I keep coming back to the same question with MU.
Everyone is modeling this massive HBM/memory cycle based on AI demand continuing to explode. Fair enough. But what happens if Nvidia’s response to the memory bottleneck isn’t simply *buy more memory*?
There’s already reporting that Nvidia is testing lower-memory configurations for Rubin Ultra. Separately, TrendForce previously reported that Nvidia was only getting enough LPDRAM to cover roughly 60% of its projected SOCAMM requirements and responded by reducing the memory configuration.
That got me thinking.
Back-of-the-napkin math: if you only have enough memory to build 60% of the systems you planned, but you figure out how to cut the memory requirement per system in half, suddenly that same memory supply theoretically supports 120% of your original production target.
Obviously the real world isn’t that clean. There are other bottlenecks, different types of memory, yields, packaging, etc.
But that’s not really my point.
My question is whether this memory shortage is inadvertently forcing Nvidia to figure out how to become **less dependent on HBM per GPU**.
Because if Nvidia can get similar real-world performance using less HBM through better cache management, KV-cache optimization, Vera system memory, NVLink, storage offload, etc., why would they ever go back to stuffing maximum HBM into every GPU once the shortage is over?
Especially if the alternative lets them ship more GPUs.
And this is where I wonder if people are looking at this backwards.
Nvidia doesn’t make money by maximizing the amount of Micron memory attached to each GPU. Nvidia makes money by maximizing the number of extremely expensive GPUs/racks it can ship and the performance customers get from them.
If 288GB of HBM lets Nvidia ship 60 units but 192GB plus better memory/cache management lets them ship 90 or 100, which one do you think Nvidia wants?
Then take it another step.
What if Nvidia gets really good at this?
Cache → HBM → system DRAM/LPDDR → NVMe.
Nvidia already controls the GPU, CPU, NVLink, networking, BlueField, CUDA and increasingly the rack itself. If Nvidia develops the software that intelligently decides what data belongs in each memory tier, isn’t that actually an even **stronger Nvidia moat**?
At that point you’re not buying a GPU with a bunch of memory attached to it. You’re buying Nvidia’s entire proprietary AI computer, and Nvidia decides how much expensive HBM is actually necessary.
Which brings me back to MU.
What exactly do we expect Jensen to say on the next earnings call if somebody asks about the memory bottleneck?
Option A:
*“Yes, memory is constraining how many Rubin systems we can ship.”*
Probably not great for Nvidia initially. But the market can eventually look through that because those GPUs aren’t necessarily cancelled. They’re delayed.
And ironically that’s still not necessarily great for MU’s valuation because now everyone knows Nvidia has a serious incentive to engineer around the bottleneck.
Option B:
*“We’re testing multiple memory configurations that allow us to maintain performance while using less memory.”*
Great answer for Nvidia.
I’m not sure that’s a great answer for Micron.
And then there’s Option C, which would concern me the most as an MU shareholder:
Nvidia basically says the testing is working, software/cache/memory-tiering improvements are allowing them to achieve their targets with materially less HBM, and they intend to carry those lessons into future architectures.
Now you’re not talking about a temporary supply problem anymore.
You’re potentially talking about **lower HBM content per GPU becoming structural**.
Yes, Nvidia could ship so many additional GPUs that total HBM demand still goes up. I’m not arguing HBM suddenly disappears.
I’m questioning the assumptions baked into MU’s future earnings.
If Wall Street is modeling:
more GPUs × more HBM per GPU × higher HBM prices
and Nvidia turns that into:
**WAY more GPUs × LESS HBM per GPU × less dependence on the memory suppliers**
how much of Micron’s future HBM earnings and scarcity premium needs to be rerated?
And there’s another weird part.
If more data gets pushed out of HBM into cheaper memory tiers and fast storage, maybe the loser isn’t “memory” broadly. Maybe the mix just changes. HBM loses some content while LPDDR/DRAM and enterprise NAND pick some of it up.
Which could make this a completely different conversation for MU versus something like SNDK.
So… riddle me this:
**If Nvidia proves it can get close to 100% of its desired GPU output by reducing memory per system instead of waiting for the memory manufacturers to catch up, what should Micron actually be worth?**
Because Nvidia has every financial incentive in the world to solve this permanently.
More GPUs shipped.
Less dependence on three memory suppliers.
Potentially lower BOM.
Potentially higher margins.
Stronger proprietary rack architecture.
Less chance memory screws up Feynman or whatever comes next.
Why *wouldn’t* Nvidia keep investing in that?Maybe I’m missing something.
Asking for a friend. Thanks in advance🤔