What is HBM?
HBM
High Bandwidth Memory (HBM) is a type of computer memory that vertically stacks DRAM dies to achieve higher bandwidth and lower power consumption, used in high-performance computing and AI accelerators.
How it developed
- Aug 2026 - Jon Y said the design philosophy behind HBM is that many slower channels will beat a few fast ones.
- Aug 2026 - Jon Y said AMD approached SK hynix to co-produce a new industry standard architecture, leading to HBM’s development.
- Aug 2026 - Jon Y said SK hynix had to co-develop everything from scratch, including molding equipment with Japan’s Towa and EMC material with NAMICS, which required some fifty iterations.
- Jul 2026 - Ben Pouladian said that in HBM development, adding an extra layer beyond the highest of 14 was too hard, as aligning everything became difficult.
- May 2026 - Andrew Feldman said Micron was producing numbers with 80-85% gross margins, indicating high profitability in the HBM market.
- Aug 2026 - Jon Y said with supply booked out for the next two years and surging demand for more intelligent models, more memory fabs are needed.
In the evidence
Every line below is attributed to a named speaker.
SK hynix required approximately 50 development iterations of EMC material with NAMICS just to enable MR-MUF HBM packaging.
“The EMC material was developed in partnership with the Japanese company NAMICS, and required some fifty iterations.”Jon Y · 13 Aug 2026
HBM achieves higher total bandwidth by stacking many wide, slower DRAM channels rather than pushing a few channels to extreme speed.
“The design philosophy is that many slower channels will beat a few fast ones.”Jon Y · 13 Aug 2026
HBM was originated by AMD approaching SK hynix, not by Nvidia, despite Nvidia dominating the AI GPU narrative today.
“This led AMD to approach SK hynix to co-produce a new industry standard architecture.”Jon Y · 13 Aug 2026
Towa (Japan): supplier of molding equipment co-developed with SK hynix specifically for MR-MUF HBM packaging, a critical and proprietary process step.
“Since no equipment or material then existed for it, SK hynix had to co-develop everything from scratch. The molding equipment, for example, was produced with Japan's Towa. The EMC material was developed in partnership with the Japanese company NAMICS, and required some fifty iterations.”Jon Y · 13 Aug 2026
NAMICS (Japan): supplier of EMC underfill material co-developed with SK hynix for HBM MR-MUF packaging, representing a hard-to-replicate materials partnership.
“Since no equipment or material then existed for it, SK hynix had to co-develop everything from scratch. The molding equipment, for example, was produced with Japan's Towa. The EMC material was developed in partnership with the Japanese company NAMICS, and required some fifty iterations.”Jon Y · 13 Aug 2026
Micron is posting 80 to 85 percent gross margins on HBM memory, signaling severe AI-driven supply constraint.
“Micron producing numbers where they have 80 85% gross margins.”Andrew Feldman · 26 May 2026
HBM design philosophy in one line: many slow channels beat a few fast ones.
“The design philosophy is that many slower channels will beat a few fast ones.”Jon Y · 13 Aug 2026
HBM memory stacking is currently capped at 14 layers after attempts to reach 16 layers failed due to alignment difficulty.
“They wanted to go up I think highest is like 14. They wanted to go to 16 and they weren't able to do it. Adding that extra layer was just too hard aligning everything.”Ben Pouladian · 14 Jul 2026
HBM memory supply is reportedly booked out for the next two years as of the episode date (August 2026).
“With supply booked out for the next two years - and end demand for more intelligent models continuing to surge - more memory fabs are needed.”Jon Y · 13 Aug 2026