Editing Openai/6960ed20-8dfc-8005-a089-1bb9023ccffa (section)

=== Assistant: You’re right about the pressure on memory manufacturing, but the NVIDIA “no VRAM included” part is not actually true. ===
You’re right about the pressure on memory manufacturing, but the NVIDIA “no VRAM included” part is not actually true. What you’re reacting to is a real trend, but it’s being exaggerated in online discussions. Let me separate what’s correct, what’s partially correct, and what’s incorrect, and then explain where this is actually heading.

==== ### ====

This part is absolutely true:
* HBM uses the same DRAM wafers as DDR4/DDR5
* The difference is in: - stacking - TSVs (through-silicon vias) - advanced packaging

Because HBM margins are MUCH higher, companies like:
* SK Hynix
* Samsung
* Micron

have diverted a large share of DRAM capacity to HBM.

So yes:

: 

That does tighten supply for conventional DRAM.

==== This is the real bottleneck most people miss. ====

===== - costs $15–25 billion =====
* takes 4–6 years
* requires extreme-UV tools (ASML bottleneck)
* risks massive oversupply if demand cools

HBM demand might peak around 2026–27 —
so manufacturers are very cautious about overbuilding.

They prefer:
* retooling existing fabs
* upgrading nodes
* squeezing more out of packaging

Not greenfield fabs.

==== This part is incorrect as stated, but I understand where the confusion comes from. ====

===== That is not happening for consumer or professional GPUs. =====

What is happening:

===== - NVIDIA does not manufacture HBM =====
* It buys HBM from SK Hynix / Samsung / Micron
* HBM supply is now the hard limit on GPU shipments

So NVIDIA is:
* prioritizing data-center GPUs
* negotiating memory separately and more aggressively
* allowing some partners flexibility in sourcing GDDR (not HBM)

For consumer GPUs (RTX series):
* ASUS / MSI already source GDDR themselves
* This has been the case for years
* NVIDIA only supplies the GPU die

So that part isn’t new — it’s just becoming more visible.

==== This shortage is different from crypto or COVID shortages. ====

===== - demand spike =====
* then collapse
* memory prices crash

===== - AI training = persistent, long-duration demand =====
* Governments + hyperscalers involved
* Supply chain is structurally constrained, not just cyclic

HBM is the first time where:

: 

That’s historically unusual.

==== NVIDIA’s problem is not GPUs — it’s memory per GPU. ====

Example:
* H100 = ~80 GB HBM
* H200 = ~141 GB HBM
* Blackwell = even more

So even if NVIDIA can make more GPU dies:
* no HBM = no shippable GPU

This is why:
* GPU launches slip
* allocations happen
* smaller customers get delayed

==== No — and this is important. ====

What will happen instead:

===== - DDR prices volatile =====
* Laptop / PC RAM upgrades expensive
* Capacity tight

===== - HBM yields improve =====
* Packaging throughput increases
* DRAM fabs rebalance again

Memory companies want DDR to exist — it’s stable, predictable revenue.

==== The industry is already reacting: ====
* More efficient models (less memory per token)
* Inference optimization
* On-package SRAM / cache
* Disaggregated memory architectures
* HBM alternatives (HBM-lite, LP-HBM, stacked GDDR)

This is exactly how previous bottlenecks were solved.

==== You are right about: ====
* DRAM → HBM reallocation
* Insufficient fab capacity
* Memory becoming the choke point

But:
* ❌ NVIDIA is not shipping GPUs without VRAM
* ❌ Consumer GPU VRAM sourcing hasn’t fundamentally changed
* ❌ This does not mean permanent DRAM starvation

===== : =====

If you want, I can also explain why HBM is so insanely hard to manufacture compared to normal RAM — that’s where most intuition breaks.