Let's start this week with a question: What's the most important part of an AI chip? If you say it's the GPU, then you are wrong. More than the GPU, it's the memory in it that holds significance. | The world is experiencing a memory revolution. The formerly boring low-margin business has just evolved into a most profitable segment in the semiconductor industry. We're talking 60% margins on products that used to scrape by at 20%. And only three companies hold the leash on the supply. |
|
| | 5 December Value Traps to Watch Out For | For more than two decades, the semiconductor industry had kept a secret called the "memory wall." | GPU performance exploded 60,000x over two decades. But traditional DRAM bandwidth had improved, but only 100x in the same period. | The gap was manageable. But only when we were just gaming or browsing. The new AI models have a thirst for terabytes of data per second. The GPU is burning electricity, time, and capital, and still not quenching that thirst. | Enter High-Bandwidth Memory (HBM) into the picture. With the Vertical Revolution, HBM is solving the memory wall by completely reimagining memory architecture. | In contrast to the old flat memory chips, HBM is stacking 8-16 memory layers vertically and places them right next to the GPU, with each stack delivering more than 1 TB/s of bandwidth and holding up to 48 GB per stack in HBM3E systems. | HBM4 increases the data lanes by two times, from 1,024 to 2,048, pushing throughput even higher. And it uses 40-50% less power than DDR4 for the same work. | Here's what you need to understand about the current state of things – without HBM, even the most powerful GPU is useless. |
|
| | | | | Amazon Prime Members: Do Not Miss This Bonus | | If you spend a good amount on Amazon, this card could easily put $100s back in your pocket every year. Even better, you could get approved extremely fast — and if you are, you'll receive an insanely valuable welcome bonus deposited straight into your Amazon account for immediate use. | No hoops or extra spending. Get approved, and the bonus is yours. If you shop at Amazon or Whole Foods, this card could help you earn meaningful cash back from purchases you're already making. This offer won't last forever, and for Prime members, it's basically a no-brainer. | Amazon Prime members: See what you could get, no strings attached |
| |
| | |
| Memory Is Now a Hard Ceiling | AI workloads consume memory in three ways: | Model weights (the trained parameters) Intermediate calculations (activations during processing) Training memory (optimizer states)
| It takes 8 GB just to load a single 4 billion parameter model. Jumping to real systems, there is a 30% overhead. With such a huge requirement, "out-of-memory" errors often halt training. | Memory is irreplaceable. Without memory, there is no model, no training, and no data center expansion. |
|
| | The Supply Shock in Four Numbers | Now let's see how extreme the demand shock really is. | HBM demand doubled in 2025 after a 200% jump in 2024 DRAM inventories: 2-4 weeks down from 13-17 weeks earlier DRAM prices: +8-13% in Q4 2025 HBM prices: +13-18% in the same period
| All three major memory suppliers have kept a houseful board on HBM production through 2026. | More than a temporary shortage, the gap is structural. |
|
| | Why You Can't Just Make More | A single HBM stack contains 12 to 16 layers. Each layer must be aligned with perfection and over thousands of microscopic connections must work without flaws. One defect and the entire stack is out for trash. | Yields are also far lower than standard DRAM. Stable yields for HBM3E 12-Hi stacks require at least two quarters of learning. Early HBM4 pilot yields, at around 65%, are terrible by memory standards. | Let's say you went through these monumental tasks and successfully made an HBM. But you are still not done because you need to assemble the HBM stacks with GPUs using CoWoS packaging, which is booked till mid-2026. | The chips exist, but can't be shipped without packaging slots. |
|
| | The Three-Company Oligopoly | At the moment, the entire HBM sector is dominated by three companies. | SK Hynix: 62 to 64% market share | HBM represents 42% of its DRAM revenue Estimated HBM revenue is around $20.7 billion HBM margins: ~60% Primary supplier for NVIDIA H100 and H200
| Micron: 21% market share | Went up from ~5% in 2024 to 21% in Q2 2025 HBM now 23% of Micron's DRAM revenue Quarterly HBM revenue stands at $2 billion Exited consumer memory, pulling 25% of global DRAM supply from low-margin products
| Samsung: 15 to 17% market share | Making heavy investments to catch up with its competitors HBM revenue: ~$9 billion Planning more than $20 billion annual investment Targets 30% or more of NVIDIA's HBM4 volume in 2026
| Together: ~95% of global DRAM, nearly 100% of HBM. That's your oligopoly. |
|
| | The Pricing Power That Changed Everything | HBM margins close to 60%, and standard DRAM margins are climbing up toward 40%. | With HBM accounting for over 30% of total DRAM market value despite being only ~10% of bit volume, the companies are basically selling less product for way more money. | It keeps getting interesting. Micron reported a 162% quarter-over-quarter price increase in some segments. | Beyond a memory company, the chips business are cash printing machine now. |
|
| | The Bottom Line | Memory has become both the fuel that powers AI and a machine that produces a margin that the memory industry hasn't seen in decades. The companies controlling the supply of HBM will remain at the core of the global tech industry, as long as AI demand keeps growing. | To make the most out of it, here's your investment thesis in one sentence: In the AI race, the fastest chef still waits for ingredients. | And right now, only three companies control the kitchen. |
|
| | | | | Important disclosures: This newsletter is provided for informational purposes only and does not constitute investment advice. All investments involve risk, including possible loss of principal. Please consult with your financial advisor before making investment decisions. |
| |
| | |
|
|
No comments:
Post a Comment