Google and NVIDIA to Build AIDC That Operates According to Power Supply [Global AI Briefing]

[Financial News] Global artificial intelligence (AI) companies such as Google and NVIDIA have decided to develop AIDC operational technology that flexibly adjusts operations according to power supply conditions and establish global standards to resolve the issue of securing power, which is emerging as a bottleneck in the expansion of AI data centers (AIDCs).
The proposal is to operate AIDCs not as electricity-guzzling monsters at every moment, but as facilities that automatically regulate power usage according to grid conditions.
On the 16th (local time), Google, NVIDIA, and AI data center power management company Emerald AI officially announced the launch of the AI Energy Management Alliance (AEMA), which includes 18 companies and organizations in the AI, data center, and power sectors, such as National Grid plc, AES, and NRG Energy.
AEMA explained that it is developing technology to transform AIDC operating methods into power-flexible systems and establish standards.
Until now, AIDCs have been large power consumers that used electricity stably 24 hours a day, but AEMA's approach is to have AIDCs automatically increase AI computations during times or in regions with surplus power supply and slow down low-priority training or computations when power demand surges, thereby reducing power consumption. The plan also envisions reducing the power drawn from the grid by utilizing batteries or self-generation.
NVIDIA explained that it is already applying power-flexible GPU operating methods to actual AIDCs.

The method works by having software automatically slow down low-priority tasks while continuing core AI inference work when a power company sends a signal to reduce power usage. NVIDIA stated that in a real-world demonstration, it automatically reduced power usage from 4 MW to 3 MW while maintaining high-priority tasks and responded to more than 200 power adjustment signals each time.
AEMA plans to establish standards to measure how quickly and by how much AID can actually reduce power consumption. It added that it will pursue policies to create common standards for how quickly and reliably AID can regulate power consumption and to allow rapid grid connections for facilities that demonstrate this capability.
AEMA is understood as a strategy to address the bottleneck in building AI infrastructure, where securing power is hindering development—not only because difficulties in securing power grids are delaying AIDC construction in the United States and around the world, but also because opposition from local residents is escalating into political issues.
In the United States, there are cases in which large-scale AIDCs wait more than five to ten years to be connected to the power grid. This is because power companies must secure power plants and transmission networks in advance to ensure they can supply electricity even when data centers demand maximum power.
On the other hand, since the actual power grid is not used at maximum capacity 24 hours a day throughout the year, adjusting AIDC power usage to match demand could reduce the burden of building new power plants or substations for AIDC construction.
AEMA analyzed that the average utilization rate of the U.S. power grid is approximately 50%.
Industry experts predict that future competition in the AIDC industry will be measured not only by simple capacity expansion but also by securing operational technologies such as power flexibility.
 
[email protected] Lee Gu-sun Reporter