KT Cloud Builds Massive AI Training GPU Infrastructure... Full-Scale Market Entry

Building Hyperscale AI Computing (HAC) Infrastructure
Collaboration with Domestic Company 'More' to Develop AI Framework
Developing Domestic AI Cloud Semiconductor Chip with 'Rebellion'

KT Cloud employees are inspecting equipment at the IDC where the Hyperscale AI Zone has been established. Photo by KT Cloud

[Asia Economy Reporter Oh Su-yeon] KT Cloud announced on the 25th that it will build a hyper-scale artificial intelligence (AI) training GPU infrastructure at the KT Daedeok 2 Research Center by December.

KT Cloud plans to establish a hyper-scale AI computing (HAC) infrastructure in this project to verify optimal performance and operational efficiency, and to actively target the hyper-scale AI business market.

Hyper-scale AI is designed to think autonomously like humans by learning large-scale data based on GPU infrastructure capable of massive computations. KT has decided to promote the hyper-scale AI business in earnest and secure GPU computing infrastructure for timely training by simultaneously introducing KT Cloud’s HAC and NVIDIA’s GPU appliances.

Through participation in this project, KT Cloud will secure key business references for hyper-scale AI with HAC. They plan to review and supplement hyper-scale AI training performance improvements and retraining results to elevate their technology to a global top-tier level. In collaboration with the domestic company 'Moreh,' they are developing an optimized AI framework, and together with 'Rebellions,' they are also advancing the development of domestic AI cloud semiconductor chips.

KT Cloud’s HAC is the world’s first pay-as-you-go GPU service launched in December last year. It boasts strengths such as business agility, cost efficiency, development flexibility and continuity, and programming compatibility necessary for scaling AI models. HAC uses AMD products, which are well-known for their cost efficiency, as its GPUs.

Introduction to Hyperscale AI Computing (HAC) Service. Photo by KT Cloud

HAC logically combines multiple GPUs to be used as a single GPU and supports clustering of hundreds to thousands of GPUs. When large-scale computations are needed, it allows flexible service use by allocating only the required GPU resources through multi and dynamic allocation technology and returning them immediately after computation. Since costs are charged only for the period and resources allocated for GPU use based on the cloud, it significantly reduces the burden compared to building expensive equipment. Additionally, it guarantees programming compatibility that allows reuse of existing development sources such as code reuse, as well as flexibility and continuity to elastically change the scale of resources needed during development and continue from past training results.

As a result of KT Cloud’s hyper-scale AI language model training test, HAC showed a cost reduction effect of 30-50% compared to existing on-premise services under equivalent conditions. KT Cloud plans to continuously expand its server farms and expects to provide equivalent computing resources at up to one-tenth the cost efficiency in the future.

Yoon Dong-sik, CEO of KT Cloud, said, “HAC is an innovative technology and service that provides the large-scale infrastructure essential for hyper-scale AI research and service development. It lowers the barrier to the hyper-scale AI industry and is creating success stories with AI specialized companies and startups that require large-scale GPU infrastructure. Moreover, it will lead the growth of the hyper-scale AI industry together with leading domestic and international AI companies.”

Meanwhile, according to IT market research firm IDC, the global AI solutions market size is expected to reach $450 billion (approximately 648 trillion KRW) this year. This represents a growth of over 17% compared to $383.3 billion (approximately 552 trillion KRW) last year, and the growth trend is expected to continue for more than five years. Among them, the AI hardware market, with revenue of $1.88 billion (approximately 2.7 trillion KRW) last year, is smaller in scale compared to the overall market but shows the steepest growth. IDC explains that the revenue growth rates for AI servers and storage are 39.1% and 32.9%, respectively, as companies build dedicated AI systems according to the computing and storage demands of AI models and data sets, resulting in this growth trend.

This content was produced with the assistance of AI translation services.