Where is the AI infrastructure investment cycle right now? If Act 1 — the explosive surge in GPU demand — was symbolized by Nvidia's stock price, the market has now begun asking who the lead of Act 2 will be. The history of SK Hynix's HBM (High Bandwidth Memory) pioneering and the "intelligence export" discourse within Korea's AI ecosystem can be read as field-level signals that the center of gravity of the CapEx cycle is migrating from silicon hardware toward software and model layers.

What HBM Attests About the CapEx Cycle's Structure

In AI computing infrastructure, GPUs and memory cannot be thought of separately. For accelerators like Nvidia's H100 and B200 to deliver performance, ultra-fast HBM must be attached alongside them. The Korea Economic Daily's account of SK Hynix Chairman Chey Tae-won's footsteps captures this structure concisely. The essence of HBM pioneering — spanning from the Cheongju factory to New York partnerships — was designing the memory architecture preemptively before AI accelerator demand even materialized. In other words, the AI CapEx cycle is not the story of a single GPU fabless company; it is a simultaneous investment cycle across the entire supply chain, extending from memory and packaging to cooling and power.

From an investor's perspective, this structure carries important implications. In the early cycle, the bottleneck was in GPUs, so Nvidia's pricing power was overwhelming. As datacenter operators begin accumulating GPU inventory, however, the next bottleneck shifts to HBM supply, then to power and cooling infrastructure, and finally to the software layer that leverages all that hardware. This is why tracking "where the current bottleneck lies" is the core framework in CapEx cycle analysis.

At the current moment, HBM supply is effectively an oligopoly held by SK Hynix, Samsung, and Micron, with SK Hynix's leading technology particularly pronounced in the highest-spec product families, as Korea Economic Daily's TechX column notes. As long as this oligopoly structure holds, the HBM layer can maintain considerable pricing power in the cycle following GPUs. The conditions under which this judgment fails are equally clear: if Chinese memory manufacturers catch up on HBM technology faster than expected, or if AI model architectures evolve in a direction that reduces memory intensity.

HBM memory semiconductor production facility

The 'Intelligence Export' Discourse and What It Signals About the Software Layer Transition

The next stage after the hardware cycle is ultimately when the software and model layer becomes monetized. The most direct articulation of this transition came from Ha Jeong-woo, director of Naver's AI research lab. According to Korea Economic Daily reporting, he argued that "Korea should lead with its own AI models and become an 'intelligence-exporting country.'" The phrase "intelligence export" is not mere rhetoric — it points to an industry center-of-gravity shift from hardware exports (semiconductors, equipment) to software and model exports (APIs, services, fine-tuning).

In the AI CapEx cycle context, this argument is significant. The reason hyperscalers (Google, Microsoft, Amazon, Meta) are pouring hundreds of billions of dollars into datacenters is ultimately to capture the margins of model-based services. In other words, CapEx is deployed in hardware but the returns are generated in the software layer. For Korea to become an "intelligence-exporting country," it needs to simultaneously build a competitive model and service layer on top of the hardware supply chain exemplified by HBM.

Writer Song Gil-yeong's diagnosis helps gauge the speed of this transition. In Korea Economic Daily reporting, he stated that "in the AI era, the 'Kim bujang' (middle-manager archetype) will disappear, and AI adoption is not a matter of choice." The AI replacement of middle-management functions within corporations signals that software layer demand is already materializing in the B2B market. When the companies executing CapEx begin proving ROI, the scale and direction of the next investment cycle will be determined.

AI datacenter server racks

Mid-Cycle Check: Framework and Counter-Scenarios

The key variables for assessing the AI CapEx cycle are summarized in the table below. Distinguishing for each layer which stage it is currently in, and what conditions would accelerate or reverse the transition, is the judgment framework investors need.

Layer Current Phase Acceleration Condition Reversal (Failure) Condition
GPU / Accelerators Supply expansion phase; pricing power may gradually decline Sudden demand surge from new model architectures Accelerated in-house chip development by hyperscalers
HBM / Packaging Oligopoly supply; first-mover premium maintained Expanded AI inference demand (growing HBM capacity need) Chinese vendor tech catch-up; memory-efficient architectures
Datacenter infrastructure (power, cooling) Early bottleneck; lead times lengthening Relaxed energy policy in major countries; small modular reactor commercialization Power procurement failure delays datacenter expansion
Software / Model layer Early monetization; ROI proof stage Accelerating B2B automation adoption (replacement of middle management) Regulatory tightening; demand disappointment from model performance plateau

AI CapEx cycle returns are deployed in hardware and recovered in the software layer — right now we are somewhere in between.

The most notable structural change at the current moment is that hyperscalers are increasing investment in developing their own AI chips. Google's TPU, Amazon's Trainium, and Meta's MTIA are all attempts to reduce Nvidia dependence. The more successful this in-house development becomes, the weaker the GPU layer's pricing power — but paradoxically, demand for general-purpose memory components like HBM can persist, since any chip still needs high-speed memory. This is why the HBM supply chain that SK Hynix built from Cheongju to New York becomes a strategic asset in this configuration.

The timing of the software layer transition is governed by B2B adoption speed. The rate at which Song Gil-yeong's diagnosis — "AI adoption is not optional" — translates into actual purchase decisions on the factory floor will be the rationale for hyperscalers executing the next round of CapEx. Conversely, if ROI proof is slower than expected, it cannot be ruled out that the 2026–2027 CapEx cycle will first pass through a hardware inventory adjustment phase.

What to Watch

  • HBM market share changes: Market share among SK Hynix, Samsung, and Micron in HBM3e and above — a leading indicator of whether competition is intensifying
  • Hyperscaler CapEx guidance: Quarterly capital expenditure announcements from Google, Microsoft, Amazon, and Meta — the direct signal of cycle continuation or adjustment
  • In-house chip development progress: Google TPU and Amazon Trainium's actual workload share — a gauge for the rate at which GPU layer pricing power weakens
  • AI B2B adoption rate: Share of AI software and services in corporate IT spending — confirming when software layer monetization begins
  • Power and cooling lead times: US and European datacenter power permitting timelines — whether infrastructure bottlenecks are constraining CapEx execution
  • Chinese HBM technology indicators: Reports of HBM sample shipments from Chinese memory manufacturers such as CXMT — signals of oligopoly structure fracture

References