Table of Contents
In its latest “AI Infrastructure Insight” article, SK Hynix argues that the expansion of AI workloads is changing how data centers are designed and operated
In sum – what to know:
Integrated infrastructure – AI performance increasingly depends on compute, memory, storage, networking, power and cooling working together.
Rack-scale design – Data center architecture is moving beyond individual servers toward rack- and cluster-scale systems.
Infrastructure investment – AI is expanding investment beyond chips into networking, power, cooling, construction and broader data center infrastructure.
AI data centers are evolving from facilities built primarily around servers and storage into integrated infrastructure in which compute, memory, storage, networking, power and cooling must work together, according to SK Hynix, the semiconductor arm of Korea’s SK Group.
In its latest “AI Infrastructure Insight” article, the company argues that the expansion of AI workloads is changing how data centers are designed and operated.
As AI models become larger, both computational demand and data movement are increasing. Training requires massive datasets to be repeatedly read and processed, while inference depends on retrieving information quickly in response to user requests.
At the same time, AI applications are expanding across areas including search, workplace productivity, content creation, manufacturing, finance, and healthcare, increasing the volume and frequency of requests handled by data centers.
SK Hynix said this is shifting the focus of data center design beyond the performance of individual processors. Compute resources can only deliver their full potential when storage can supply data quickly, memory can deliver data efficiently, networks do not become bottlenecks, and power and cooling can keep pace, the firm said.
The company identifies five key components of AI data centers: compute, memory, storage, networking, and power and cooling.
Compute resources, including GPUs, AI accelerators and CPUs, remain the engines for AI training and inference. However, SK Hynix said their performance increasingly depends on their ability to receive a stable supply of data and connect with other infrastructure resources with minimal latency.
Memory is also taking on a more critical role. SK hynix describes memory as the layer connecting compute and data, with High Bandwidth Memory (HBM) positioned close to accelerators to provide high bandwidth, while server DRAM provides greater capacity and is shared across the server.
Storage is becoming increasingly focused on the speed of storing and retrieving data, rather than simply long-term data retention. Networking has similarly evolved beyond a communications layer, with latency, bandwidth and data transfer efficiency becoming important to the scalability of AI data centers.
Power and cooling are also becoming foundational design requirements as higher-performance systems are deployed at greater density, increasing both power consumption and heat generation.
From servers to racks and clusters
SK hynix said the scale of AI workloads is also changing the primary unit of data center design.
Traditional data centers focused heavily on individual server performance, but large-scale AI training and inference require multiple accelerators and servers to exchange data, coordinate computation and divide workloads across the system.
As a result, architecture is moving from server-centric designs toward rack-scale and ultimately cluster-scale systems.
At the rack level, compute, memory, power delivery, cooling and networking need to be designed together. At the cluster level, multiple racks must operate as a single computing system, increasing the importance of network architecture and data movement efficiency.
The company also points to changes in where and how data centers can operate. It cites Microsoft’s Project Natick as an example of an underwater data center experiment and Google’s Project Suncatcher and Starcloud’s satellite efforts as examples of AI computing being explored beyond Earth.
SK Hynix said these examples point to a broader change in how data centers are conceived, with questions increasingly focused on optimizing power delivery, cooling, networking and data movement as an integrated system.
The company concludes that AI infrastructure competitiveness depends not on individual components alone, but on how effectively compute, memory, storage, networking, power and cooling are connected and optimized as a unified system.
SK Hynix is set to dramatically increase its memory capacity. At Computex 2026, SK Group chairman Chey Tae-won, who also chairs SK Hynix, told reporters that the company plans to double its memory wafer production capacity within five years. Chey described the move as a response to a structural shortage of AI memory chips he expects to persist until roughly 2030
