power profile wave form, with a fast ramp-up in power at the workload start, followed by many successive sharp up-and-down power cycles until there is a fast ramp-down at workload end. class=lazyload wp-image-103699 data-srcset=https://developer-blogs.nvidia.com/wp-content/uploads/2025/07/image3-4-png.webp 1999w, https://developer-blogs.nvidia.com/wp-content/uploads/2025/07/image3-4-300x122-png.webp 300w, https://developer-blogs.nvidia.com/wp-content/uploads/2025/07/image3-4-625x254-png.webp 625w, https://developer-blogs.nvidia.com/wp-content/uploads/2025/07/image3-4-179x73-png.webp 179w, https://developer-blogs.nvidia.com/wp-content/uploads/2025/07/image3-4-768x312-png.webp 768w, https://developer-blogs.nvidia.com/wp-content/uploads/2025/07/image3-4-1536x623-png.webp 1536w, https://developer-blogs.nvidia.com/wp-content/uploads/2025/07/image3-4-645x262-png.webp 645w, https://developer-blogs.nvidia.com/wp-content/uploads/2025/07/image3-4-500x203-png.webp 500w, https://developer-blogs.nvidia.com/wp-content/uploads/2025/07/image3-4-160x65-png.webp 160w, https://developer-blogs.nvidia.com/wp-content/uploads/2025/07/image3-4-362x147-png.webp 362w, https://developer-blogs.nvidia.com/wp-content/uploads/2025/07/image3-4-271x110-png.webp 271w, https://developer-blogs.nvidia.com/wp-content/uploads/2025/07/image3-4-1024x415-png.webp 1024w, https://developer-blogs.nvidia.com/wp-content/uploads/2025/07/image3-4-960x389-png.webp 960w data-sizes=(max-width: 1999px) 100vw, 1999px />Figure 1. Thousands of GPUs processing an AI workload consume power simultaneously, and synchronously shift between higher and lower power states. After the workload completes, the GPUs idle simultaneously, too.
Visualizing the individual GPUs as rows on a heatmap illustrates why AI data centers pose unique power challenges to the power delivery grid. (See Figure 2 below.) Traditional data center workloads operate asynchronously across the compute infrastructure. The AI training workload heatmap highlights how GPUs operate synchronously, causing the total power drawn by a GPU cluster to mirror and amplify the power pattern of a single node.
data-src=https://developer-blogs.nvidia.com/wp-content/uploads/2025/07/image1-2-png.webp alt=Two charts showing a GPU power consumption heat map. Traditional data center workloads show a random heat map pattern of power consumption over time across a GPU cluster. AI training workloads show color bars as all GPUs enter the same power consumption state across the cluster, ramping up and down as a whole. class=lazyload wp-image-103700 data-srcset=https://developer-blogs.nvidia.com/wp-content/uploads/2025/07/image1-2-png.webp 967w, https://developer-blogs.nvidia.com/wp-content/uploads/2025/07/image1-2-300x269-png.webp 300w, https://developer-blogs.nvidia.com/wp-content/uploads/2025/07/image1-2-625x560-png.webp 625w, https://developer-blogs.nvidia.com/wp-content/uploads/2025/07/image1-2-128x115-png.webp 128w, https://developer-blogs.nvidia.com/wp-content/uploads/2025/07/image1-2-768x689-png.webp 768w, https://developer-blogs.nvidia.com/wp-content/uploads/2025/07/image1-2-645x578-png.webp 645w, https://developer-blogs.nvidia.com/wp-content/uploads/2025/07/image1-2-335x300-png.webp 335w, https://developer-blogs.nvidia.com/wp-content/uploads/2025/07/image1-2-100x90-png.webp 100w, https://developer-blogs.nvidia.com/wp-content/uploads/2025/07/image1-2-362x325-png.webp 362w, https://developer-blogs.nvidia.com/wp-content/uploads/2025/07/image1-2-123x110-png.webp 123w, https://developer-blogs.nvidia.com/wp-content/uploads/2025/07/image1-2-602x540-png.webp 602w dat
