Companies that fail to address the token burden and memory limits caused by AI's rapidly rising computing demands could find that adopting AI threatens their competitiveness rather than strengthens it, according to an assessment presented at an industry forum. As the era of agentic AI — systems that make decisions and act on their own — takes hold, surging computing costs and hardware bottlenecks have emerged as key challenges for businesses. A token is the basic unit of data processed by AI systems.
Dell Technologies Korea held the Dell Technologies Forum (DTF) 2026 at COEX in Seoul on the 25th, where it unveiled a next-generation AI infrastructure strategy built around these themes. At the event, which drew 43 global partner companies, Dell and SK hynix (000660.KS) each presented software and hardware solutions to support the agentic AI era.
In a keynote address, Dell Technologies Korea President You Sang-mo identified companies' lack of execution as the central challenge in today's AI market. "While 61% of business leaders feel they need to integrate AI into the workplace, 64% do not yet have a clear roadmap for doing so," You said. He added that 73% of respondents want security and cyber resilience built in from the earliest stages of AI design, stressing the importance of designing integrated infrastructure tailored to business goals.
Matt Dunphy, a Dell senior vice president, cited cost efficiency as the biggest challenge to the spread of agentic AI. In an environment where multiple AI agents exchange data and carry out tasks on their own, the volume of tokens to be processed rises sharply, which can translate into enormous costs.
"Data must be structured in advance to minimize the amount of information passed to AI models," Dunphy said. "A hybrid infrastructure strategy that combines on-premise systems and the cloud according to cost efficiency is essential."
Joo Young-pyo, a vice president at SK hynix, laid out the direction of next-generation memory technology needed to handle surging AI computing demand. "Agentic AI requires processing hundreds of tokens per second, pushing the bandwidth and power efficiency of existing systems to their limits," Joo said.
To address this, SK hynix presented three core technologies: 3D vertical stacking, processing-in-memory (PIM) and memory pooling. The approach stacks memory vertically to overcome the limits of physical space, and combines storage and computing functions to reduce the power consumed when data is moved.
The company also plans to use memory pooling technology based on Compute Express Link (CXL) to cut the idle time in which high-performance GPUs wait due to data transmission delays, improving the operating efficiency of AI infrastructure.
