The AI industry's focus has fundamentally shifted from creating the largest language models to deploying those models affordably and efficiently in real-world settings. At GTC 2026, the world's most intensive artificial intelligence conference, this transformation became impossible to ignore. What was once a race to build bigger and smarter models has become a race to run those massive systems cost-effectively in data centers and manufacturing facilities. Why Is Infrastructure Now More Important Than Raw AI Power? For years, the AI narrative centered on model size and capability. Companies competed to announce larger language models (LLMs), which are AI systems trained on vast amounts of text data to understand and generate human language. But at GTC 2026, the conversation on the exhibition floor told a different story entirely. Nearly every booth, regardless of company, gravitated toward discussions about data centers and infrastructure rather than model performance. NVIDIA CEO Jensen Huang stressed during his keynote address that the age of AI agents, which are autonomous software systems that can perform tasks independently, will transform the world. This messaging effectively signaled a restructuring of the entire AI industry, moving away from pure model competition toward practical deployment challenges. The shift reflects a hard economic reality: deploying massive AI models is expensive. Companies now need solutions that reduce costs, improve efficiency, and integrate seamlessly into existing infrastructure. Korean research teams at the conference showcased work on lightweight medical AI, robotic process simulation, and GPU communication optimization, all centered on cost reduction and efficiency. What Role Is Memory Playing in This New AI Race? At the heart of this infrastructure revolution lies a component most people have never heard of: high-bandwidth memory (HBM). HBM is a specialized type of memory that stacks multiple data storage chips vertically to dramatically increase how fast information can be processed. SK Hynix, a South Korean semiconductor company, has become NVIDIA's critical partner in this space, commanding a dominant share of the HBM market. The partnership between NVIDIA and SK Hynix began in 2013 when SK Hynix first developed HBM technology. When NVIDIA decided to use HBM in its AI accelerators, the two companies began collaborating closely. As AI adoption exploded following the ChatGPT boom, both companies emerged as defining icons of the AI era, with SK Hynix evolving from a simple component supplier into a true co-innovator in AI chip performance. SK Hynix's booth at GTC 2026 demonstrated this evolution vividly. The company showcased a 1,000,000 times enlarged model of HBM to help non-experts understand the product's structure. More impressively, visitors could physically place memory components into a mock-up of NVIDIA's next-generation AI accelerator platform called Vera Rubin, with lights turning on to show how power flows through the system. This hands-on approach made abstract semiconductor technology tangible and understandable. How Are Companies Customizing Memory Solutions for AI? One of the most significant innovations on display was custom HBM (cHBM), which represents a fundamental shift in how memory is designed for AI systems. Rather than offering one-size-fits-all memory chips, SK Hynix is now working with individual customers to create memory solutions tailored to their specific AI needs. The core innovation in cHBM is something called Stream DQ Architecture, implemented on the base die, which is a chip mounted at the bottom of the HBM package. Think of it this way: if HBM is a workbench and the GPU (graphics processing unit) is the worker, Stream DQ Architecture is like laying a conveyor belt on that workbench to pre-process materials before the worker touches them. In traditional systems, the GPU had to handle all data preparation itself. With cHBM, those pre-processing tasks are offloaded to the base die inside the HBM stack, reducing the GPU's burden and accelerating computation. According to SK Hynix, this architecture can improve maximum inference throughput, which is how many predictions an AI model can make per second, by around seven times. This is particularly important for large language model inference, the process of running a trained AI model to generate responses. By enabling customer-specific differentiation in HBM, cHBM pushes SK Hynix beyond the role of a mere supplier toward that of a true technology partner in AI chip performance innovation. What Storage Solutions Are Emerging for AI Data Centers? Another critical piece of the infrastructure puzzle is storage. NVIDIA's Vera Rubin rack incorporates a dedicated storage layer called ICMS (In-Context Memory System), designed to hold conversational and contextual data for AI. As LLM usage has gone mainstream, the volume of context data has exploded, yet storing this data in expensive HBM is highly inefficient. NVIDIA's solution is to carve out a separate ICMS tier and populate it with large-capacity enterprise SSDs (solid-state drives), which are faster and more reliable than traditional hard drives. Industry expectations suggest that each Vera Rubin rack could require on the order of 9,600 terabytes of storage capacity. In practical terms, that means each Vera Rubin rack sold translates into incremental demand for both HBM and eSSD from SK Hynix. SK Hynix anticipated this trend by exhibiting eSSDs optimized for Vera Rubin's direct liquid-cooled environment, such as the PEB210 E1.S model. NAND flash memory, once dubbed the "forgotten memory," is making a comeback in AI data centers, and SK Hynix is moving quickly to prepare for this market shift. Steps to Understanding AI Infrastructure's New Priorities - Model Deployment Over Model Size: The AI industry is shifting focus from building larger models to deploying existing models more cost-effectively and efficiently in real-world data centers and manufacturing environments. - Memory as a Competitive Advantage: High-bandwidth memory and custom memory solutions are becoming critical differentiators in AI performance, with companies like SK Hynix moving from component suppliers to technology partners. - Storage Infrastructure Expansion: As AI systems process more conversational and contextual data, enterprise SSDs and specialized storage systems are becoming essential components of AI data center architecture. - Customer-Specific Customization: Rather than offering standardized components, leading semiconductor companies are now designing memory and storage solutions tailored to individual customer needs and AI chip roadmaps. How Is Leadership Positioning Companies for the Next Phase of AI? SK Hynix Chairman Chey Tae-won's presence at GTC 2026 was far more than ceremonial. Just a month before the event, Chey held a one-on-one meeting with Jensen Huang in Silicon Valley. Reports indicate that the two leaders discussed not only stable HBM4 supply, which is the latest generation of high-bandwidth memory, but also an expanded collaboration roadmap spanning cHBM, eSSD, and comprehensive AI data center solutions. Around the same time, within a single week, Chey also met with the CEOs of five global tech giants, including Broadcom, Microsoft, Meta, and Google. These discussions went far beyond simple HBM supply. Topics reportedly ranged from co-design of customized memory solutions aligned to each company's next-generation AI chip roadmap to joint work on AI data center architectures. At a time when the AI ecosystem is undergoing a structural transition, Chey is actively moving to lock in multiple partnerships in advance. "The age of AI agents will transform the world," stressed Jensen Huang, signaling a restructuring of the AI industry. Jensen Huang, CEO at NVIDIA Chey's presence at GTC served as a very public declaration of SK Hynix's technical capabilities before the world's leading tech companies. During his tour of the NVIDIA booth, accompanied by relevant executives, he persistently asked "why does it have to be this way?" This questioning approach reflects a company actively seeking to innovate and improve every aspect of AI infrastructure, rather than simply accepting existing solutions. The shift from model-centric competition to infrastructure-centric competition represents a maturation of the AI industry. As models become commoditized and more widely available, the real competitive advantage lies in who can deploy them most efficiently, cost-effectively, and at scale. SK Hynix and NVIDIA's deepening partnership, along with the broader industry focus on infrastructure, suggests that this phase of AI competition will be won not by those who build the biggest models, but by those who build the most efficient systems to run them.
