Tim, great piece — the harvest read feels right. The wrinkle I keep wrestling with is the LCD endgame parallel. Korean displaymakers exited LCD on similar logic, and a decade later the Chinese cartel is setting panel prices for Samsung's and LG's TV businesses and funding the OLED attack. CXMT is already sampling HBM3 to Huawei, targeting HBM3E in 2027, and building a Shanghai fab 2-3x the size of Hefei — even if their HBM only ever satisfies Chinese demand, that erases the China revenue line for the Koreans and funnels HBM-margin cash into the same subsidy-fueled global expansion that took LCD from Korea.
The other piece I can't square is whether the "new floor" of LPDDR5X SOCAMM2, server DDR5, and GDDR7 is really a floor at all. Past cycles were cushioned by demand spread across billions of devices and thousands of OEMs, but today's volume comes from maybe five hyperscaler wallets that also drive HBM — so every tier corrects together when AI capex eventually digests. The one thing that might make the harvest genuinely rational is the 3-5 year LTAs Samsung is now pushing.
Curious whether you think those LTAs are durable enough to make the LCD scenario second-order, or whether they just delay the reckoning.
I’m not sure how it’ll play out. I do think this AI boom is a structural change. But I also feel we’re yet to see how the end-device (client) scenario changes going forward. The dot-com boom changed many things under the hood, leading not only to fiber everywhere but a viable market for consumer laptops, and eventually mobile internet and finally smartphones. As a result, a smartphone chip can now power a laptop — and that laptop has a use. This was inconceivable before the dot-com infra roll out.
So I think that’s the lens through which we should look at AI: what client/edge uses will come, and what flavors of memory will follow.
They're not doing this because they're evil. They're doing this because they can't predict if the current demand will reliably stay the same. The RAM industry has gone through multiple boom and bust cycles.
It's definitely true that advent of AI/LLM's represent a permanent qualitative change to society/economy, but to echo Yury the memory suppliers have deep muscle memory, scar tissues, whatever the best analogy is for the volatile boom/bust cycles which did as much to push other (mostly Japanese/EU) suppliers out of the market. They can invest now to expand capacity and then in 2 years when the capacity comes online they may find customers backing away from previous promises to help the investment pay off.
Furthermore, the most visible "end customers" for their own end-customers are the not-so-confidence-inspiring trio of Sam Altman ("podcast bro", as TSMC execs reportedly called him after Altman laughably proposed trillions for fab investments), Masayoshi Son, and Elon Musk.
With the move to IPO, we now have a much better measure of the WeWork-levels of incinerated cash by OpenAi and Anthropic (and most assuredly the smaller players as well). From my own personal experience, people who make a living from the manipulation of atoms are not as prone to the messianic fervor that often afflicts entrepreneurs in bits - note that the "tech evangelist" title is much less common at true hardware technology companies (they seem to be much more common in software and related internet/mobile, or consumer products, not stuff like IC's, as it is difficult to "evangelize" with a straight face about products that basically amount to applied material science)
Tim, great piece — the harvest read feels right. The wrinkle I keep wrestling with is the LCD endgame parallel. Korean displaymakers exited LCD on similar logic, and a decade later the Chinese cartel is setting panel prices for Samsung's and LG's TV businesses and funding the OLED attack. CXMT is already sampling HBM3 to Huawei, targeting HBM3E in 2027, and building a Shanghai fab 2-3x the size of Hefei — even if their HBM only ever satisfies Chinese demand, that erases the China revenue line for the Koreans and funnels HBM-margin cash into the same subsidy-fueled global expansion that took LCD from Korea.
The other piece I can't square is whether the "new floor" of LPDDR5X SOCAMM2, server DDR5, and GDDR7 is really a floor at all. Past cycles were cushioned by demand spread across billions of devices and thousands of OEMs, but today's volume comes from maybe five hyperscaler wallets that also drive HBM — so every tier corrects together when AI capex eventually digests. The one thing that might make the harvest genuinely rational is the 3-5 year LTAs Samsung is now pushing.
Curious whether you think those LTAs are durable enough to make the LCD scenario second-order, or whether they just delay the reckoning.
Good analysis 👍
I’m not sure how it’ll play out. I do think this AI boom is a structural change. But I also feel we’re yet to see how the end-device (client) scenario changes going forward. The dot-com boom changed many things under the hood, leading not only to fiber everywhere but a viable market for consumer laptops, and eventually mobile internet and finally smartphones. As a result, a smartphone chip can now power a laptop — and that laptop has a use. This was inconceivable before the dot-com infra roll out.
So I think that’s the lens through which we should look at AI: what client/edge uses will come, and what flavors of memory will follow.
Great color, thanks Tim. - Brad
They're not doing this because they're evil. They're doing this because they can't predict if the current demand will reliably stay the same. The RAM industry has gone through multiple boom and bust cycles.
It's definitely true that advent of AI/LLM's represent a permanent qualitative change to society/economy, but to echo Yury the memory suppliers have deep muscle memory, scar tissues, whatever the best analogy is for the volatile boom/bust cycles which did as much to push other (mostly Japanese/EU) suppliers out of the market. They can invest now to expand capacity and then in 2 years when the capacity comes online they may find customers backing away from previous promises to help the investment pay off.
Furthermore, the most visible "end customers" for their own end-customers are the not-so-confidence-inspiring trio of Sam Altman ("podcast bro", as TSMC execs reportedly called him after Altman laughably proposed trillions for fab investments), Masayoshi Son, and Elon Musk.
With the move to IPO, we now have a much better measure of the WeWork-levels of incinerated cash by OpenAi and Anthropic (and most assuredly the smaller players as well). From my own personal experience, people who make a living from the manipulation of atoms are not as prone to the messianic fervor that often afflicts entrepreneurs in bits - note that the "tech evangelist" title is much less common at true hardware technology companies (they seem to be much more common in software and related internet/mobile, or consumer products, not stuff like IC's, as it is difficult to "evangelize" with a straight face about products that basically amount to applied material science)