AI is Not Just a Chip Story – but a Memory Revolution

People often lack “long-term vision” to clearly see the present. When a trend has revealed enough data, the crowd tends to doubt because… the price has already increased significantly.
Over the past three years, the VRAM capacity on each NVIDIA GPU has increased nearly 3.6 times. But what’s truly remarkable is not just the GPU — but the amount of accompanying HBM memory. In the next three years, the demand for HBM is forecasted to increase up to 5 times.

So the question is no longer: “Has the memory stock already risen a lot?” But rather: “Do you truly believe AI is a long-term revolution?”

When NVIDIA Grows, Memory Cannot Stay Out
If you think memory stocks cannot cycle, then it means:
NVIDIA’s GPU sales will not continue to explode
AI demand will soon cool down
Global token growth is just a temporary phenomenon
But the reality shows AI is still expanding exponentially. Jensen Huang once emphasized that the computational demand for AI could increase by thousands of times. If compute increases by thousands, then memory — the fuel for compute — cannot just increase in small increments.
AI consumes three main components:
Compute (Processing)
Storage/Memory (Memory)
Optical Interconnect (Optical Connectivity)
If you believe compute is a long-term growth story, then memory by default is also a long-term growth story. These two factors are inseparable.

The “Drain” Effect from NVIDIA: The “Depletion” of Supply
Moving into the Blackwell Ultra and Vera Rubin generations:
New GPUs consume up to 288GB of HBM3e per card
High-end versions can reach 384GB – 512GB
Vera CPUs come with upgraded system DRAM up to 1.5TB per socket
This means: a high-end AI rack today consumes DRAM much more than traditional servers.
More importantly, NVIDIA uses:
Large-scale upfront cash payments
Long-term “take-or-pay” contracts
=> Locking in most of the high-end HBM capacity for the next 2–3 years.
Consequences:
Other hyperscalers like Google or Amazon must compete for the remaining supply
Traditional servers and consumer electronics are forced to compete for DRAM resources
The industry is entering a structurally shortage state, no longer just a typical cyclical shortage.

Where Are the Big Three Memory Companies?
SK Hynix
HBM capacity in 2026 is nearly sold out
Pushing hard on M15X plant and Yongin cluster
But large-scale production only truly booms from late 2027–2028
Actual orders have already exceeded theoretical supply capacity
Samsung Electronics
Target to increase HBM growth by over 3 times
Massive investments in P5H
HBM4 depends on advanced packaging yield and 4nm logic process
Significant capacity is only expected around 2028
Micron Technology
Capex surged to over $25 billion
HBM4 enters high-volume production
Most of the supply in 2026–2027 is already tied up in long-term contracts
The US super-factory will only contribute significantly from 2028

Physical Barriers: Cannot “Open the Valve” Immediately
There are two hard limits:

1. Cleanroom construction cycle: From start to standard-compliant takes 18–24 months. Cannot be shortened by money.
2. HBM4 consumes 3–4 times more wafers than regular DRAM: HBM4 uses a 2048-bit interface and requires base die logic on advanced processes. This prolongs manufacturing cycles and increases wafer consumption.
   => Despite record-breaking investments, actual bit growth remains limited by physical constraints.

Conclusion: Memory Is Transforming into Growth Stocks
NVIDIA’s valuation reflects an assumption: AI is a long-term revolution.
If you accept that assumption, then it’s undeniable:
Memory is the fuel of AI
HBM is a strategic resource
High-end DRAM is undergoing a global supply-demand restructuring
Essentially, compute and memory are two sides of the same coin. They will either both become long-term growth drivers or both collapse together.
The market may still be skeptical. But data on capacity, binding contracts, and physical barriers show: this cycle may no longer be a “cycle” in the traditional sense.