Varun Daiya

building @thinkcompact

[Varun Daiya @varundaiya]

@jimcramer most of the hbm side gets priced months ahead on annual deals, so the beat feels more like plain server ddr5 going up a ton

[Varun Daiya @varundaiya]

@teslayoda lol to be fair saying output is fully committed thru 2029 is kinda sober, a new dram fab takes a couple years before it ships a single good wafer

[Varun Daiya @varundaiya]

@mhdfaran and its paid per token not per model, long context kv cache eats hbm gigabytes way before the compute runs out so every new user is more memory

[Varun Daiya @varundaiya]

@JonahLupton every mi accelerator amd ships carries a pile of hbm stacks, so a good chunk of amds own bom ends up as hynix and samsung revenue anyway

[Varun Daiya @varundaiya]

@MentoviaX hbm4 overtaking 3e that fast makes sense, the 2048 bit interface doubles bandwidth per stack without needing crazy pin speeds

[Varun Daiya @varundaiya]

@firstadopter if power is locked in for 27 the choke point just moves back to packaging, stacks only ship once theres cowos capacity to mount them on

[Varun Daiya @varundaiya]

@jukan05 tpus are stacked with hbm too so whichever chip openai lands on its mostly the same three memory vendors getting paid

[Varun Daiya @varundaiya]

@danielnewmanUV 80% op margin on a memory company still breaks my brain, this used to be the industry that lost money every other downturn

[Varun Daiya @varundaiya]

@DYKA01 mid 20s dram bit growth in one quarter is a lot when node shrinks barely add bits anymore, guess most of it is m15x and 1c wafers coming online

[Varun Daiya @varundaiya]

@tengyanAI five year terms with pricing inside an agreed band is wild for dram, contracts used to reset every quarter

[Varun Daiya @varundaiya]

sk hynix just printed 83T won op profit at an 80% margin, higher than tsmc. this is the memory wall showing up on an income statement. flops per gpu grew way faster than hbm bandwidth for a decade so decode is basically waiting on memory, and the memory guys get to charge for the wait

[Varun Daiya @varundaiya]

@pequityresearch @firesidealpha card spend is a nice early read but i wonder how much api usage slips past it via cloud marketplace billing

[Varun Daiya @varundaiya]

@Prince_Canuma lol fair, backtracking in the trace is fine. it only gets funny when the backtrack lands after the tool call already fired

[Varun Daiya @varundaiya]

@scaling01 curious how long it stays consistent when you walk away and come back, thats where world models usually start quietly redecorating the room

[Varun Daiya @varundaiya]

@PhotonCap hbm stacks keep getting taller and bump pitch keeps shrinking, so a slip in one generation says little about whether hybrid bonding wins long run

[Varun Daiya @varundaiya]

@giffmana perfect hashing over a fixed key set is such a nice trick, lookup ends up a couple multiplies and one compare with no collision handling

[Varun Daiya @varundaiya]

@BenBajarin the annoying part up there is heat not power, you can only get rid of it by radiating so radiator area grows with the cluster

[Varun Daiya @varundaiya]

@tensornestAI In, would love to try!

[Varun Daiya @varundaiya]

@mrm8488 routing is mostly a classification problem so sub billion is plenty if the action space is tight. hard part is getting it to escalate instead of trying the task itself