An observation about your point on memory and its extended need. Better/faster C/GPU always needs better/faster memory and better faster memory always allows for better/faster C/GPU. So they are married and investing in best in class companies in both, or those that do both, always a good bet.
The timing of this conversation is impeccable with TurboQuant dropping this week and rattling memory stocks. Val's point about the memory cycle being driven by AI demand still holds but the nuance that is getting lost in the sell-off is the difference between capacity vs. bandwidth demand.
TurboQuant compresses KV caches 6x. Sounds like a memory killer. But in practice, every team running long-context inference will use that compression to fit longer sequences into the same HBM footprint, not to buy less HBM. The constraint shifts from "how much memory" to "how fast can I fill it," which actually increases the premium on high-bandwidth memory, not decreases it.
I would be willing to bet that the memory cycle thesis Doug and Val laid out here is deep enough that it survives TurboQuant just fine.
hey quick question - i'm finally leaving substack as a consumer. i have an embarrassing number of subscriptions but between the cozying up to nazis and then this latest polymarket nonsense it's way past time i stopped giving them money -- but i don't want to stop giving YOU money. have you thought about moving to a better platform? there are a bunch out there that make it very easy (and i think more profitable for you as well, unless substack is offering you specific incentives which they do sometimes to retain talent)
An observation about your point on memory and its extended need. Better/faster C/GPU always needs better/faster memory and better faster memory always allows for better/faster C/GPU. So they are married and investing in best in class companies in both, or those that do both, always a good bet.
The timing of this conversation is impeccable with TurboQuant dropping this week and rattling memory stocks. Val's point about the memory cycle being driven by AI demand still holds but the nuance that is getting lost in the sell-off is the difference between capacity vs. bandwidth demand.
TurboQuant compresses KV caches 6x. Sounds like a memory killer. But in practice, every team running long-context inference will use that compression to fit longer sequences into the same HBM footprint, not to buy less HBM. The constraint shifts from "how much memory" to "how fast can I fill it," which actually increases the premium on high-bandwidth memory, not decreases it.
I would be willing to bet that the memory cycle thesis Doug and Val laid out here is deep enough that it survives TurboQuant just fine.
hey quick question - i'm finally leaving substack as a consumer. i have an embarrassing number of subscriptions but between the cozying up to nazis and then this latest polymarket nonsense it's way past time i stopped giving them money -- but i don't want to stop giving YOU money. have you thought about moving to a better platform? there are a bunch out there that make it very easy (and i think more profitable for you as well, unless substack is offering you specific incentives which they do sometimes to retain talent)
here’s an example for ghost: https://www.alexhyett.com/newsletter/substack-to-ghost/
and for beehiiv: https://www.fingers.email/p/fingers-has-moved-platforms-here-s-why