South Korea should propose international AI data center standards in 2027
Concerns that big technology companies' criteria could become established amid a lack of AI data center inference efficiency standards
AI competition is extending beyond technology development to international standards that establish the rules. The United States, the European Union (EU), China and Japan are bringing together government and corporate experts to incorporate their technologies and industrial experience into standards. As inference demand grows, the scope of competition is expanding to AI data centers. South Korea ranks among the leading countries in participation at international meetings, but its ability to lead standards development remains limited. Digital Daily examines the state of AI standards competition and the challenges South Korea needs to address through perspectives from those involved in international standardization.
As the focus of AI use shifts from training to inference, international criteria for measuring the efficiency of AI data centers (AIDCs) have emerged as a new standardization task. Existing indicators such as power usage effectiveness (PUE) cannot readily distinguish how much useful output was produced using the same amount of electricity. With big technology companies' own criteria potentially becoming established before international criteria are in place, there are calls to accelerate standards proposals supported by measurement methods and empirical data.
The criteria used to evaluate AI services also change in the inference era. During training, the main subjects of evaluation were datasets, model architecture and training performance. During inference, response speed, throughput and power consumption jointly determine service quality. Service reliability and cost per inference have also emerged as major evaluation criteria. As evaluation criteria change, the measures used to assess AI data center efficiency need to change accordingly.
Measurement methods first as standards work shifts to inference
Recent AI standards discussions are also expanding from models to service operations. The AI subcommittee (SC 42) under the Joint Technical Committee (JTC 1) of the International Organization for Standardization (ISO) and the International Electrotechnical Commission (IEC) has begun work on inference standards. ISO/IEC TS 25258, led by South Korea, is a representative example. This technical specification establishes a framework for organizing inference in AI systems that process multiple tasks simultaneously. It is currently at the committee draft (CD) stage, with development targeted for completion in 2027.
Even with a framework for organizing inference, standards remain incomplete without measures for comparing efficiency. Power consumption, response speed, throughput and cooling costs differ among data centers even when they run the same AI service. Without specifying which tasks to measure and under what conditions, it is impossible to determine which is more efficient. This is why measurement methods are considered the most urgent task in standardization discussions.
International standardization starts with establishing measurement methods that allow countries to make comparisons under the same conditions, rather than mandating a particular technology. Data center designs and equipment configurations vary by country, making international agreement difficult if a particular technology is prioritized. Common measurement methods are needed to compare countries' technologies against the same criteria.
Inference efficiency cannot be judged solely by electricity savings. Lower electricity use does not necessarily mean higher efficiency if responses are slow or outputs are less useful. Assessment must jointly consider how much electricity, computing resources and cost are required to produce a useful answer.
Standardization work begins by defining the characteristics of inference workloads. The type of workload used as a reference must be determined to clarify what is being measured. Comparisons among data centers require measurement methods that produce the same results regardless of who performs the measurement. Discussions on supplementing existing efficiency indicators can then proceed on that basis.
Limitations of PUE and the possibility that big technology companies' criteria could become established first
International data center standards already exist. The ISO/IEC 22237 series covers data center facilities and infrastructure, while the ISO/IEC 30134 series addresses resource efficiency indicators. A revised edition of ISO/IEC 30134-2, which defines PUE, was also published in January this year.
However, this indicator alone cannot readily measure AI data center efficiency. AI data centers typically operate densely installed high-performance accelerators and consume substantial amounts of electricity. They also require specialized cooling equipment. Their allocation of resources among different workloads also differs from that of conventional data centers. PUE indicates facility efficiency through the ratio of total data center power consumption to IT equipment power consumption, but it does not show how much useful inference output was produced with that electricity.
The productivity of AI computing is an area not covered by existing standards. Rather than discarding existing indicators, a practical approach is to add a measurement framework that reflects AI workload characteristics or revise existing standards.
If international standards are delayed, big technology companies' own criteria could become established first. Meta, OpenAI and others operate very large data centers according to their own criteria. When companies with substantial investment resources and control over supply chains adopt particular infrastructure architectures or interfaces, equipment and service suppliers follow. These approaches can readily become de facto market standards. However, internal corporate operating specifications differ in nature from international standards agreed upon by multiple countries and stakeholders.
International standards would allow all market participants to evaluate AI data centers against the same criteria. Data center developers would find it easier to compare the performance and efficiency of their facilities with those of others, while equipment suppliers could make products to a standard and supply them to multiple data centers. Institutions and companies commissioning data centers could also use standards as a basis for equipment purchases and acceptance inspections. For South Korea, this would be an opportunity to incorporate domestic technology into international criteria rather than follow operating practices established by global companies.
Five candidate international standards for AI data centers proposed by Jo Young-im, a professor in Gachon University's Department of Computer Engineering and head of South Korea's delegation to ISO/IEC JTC 1/SC 42.
SC 39 and SC 42 linkage; empirical data as the key issue
International discussions are being advanced jointly by SC 39 and SC 42. SC 39 is the subcommittee responsible for sustainability in IT and data centers. It covers facilities, power efficiency, water use and energy reuse. It is currently developing ISO/IEC TR 25953, a technical report (TR) connecting AI and data center sustainability, at the committee draft stage. A joint advisory group (JAG) on AI and sustainability has also been established between the two committees.
AI data center standards require the expertise of both committees. SC 39 has strengths in facility and energy efficiency, while SC 42 has strengths in AI computing and inference. However, which committee will lead the new standards and how they will be developed together have not been decided. Whether to undertake joint development will be determined through formal procedures separate from the joint advisory group.
Several procedures are required to propose domestic criteria as international standards. Establishing AI data center policies or power and cooling guidelines does not automatically make them Korean Industrial Standards (KS) or international standards. The scope of existing KS and international standards must be analyzed, and newly required measurement methods and requirements must be specified. Procedures involving the Korean Agency for Technology and Standards and domestic technical committees then follow. Establishing KS and proposing international standards can also be pursued simultaneously.
The process also takes considerable time. A typical international standards project takes 3 to 5 years from approval to publication. The schedule becomes longer if operating data, test methods, co-proposing countries and a project leader are not prepared together from the outset.
Empirical validation determines whether a country can take the lead. South Korea has capabilities across semiconductors, power equipment, cooling, networks and data center operations. However, claims of superior cooling or power efficiency alone are insufficient to persuade other countries. Internationally comparable test results are needed for South Korean technology to be incorporated into standards.
Jo, a professor in Gachon University's Department of Computer Engineering and head of South Korea's delegation to ISO/IEC JTC 1/SC 42, identified a national standardization roadmap addressing both AI data center energy and resource efficiency and inference performance as a priority. Jo also called for accelerating efforts to secure industrial measurement data and recruit co-proposing countries.
Jo proposed putting forward at least 1 candidate international standard for AI data centers by the second half of 2027. The proposed schedule is to select priority tasks in the first half of 2027, review whether they overlap with existing international standards, and submit proposals in the second half.
Jo identified 5 candidate areas: measuring the energy efficiency of AI inference workloads, cooling performance for high-density AI infrastructure, inference service quality assessment, AI workload-based resource management, and validation of AIDC operating data. Among these, Jo prioritized reviewing a method that measures inference service performance and energy consumption together.
Distinguishing the roles of new and existing standards remains a task. The ISO/IEC 30134 series addresses the power efficiency of data center facilities through indicators such as PUE. To be persuasive in international discussions, new measurement methods must clearly define their scope as measuring the electricity and resources used to produce inference outputs.
Jo said, "Taking the lead in standards competition does not mean submitting a document title before other countries." Jo added, "Taking the lead in practice means being the first to define an internationally relevant problem, secure empirical data and propose test methods that multiple countries can accept."
Original Korean article: Digital Daily
English brief by Korea Wind Intelligence, koreawindintel.com
