Adjusting AI workloads to grid conditions
“It is essential to design and build power infrastructure and artificial intelligence (AI) infrastructure together, rather than as completely independent systems.”
As rapid expansion of AI data centers increases pressure on electricity grids, speakers proposed responding not only by building new power plants and transmission networks but also by allowing data centers to adjust electricity consumption flexibly. They said South Korea, where AI demand is growing rapidly, also needs to combine greater grid investment with more efficient operation of existing infrastructure.
Mansi Shah, head of product at Emerald AI, emphasized this approach during a presentation at the Bitgaram International Exposition of Electric Power Technology (BIXPO 2026), held at the Kimdaejung Convention Center in Gwangju on November 5. “Given the scale and speed of AI growth, we cannot wait 5, 7 or 8 years for power infrastructure,” Shah said.
Shah explained that not all AI data center tasks require the same level of power supply. Some tasks, such as fraud detection, need immediate processing, but some AI tasks can be delayed slightly without significantly affecting service.
“AI workloads themselves have inherent flexibility,” Shah said. “Even if power is limited, the work continues, only somewhat more slowly.”
Shah suggested limiting GPU power consumption, shifting tasks to times when the grid is less congested, and distributing work to data centers in other regions as ways to adjust electricity use.
“Using data centers flexibly can allow them to begin operating sooner on the existing grid, even before new transmission lines, substations and generating facilities are completed,” Shah emphasized.
Shah also described practical applications. Emerald AI conducted initial demonstrations in the United States, the United Kingdom and the European Union (EU). At a California data center, software adjusted about 35,000 AI tasks to reduce power consumption, Shah explained.
“We maintained a reduction in power consumption of about 15% and gradually increased consumption after grid conditions returned to normal,” Shah said. “This enabled the data center to secure about 25% more power immediately.”
Shah urged the application of similar approaches to data center development in South Korea. “Most data centers in South Korea want to be located in the Seoul metropolitan area,” Shah said. “Where to obtain power is the key constraint.”
“Power flexibility works in practice and can be a powerful tool,” Shah added, expressing willingness to pursue demonstration projects with South Korean power companies and others.
Grid investment and digital operations
However, speakers also cautioned that adjusting electricity demand alone cannot replace investment in new facilities.
Mark McGranaghan, a fellow at the US-based Electric Power Research Institute (EPRI), emphasized that increased grid investment and digital transformation must proceed together. “We cannot solve everything with flexibility. We need to build transmission and distribution lines,” McGranaghan said.
Citing analysis by European electricity industry association Eurelectric, McGranaghan said, “European distribution grids need about €67 billion annually. That is at least twice the amount invested each year over the past 5 years.” Factors behind the need for increased investment included not only data center expansion but also electric vehicle adoption, renewable energy growth and grid resilience to climate change.
McGranaghan stressed that greater investment must be accompanied by better use of existing facilities. This requires a platform to forecast and adjust electricity use across resources including data centers, electric vehicles and residential energy storage systems.
“Digitalization is the key to making all of this work,” McGranaghan said. “Sensors and data on the grid are essential to implementing AI.”
McGranaghan also discussed Korea Electric Power Corporation's advanced distribution management system (ADMS), saying a demonstration at the KEPCO Human Resources Development Institute was “an excellent starting point for the technology needed in future control centers.” McGranaghan added, “We need to optimize grid operations by integrating management, forecasting and coordination,” and urged that the benefits of new operating technologies be incorporated from the investment planning stage.
Building trust in AI applications
In a panel discussion moderated by Seoul National University professor Lee Jong-soo after the presentations, trust was emphasized as a prerequisite for cooperation between AI and the power industry. Participants also suggested using virtual simulations to verify safety and performance before introducing AI in industrial settings, and improving acceptance among on-site personnel.
“Making grid operators' work easier and more accurate is the first step toward building trust before introducing AI into critical infrastructure,” McGranaghan said.
Shah also emphasized, “Power companies and data centers need safeguards in place and must share sensitive data to find trustworthy solutions.” Shah added, “If data centers operate flexibly and bear more of the cost, communities could obtain a more reliable grid at the same or lower cost.”
Original Korean article: Edaily
English brief by Korea Wind Intelligence, koreawindintel.com
