"Not only graphics processing units (GPUs) but also central processing units (CPUs) are becoming scarce, and next year it will be extremely difficult to get the PCs we usually think of."
Shin Jeong-gyu, CEO of Lablup, said this in a lecture at SMARTCLOUD SHOW 2026, the country's largest tech conference, held at the Westin Josun Hotel in Sogong-dong, Seoul, on the 26th. With the spread of agentic artificial intelligence (AI), shortages and bottlenecks will intensify across computing resources, not only GPUs but also CPUs, memory, storage, and power.
Lablup is a corporations specializing in AI infrastructure management and operations software, founded in 2015. It is developing technology that determines which AI models and computing resources such as GPUs are needed depending on the difficulty of the tasks AI will perform, and allocates the necessary resources efficiently.
According to CEO Shin, agentic AI does not stop at answering users' questions but performs actual tasks in place of people. In this process, not only GPUs that run AI models but also CPUs and memory to handle tasks, storage devices, and software environments are needed together. Shin explained that the spread of agentic AI is similar to a situation where the number of employees working at a company suddenly increases by dozens of times.
The problem is that the required computing resources increase rapidly as the number of agents grows. Based on this year, Shin said about 20 agents operate in the background per person, but by the end of next year, some corporations could have around 100 agents per person running at all times.
Another factor fueling bottlenecks is that other hardware is advancing relatively more slowly compared with GPUs. Shin explained that while GPU performance has risen rapidly, the performance of CPUs, memory, storage devices, and PCI Express (PCIe) connecting them has not improved at the same pace. Shin said, "Even if the number of cores on a CPU increases, if the memory bandwidth shared by each core or the processing speed of storage devices does not increase sufficiently, bottlenecks emerge elsewhere."
Power issues at AI data centers are also growing. According to Shin, in the past, the power supplied to a single server rack in a data center was around 4 kilowatts (kW), but to accommodate Nvidia's next-generation AI accelerator "Vera Rubin," up to 250 kW per rack is required.
Safely isolating and running agents is also a new challenge. Traditional virtual machines (VMs) impose greater system overhead as more run simultaneously. Containers can also separate individual tasks, but because they share resources such as memory and storage, there are limits to fully isolating them.
Shin predicted that the shortage of computing resources driven by the spread of agentic AI will not end in the short term. "We are rapidly deploying all resources, but because demand is overwhelmingly outpacing production, I expect the situation to last," he said. "It will be a once-in-a-generation level of change that we experience only a few times in our lives."