$SK Hynix(SKHY.US) (SK Hynix) and $NVIDIA(NVDA.US) (NVIDIA) are becoming the core beneficiaries of the "computing power bottleneck shift" in the AI era.

Over the past two years, the market has primarily focused on GPU quantities: how much AI computing power can be provided by H100, H200, and Blackwell systems. However, as AI models continue to scale, new constraints are emerging—memory bandwidth and high-performance memory supply are becoming key bottlenecks in AI infrastructure.

$SK Hynix(SKHY.US): Structural changes are occurring in AI memory demand

SK Hynix's DRAM revenue has grown more than 5-fold over the past three years, with the core driver coming from AI server demand, especially HBM (High Bandwidth Memory).

Traditional servers primarily rely on standard DRAM, while AI accelerators require large amounts of high-speed memory to feed the GPUs.

Taking the $NVIDIA(NVDA.US) Blackwell system as an example:

The Blackwell architecture further increases computational power, but simultaneously requires higher bandwidth and larger capacity HBM support. If GPU computing power increases but memory cannot keep pace, the problem of "computing power waiting for memory" arises.

This means:

The stronger the GPU, the higher the dependence on HBM.

The special aspect of HBM is:

The amount of silicon wafer consumed per GB of HBM is approximately 2–3 times that of standard DRAM.

The reason is that HBM achieves higher bandwidth through multi-layer DRAM stacking, requiring more DRAM dies and advanced packaging technology.

Therefore, when wafer capacity shifts from standard DRAM to HBM, the actual impact far exceeds a simple capacity conversion:

A wafer used for HBM may consume capacity equivalent to reducing the supply of standard DRAM by 2–3 times.

This is also why a new memory supply-demand dynamic may emerge in the AI era:

Past:
The DRAM market followed cyclical expansion → oversupply → price decline.

Now:
AI server demand drives rapid HBM growth while simultaneously squeezing traditional DRAM supply.

Supply growth rates may lag behind AI data center demand in the long term.

The impact of $NVIDIA(NVDA.US) Blackwell

Blackwell is not just a GPU upgrade, but a complete AI data center system.

Including:

GPU
HBM
NVLink high-speed interconnect
Network switching
Liquid cooling systems
Data center infrastructure

Among these, HBM has become one of the most critical limiting links in the AI chip supply chain.

Therefore, AI semiconductor competition is shifting from:

"Who makes the fastest GPU"

To:

"Who can provide a complete AI computing system, including GPU, HBM, advanced packaging, and high-speed networking."

Beneficiary chain:

First tier:
$NVIDIA(NVDA.US)
The leader in AI accelerators, controlling the GPU ecosystem.

Second tier:
$SK Hynix(SKHY.US)
One of the HBM leaders, directly benefiting from the explosion in AI memory demand.

$Micron Tech(MU.US)
A U.S. supplier of HBM and advanced DRAM.

$Taiwan Semiconductor(TSM.US)
Core of advanced packaging and AI chip manufacturing.

Third tier:
$Broadcom(AVGO.US)
AI networking chips and high-speed interconnect.

$Arista Networks(ANET.US)
AI data center networking equipment.

$Marvell Tech(MRVL.US)
High-speed connectivity and data center chips.

The core formula for future AI infrastructure competition is changing:

Past:

GPU quantity × computing power = AI capability

Future:

GPU computing power × HBM bandwidth × network interconnect × power supply = AI system capability

Therefore, HBM is not just a memory upgrade, but a strategic resource in the AI computing architecture.

As next-generation AI platforms like Blackwell and Rubin continue to expand, HBM may become one of the most important growth bottlenecks in the AI semiconductor supply chain after GPUs.