All of a sudden HBM is the leading topic of conversation. Why is that? A good number of people ask the Memory Guy that question.
Sure, the recent explosion of generative AI has been the key part of that interest, but there was even more that encouraged competition into this market at an accelerated pace. The simple answer is that HBM, a DRAM used in AI applications, remained profitable at a time when mainstream DRAM was losing money. Suddenly everyone wants to participate in this consistently-profitable market—wouldn’t you?
Here’s how that situation evolved: For the past couple of years SK hynix was the world’s only HBM supplier, and its two major competitors, Samsung and Micron, were happy to let SK hynix have that market. From their perspective, HBM was a tiny niche that was expensive to service, and greater returns could be made on standard DRAM than on HBM.
Then the DRAM market collapsed in 2022, and DRAM makers found themselves with excess production capacity and significant operating losses. The collapse was devastating, with some prices dropping almost 70% over its course, but one company, SK hynix, was faring better than its competitors. The reason for this difference was that SK hynix was selling a product whose price didn’t drop, or perhaps it dropped a little, but it remained profitable throughout the downturn.
Since then, DRAM revenues have been increasing, due in a large part to the generative AI interest surge. Hyperscale datacenters are spending heavily on the high-performance computers that are required to perform AI training on all the Internet data they can access. That growth is evident in the standard Objective Analysis market share chart shown below:
2Q24 revenues, at $23.1 billion, have risen almost as high as those of 2Q22 ($23.5B), after the market hit its 1Q23 low of $9.1 billion.
What is not obvious from this chart, though, is that SK hynix’ market share was increasing a time when Samsung’s and Micron’s shares were falling. How can you tell that? Have a close look at 2Q23 & 3Q23 of the market share chart below. This is the same data that was presented above, but shown as percentage of the market, rather than as absolute revenue numbers.
What’s that big SK hynix peak in 2Q23-3Q23? It’s a time when the market was doing very poorly, but SK hynix revenues weren’t staying in the trough as long as Samsung’s and Micron’s, whose market shares dipped at that time.
Note that Samsung regained its lost share after 3Q23, but Micron did not. Micron has been slow to enter the HBM market while Samsung responded more rapidly. Micron is hoping to reclaim its market share soon.
For those unfamiliar with HBM, it’s basically stacked DRAM chips that are largely used to accelerate high-end GPUs and other AI processors, the kind of chips that are mainly used for large language model training, but are also found in networking and high-performance computing (HPC) applications. Large language model training is the foundation for generative AI models. When you stack DRAM and connect it in close proximity to a processor, actually within the processor’s package, you can run it at a somewhat faster speed, increase the connection count for many times more bandwidth, and reduce power consumption, but it cost an awful lot to do that.
This is all explained in great detail an upcoming Objective Analysis HBM report, Will HBM Rescue the DRAM Market? This report will be available for immediate download in mid October. The report explains the technology, its market, and provides an HBM revenue forecast.
Objective Analysis monitors semiconductor markets in great depth, and provides insights that escape many market observers. Our regular clients use these insights to their advantage, and their performance testifies to that. If your company would like to explore ways that we can help you to thrive, please contact us and open a conversation.