The memory trade may be entering a completely different phase.
For the last several years, investors have been conditioned to think of memory as a cyclical industry:
Demand rises → manufacturers add capacity → oversupply → pricing collapses → production gets cut → cycle starts again.
But AI is potentially changing that equation.
Citi’s latest DRAM outlook is one of the more aggressive calls I’ve seen this cycle, projecting that the DRAM market could remain structurally undersupplied well into the early 2030s.
If those assumptions prove correct, memory may not simply be another semiconductor cycle.
It could become one of the major bottlenecks limiting AI infrastructure growth.
Citi’s Call: Demand Growing Faster Than Supply
The numbers immediately stand out.
Citi estimates:
→ 2027 DRAM demand: +30.2% YoY
→ 2027 DRAM supply: +18.8% YoY
→ Estimated 2027 deficit: ~8.7%
→ 2028 deficit: ~9.7%
And the more interesting part is the duration.
Citi’s framework suggests the DRAM market could remain in shortage conditions through 2029, 2030 and 2031.
Obviously, nobody can know today whether a four- or five-year supply deficit will actually materialize.
But the underlying mechanism is worth understanding.
Because the issue isn’t simply:
“AI needs more memory.”
The issue is that AI is changing what kind of memory is required and how much semiconductor capacity is consumed to produce it.
The HBM Problem
High Bandwidth Memory — HBM — has become one of the most important components in modern AI accelerators.
GPUs and other AI accelerators need enormous amounts of memory bandwidth.
HBM solves part of that problem by stacking DRAM dies vertically and placing them very close to the processor.
The result is extremely high bandwidth.
But there is a trade-off.
HBM consumes a disproportionate amount of manufacturing resources.
Instead of simply producing conventional DRAM chips, manufacturers have to allocate wafer capacity toward increasingly advanced HBM products.
Then comes advanced packaging.
HBM requires sophisticated stacking, bonding and packaging processes.
That means every additional generation of AI accelerators can create additional pressure on the memory supply chain.
And this is where the current cycle becomes different from previous memory cycles.
The Wafer Allocation Problem
This is one of the most important parts of the thesis.
A semiconductor manufacturer doesn’t have unlimited wafer capacity.
It has to decide where those wafers generate the greatest economic return.
And right now, HBM is extremely attractive.
So when manufacturers allocate more wafer capacity toward HBM, that capacity isn’t necessarily available for conventional DRAM.
This creates a strange situation:
Total DRAM manufacturing capacity can increase while conventional DRAM availability remains tight.
That’s because a growing percentage of the industry’s resources are being directed toward AI-related memory.
This is one reason simply saying:
“Memory companies will add capacity.”
doesn’t fully answer the problem.
The question is:
How quickly can useful capacity actually come online?
Why Supply Can’t Respond Overnight
Memory manufacturing is capital intensive.
A company can’t decide today that it needs 30% more output and have that production available next quarter.
New fabs and clean-room capacity require enormous capital investment.
Equipment has to be ordered.
Facilities have to be constructed.
Processes have to be qualified.
Yield has to improve.
Then the company has to ramp production.
And advanced memory technologies make the process even more complicated.
Citi estimates wafer capacity growth could be capped around 8% in 2027, despite much stronger demand growth.
That difference is the entire thesis.
If demand grows substantially faster than usable supply, pricing power shifts toward the manufacturers.
Server DRAM Is Becoming the Monster
The biggest structural change may be happening inside the data center.
Citi estimates server DRAM demand could increase from approximately:
226.3 billion → 341.7 billion 1Gb-equivalent units
in 2027.
That’s roughly 51% growth in one year.
And servers could represent approximately 67% of total DRAM demand by 2027 under Citi’s estimates.
Think about what is happening.
AI isn’t just increasing the number of GPUs.
Every AI server needs memory around those accelerators.
And as AI workloads become more sophisticated, memory requirements per server can increase dramatically.
That creates demand across multiple layers:
HBM → server DRAM → networking → storage → power → cooling
This is why I continue to look at AI infrastructure as a chain rather than simply a GPU story.
The Next Evolution: Continual Learning
There is another part of the thesis that I find particularly interesting.
AI workloads may increasingly move from a model that is:
Train → deploy → periodically retrain
toward systems that continuously process new information and adapt.
If continual learning becomes increasingly important, the memory requirements could expand further.
Why?
Because the system isn’t simply performing inference.
It is simultaneously handling enormous amounts of information.
That can increase demand for:
→ HBM
→ DDR5/server memory
→ SoCAMM2 and other high-performance memory architectures
→ enterprise SSDs
→ storage bandwidth
→ networking infrastructure
And this creates a multiplier effect.
More compute doesn’t necessarily mean proportionally more memory.
Some workloads can require far more memory per unit of compute.
That distinction matters.
But What About Capex?
This is where the Citi thesis gets particularly interesting.
If memory companies know there is a shortage, why don’t they simply spend aggressively and flood the market?
They are spending aggressively.
Citi estimates total DRAM + NAND capex could rise approximately 46.5% to $80.4 billion in 2027.
DRAM capex alone is projected to increase approximately 51.6% to $58.6 billion.
That’s an enormous increase.
But even that may not immediately solve the problem.
Why?
Because spending money isn’t the same thing as producing qualified wafers.
Capacity takes time.
And manufacturers also have to balance conventional DRAM against HBM demand.
What This Could Mean for Memory Pricing
If demand continues to exceed supply, the industry gets pricing power.
That is potentially very important for companies such as:
$MU
Micron has significant exposure to both conventional DRAM and HBM.
If pricing remains firm while utilization stays high, earnings can increase much faster than revenue because memory manufacturing has substantial operating leverage.
That’s one reason memory stocks can move so violently during upcycles.
The reverse is also true.
When supply exceeds demand, pricing can collapse and earnings can deteriorate extremely quickly.
So the key isn’t simply:
“DRAM demand is growing.”
The key is:
“Is demand growing faster than supply?”
$MU — The Most Direct U.S. Exposure
Micron is one of the clearest ways to express this thesis.
The company has been investing heavily in HBM and advanced memory production while benefiting from the broader AI infrastructure buildout.
If the industry remains supply constrained, Micron could benefit from:
→ Higher DRAM pricing
→ Strong HBM demand
→ Higher utilization
→ Operating leverage
→ Increasing data-center memory content
But there are risks.
Memory is still cyclical.
Competitors can increase production.
AI demand can slow.
HBM supply can ramp faster than expected.
And valuation matters.
A company can report outstanding fundamentals while the stock struggles if investors have already priced in the earnings improvement.
That is why I prefer buying MU on meaningful pullbacks rather than chasing vertical moves.
$SNDK — The Storage Side of the AI Memory Chain
SanDisk provides a different exposure.
Rather than being primarily a DRAM/HBM story, the company is positioned around NAND flash and storage.
And storage could become increasingly important as AI systems generate and process enormous datasets.
AI infrastructure needs:
Compute → Memory → Storage
The more data that is generated, stored and accessed, the greater the requirement for enterprise storage.
This is where the continual-learning thesis becomes interesting again.
More information flowing through AI systems can eventually mean more:
→ Data storage
→ Enterprise SSDs
→ High-performance storage
→ Storage bandwidth
But NAND is also highly cyclical.
So the same warning applies:
Demand growth alone isn’t enough.
The supply-demand balance determines pricing.
The Bigger Picture: Silicon-to-Substation
This is exactly why I continue to use my Silicon-to-Substation framework.
AI infrastructure isn’t one industry.
It is an interconnected chain.
Silicon
$NVDA
$AMD
$INTC
Memory
$MU
$SNDK
SK hynix
Connectivity / Photonics
$MRVL
$CRDO
$AAOI
$LITE
$COHR
Compute Infrastructure
NeoClouds and data-center operators
Power
$VST
$GEV
Nuclear
Grid infrastructure
Substation
The physical electricity infrastructure required to actually power all of it.
A bottleneck anywhere in that chain can slow the entire buildout.
And right now, memory is increasingly looking like one of the bottlenecks worth watching.
What Could Break the DRAM Shortage Thesis?
This is extremely important.
A multi-year shortage forecast should never be treated as guaranteed.
There are several ways the thesis could weaken.
AI accelerator demand slows
If GPU and accelerator deployments slow materially, memory demand could fall below current expectations.
AI efficiency improves
If models become dramatically more efficient, companies may achieve the same AI output with fewer compute and memory resources.
That could reduce the amount of infrastructure required.
HBM yields improve faster
Better manufacturing yields could effectively increase usable supply without requiring equivalent increases in wafer capacity.
New fabs ramp faster
If manufacturers bring meaningful capacity online earlier than expected, the projected deficits could shrink.
AI spending becomes more disciplined
If hyperscalers slow capex because ROI doesn’t meet expectations, the entire semiconductor infrastructure chain could experience a reset.
This is the risk I don’t want investors to ignore.
Structural doesn’t mean permanent.
What I Am Watching
For the memory trade, I would focus less on headlines and more on a handful of indicators.
1. DRAM pricing
Are contract and spot prices continuing to strengthen?
2. HBM demand
Are accelerator generations requiring increasingly larger HBM configurations?
3. Wafer allocation
How much capacity is being redirected toward HBM?
4. Capex
Are Micron, Samsung and SK hynix accelerating investment?
5. Hyperscaler capex
Is AI infrastructure spending still expanding?
6. Inventory
Are customers rebuilding inventory or beginning to destock?
7. Earnings guidance
Are memory companies raising expectations or becoming more cautious?
These indicators will tell us much more than one bullish analyst note.
AI demand is growing faster than traditional semiconductor capacity can be added, while HBM is consuming an increasing share of the industry’s resources.
If that continues, memory becomes more than another component.
It becomes a strategic bottleneck.
And that potentially makes companies exposed to DRAM, HBM and AI storage increasingly important within the broader AI infrastructure buildout.
For my Silicon-to-Substation framework, memory remains one of the layers I want to watch closely.
$MU $SNDK $WDC $NVDA $AMD $INTC $MRVL $CRDO $AAOI $SOXX
The opportunity may be measured in years.
The entries, however, still matter.
Not financial advice.