My research on Memory Makers and AI to better understand where the business is going

HBM is High Bandwidth Memory and what is being used in the AI chips.  HBM sits next to the GPU.  Currently there are only three companies that can produce HBM.

 

Sk Hynix – They are the market leader.  Largest HBM supplier.  They are the volume leader of HBM3E and HBM4

 

Samsung Electronics – They are a major supplier of HBM3E and HBM4.  They will have a larger share of HBM4

 

Micron – They are the fastest growing supplier.  They produce HBM3E and in volume production of HBM4 for Nvidia Rubin platform

 

 

HBM suppliers use stacks as a measurement for a GPU.  NVIDIA B200 has 8 HBM3E stacks. Each stack holds 24 GB.  Total memory is 8X24 GB for 192 GB.  The thing to focus on is how many stacks each GPU will use. 

 

Nvidia chips

H100 – 5 stacks

H200 – 6 stacks

B200 – 8 stacks

GB 300 – 8 stacks

Rubin  – 8 stacks

Rubin Ultra – 16 stacks

Feynman is the next generation and based on progression it will probably be 16 stacks or more

 

AMD chips

MI300A – 8 stacks

MI300X – 8 stacks

MI325X – 8 stacks

MI350X – 8 stacks

MI400 is next generation and is 12 stacks

 

Intel Chips

Gaudi 2 – 6 stacks

Gaudi 3 – 8 stacks

Falcon Shores is next, but no announcement on stacks

 

Amazon has Tranium, Tranium2,Tranium3, and Inferentia2 that use HBM, but they don’t disclose amount of stacks

 

Google chips

TPU v4 – 4 HBM stacks

TPU v5e not disclosed

TPU v5p not disclosed

Ironwood TPU – 6 stacks

 

Microsoft is working on Maia 100, but has not disclosed how much HBM it will use

 

Meta is currently working on MTIA 300/400/450/500 that wil use HBM.  They could use 4-8 HBM stacks but that is all speculation

 

Other companies of note that could be using HBM or invest in infrastructure are Apple, Broadcom, OpenAi, Tesla, xAI, Marvell, Alibaba,ByteDance, Baidu, Huawei, and IBM.

 

From the examples above it shows that HBM is being used in greater amounts per generation of GPU

 

 

Future uses for HBM besides data centers are humanoid robots, autonomous vehicles and industrial robots.  The majority of industrial robots won’t be using HBM. We probably won’t see this until 2027 at the earliest.

 

The global HBM stack demand forecast:

2026 ~20-30 million stacks

2027 ~35-50 million stacks

2028 ~55-75 million stacks

2029 ~80-100 million stacks

2030 ~100-150 million stacks

 

If chipmakers can produce more energy efficient chips in the future that also means they will replace older generations creating demand later. 

 

From an investor perspective I am very bullish.  Every hyperscaler earnings call has basically said their business is growing and they are capacity constrained.

 

Google Cloud had a revenue growth rate of 82% YOY

Azure was 43%

AWS was 37%

 

I don’t see any reason not to stay invested at this time.  Ai Infrastructure spending is turning into cloud revenue growth.  It is validating the spending which supports continued HBM demand.  Based on the earnings calls Microsoft, Amazon, and Google are making the best return on investment.  Meta is the only hyperscaler that really isn’t showing the best return on investment.  They could slow their Capex spending, but the other hyperscalers could step in for that demand if they continue to boost their revenue.