◎ Lin Renbin
As AI data centers rapidly increase power demand, the real question worth noting is not just "how much more power generation is needed" but also "can large power users become more flexible on their own?"
As AI data centers rapidly increase power demand, can large power users become more flexible? The picture shows Datong AI modular data center.
The United States is putting another idea into practice. On September 16, Google, NVIDIA and Emerald AI jointly launched the "AI Energy Management Alliance" (AEMA), focusing on the power consumption flexibility of AI data centers: in addition to increasing power generation and expanding the power grid, it also allows data centers to proactively adjust power consumption when the power grid is tight.
The concept is actually very simple. Since AI is a huge electricity user, it can not only regard itself as a "load", but can also become a "grid resource". For example, non-real-time tasks such as AI training and batch calculations can be moved to other times; batteries can support part of the demand during peak periods; or real-time dependence on the public power grid can be reduced through local and co-located power supplies. What really matters is not which technology is used in the data center, but whether it can provide the flexibility that the power grid needs in a measurable and verifiable way.
This is not a simple concept. A Duke University study pointed out that within the 22 major power balancing zones in the United States, covering about 95% of the U.S. load, the existing system may accommodate an additional 76 to 126 GW of new loads, provided that new large loads have limited flexibility to be curtailed in a short period of time. The focus of the research is not to require the data center to shut down for a long time, but to carry out limited reduction or transfer during the few hours when the power grid is most stressed.
Google, NVIDIA and Emerald AI jointly launched the "AI Energy Management Alliance" to focus on the power consumption flexibility of AI data centers. Pictured is NVIDIA.
What is truly noteworthy is not the number “100 GW”, but that it changes the way we look at large loads: large power users do not necessarily have to be supplied but can also be managed.
Google also announced in March this year that it had integrated a total of 1GW of data center demand response capabilities into long-term energy contracts with multiple U.S. utilities to reduce demand during grid peaks by shifting or limiting some machine learning workloads. This means that the AI load participates in demand response and has begun to move from research concepts to actual commercial contracts.
This is especially worthy of reference for Taiwan. When Taipower communicated with the Electric Power Research Institute (EPRI) this year, they also discussed the concept of "Power Couple—AI and power co-location" to match power supply and demand in AI data centers. If Power Couple deals with "where to use electricity", then AI load elasticity deals with "when and how much electricity is used".
The question can therefore be more specific: When applying for grid connection for large data centers in the future, in addition to stating "how much power is needed", should they also state "how much flexibility can be provided"? What workloads can be moved? How much can it be lowered during peak times? How long does it take to react and how long does it last? Can these promises be verified by the grid?
If large data centers are willing to reduce part of the load during peak hours on the grid, can they achieve more efficient grid connection arrangements? If the promise is not fulfilled, should there be corresponding responsibility and cost mechanisms? This is the next step in AI energy governance that is truly worth discussing. One of the directions AEMA is currently advocating is to provide data centers with verifiable elasticity with a more efficient grid connection path.
Because if all new AI power demand and supporting costs are shared by power plants, transmission lines, substations and general users, then the growth costs of AI may be partially passed on to the entire power system.
On the other hand, if the data center can proactively reduce the load when the power grid is tight through computing transfer, energy storage, local power supply and demand response, it will no longer just be a power user "waiting for power", but can become part of the power system.
If the data center can actively reduce the load through computing transfer and energy storage when the power grid is tight, it can become part of the power system. The picture shows Longqi ultra-high voltage substation.
Energy governance in the AI era certainly requires increasing power generation and expanding the power grid; but it cannot just increase supply without requiring large power users to increase flexibility. A truly mature AI energy policy should allow the data center to answer another question: When the power grid needs flexibility the most, how much can you provide, for how long, and can it be verified?
What the AI data center really needs is not just more electricity, but a system that clearly explains, calculates, and verifies "how much electricity is used, when it is used, and whether the power grid can use less when it is most needed."
(The author is the founder of the Taiwan Sustainable Governance Forum and an associate professor at the Cultural University)
