This is actually a very important debate for the entire AI semiconductor supply chain, not just memory stocks like

Micron Technology,

SanDisk,

Western Digital, and

Seagate Technology.

The key question is simple but very powerful:

> Does AI efficiency reduce hardware demand, or does it increase total usage?

Historically in tech, the answer is usually the second one.

---

What TurboQuant actually affects

From what analysts are saying, TurboQuant mainly:

Optimises KV cache

Improves inference efficiency

Reduces memory per query

Does NOT reduce training memory

Does NOT reduce HBM demand significantly

Mostly affects inference VRAM / system memory

So Morgan Stanley’s view makes sense: HBM (used in training GPUs) should not be heavily affected.

This means companies most exposed to HBM and AI training, especially Micron Technology, may actually be less impacted than the market fears.

---

The Jevons Paradox (very important here)

There is a famous economic concept called Jevons Paradox:

> When technology becomes more efficient, total consumption often increases, not decreases.

Examples:

More fuel-efficient cars → people drive more

Cheaper cloud computing → more software

Faster GPUs → more AI models

Cheaper storage → more data stored

So if inference becomes cheaper:

More AI agents

More queries

More applications

More edge AI devices

More data generation

More storage demand

More memory demand overall

So efficiency may increase total memory demand, not reduce it.

---

Memory demand actually comes from 3 different areas

Very important to separate:

Segment Memory Type Companies

AI Training HBM Micron

AI Inference DRAM Micron

Data Storage NAND / HDD SanDisk, WDC, Seagate

Data Centres SSD / HDD WDC, Seagate

Edge Devices NAND SanDisk

TurboQuant mainly affects inference memory efficiency, not training and not storage demand from data growth.

AI still generates massive data:

Logs

Video

Synthetic data

Training datasets

Model checkpoints

Agent memory

Enterprise data lakes

All these need storage, not just VRAM.

So companies like:

Western Digital

Seagate Technology

SanDisk

are more tied to data growth, not inference efficiency.

---

My view: Overreaction vs First Crack

I would frame it like this:

Short term

This could be:

Expectations too high

Capex very aggressive

Memory stocks ran too much

Any negative narrative triggers selloff

So short term, this looks more like positioning unwind / sentiment shift.

Long term

The real risks to memory demand are actually:

1. Custom AI chips with on-chip memory

2. Better model compression

3. More efficient architectures

4. Slower AI spending if economy weakens

Not just TurboQuant alone.

---

Big picture conclusion

I would summarise the situation like this:

If AI demand slows → memory stocks fall.

If AI becomes more efficient → AI spreads everywhere → memory demand increases.

So paradoxically:

> The biggest risk to memory is not efficiency.

The biggest risk is AI capex slowdown.

Personally, I would currently lean toward: This looks more like a scare than the end of the AI memory cycle.

The real thing to watch is:

HBM pricing

Nvidia shipments

Hyperscaler capex

NAND prices

Data centre buildouts

If those remain strong, then this drop may just be noise, not a structural crack.