Investors debate competitive moat as hyperscalers design proprietary chips
5 Yesterday 2:45 AM · 23h ago · 1 post · 7 comments · 1 source · development 5 of 6
Reddit discussion raises concerns about Micron's long-term competitive position, noting that while industry barriers and thin margins historically limited competition, major hyperscalers including Apple, Amazon AWS, Azure, and Google have begun designing their own custom chips for internal use—either directly or through current manufacturers like TSMC.
“For Micron specifically, yes it's as simple as HBM demand staying high for a long time…This means that keeping compute constant, the only way to scale intelligence is with scarcer models and more VRAM”
Mister__Mediocre, Reddit commenter · reddit ↗
Sanjay Mehrotra Micron CEOMicron Technology Memory chip makerAMD and Intel Server platform makers
The whole story articlespostscomments the bright band is this development · numbered dots are the others · click one to jump
What people said 7 voices · verbatim
-
The barrier to entry in this industry is very high. Previously margins have been thin (even negative at times). Those two combined has helped keep competion out. But despite those factors, several large companies have designed their own chips for internal use - either directly or via current manufacturers or a mix of it. Fx. Apple designed their…
-
For Micron specifically, yes it's as simple as HBM demand staying high for a long time. I don't think it depends at all on supply staying tight. In a previous era, for your application, you'd reach a point of diminishing returns when it came to RAM. That's not the case in the new era. Power is the bottleneck and power use scales with compute. This…
-
Exactly. Some people actually forget that we have used memory products way before AI as well. And a lot of them. With AI in the last 2-3 years, it's just that the amounts became ridiculously high.
-
With agentic and edge ai coming, models will need to remember context, so they'll need to store a fuckton of data. I don't think nand will stop being a bottleneck for years and years to come
-
I won't comment on the NAND side of things since I'm not quite convinced. NAND is something you definitely can have too much of, and hence there is a cliff we'll hit eventually.
-
Recent semi analysis article explains how mixture of experts models uses more kv cache throughput for decode. Although that is already known but it is a good read.
-
Whatever other bull case existed, it has been overshadowed and dwarfed by the AI case.
All 6 developments of Micron's AI Memory Boom Reaches New Heights With… →
NewswiresGoogle NewsReddit