theconductor

@aleabitoreddit Is this not just impact from Yunnan’s capacity ramp?

If you’re curious about $AXTI: It’s down 21% on the new potential dilution news. Board wanted to add 50M more shares (up to $2B worth to dilute) for shareholders vote in the 14th. 70m -> 120m shares. I say this about $IREN excessive $6B dilution and it’s the same with AXT price proposal that I hold. I would not “trust in management” to use it wisely and the fact it’s filed is a red flag. That being said: we’ll see what happens on the 14th. If it passes for 50m more share authorized dilution, I personally would not hold. But it fails, I expect a recovery.

@ContrarianCurse Not saying numbers in the out-quarters don't move higher, nor am I making a through-cycle/terminal value claim here, but it was a pretty obvious short post-EPS.

I can easily still sketch you a $MU bull case, and r/r is meaningfully improving here But also, you guys are at the “coordinated put/short seller” point of the investor base sssooooo

@bubbleboi They have enough LNG through March - even then there are coal fired plants ready to meet a grid deficit. On helium, they have 6 months worth of inventory. Tell me you don’t do channel checks without telling me you don’t do channel checks.

I’m massively short TSMC btw. They will ship zero chips next month if this trend continues.

@bubbleboi Is this not BiDi serdes?

400G per lane? Fuck are they using PAM9?

I’d like to formally apologize for shorting nvda at 15x and owning the pcb drill bit names. Also, we need more activists in semis, I want to see these OLED fab cleanrooms acquired and sold to MU

Bear pitch is -> Out year revenues contingent on reliability qual by a company with notorious reputation in this domain. Losing share in DSP and the ASIC franchise SAM eroding by day. Decent numbers from Matt last Q, index add def helped (hurt them today), think name just got squeezed. Search for 2nd derivs is real :)

#1 stock people ask me about is $MRVL. Funny situation... I am telling everyone Celestial is amazing and will be bigger than InPhi IF AND ONLY IF they pass GR-468 reliability. But (almost) nobody beleives me lol. Too much focus on Alchip and general dislike of Matt Murphy.

well, lots of things happening. some thoughts: 1. Helium is fine, I think mostly fearmongering atm, could become a legit issue if trade routes continue to get clogged (Taiwan only has so many ports), but if you do the math from wafer starts into required gas helium volumes (think this mostly used in CVD process) should be fine. 2. LNG probably larger issue here, though there is a decent amount of mothballed capacity in Taiwan (few coal plants and some private oil-fired plants). Haven't done the full math but speculate anywhere from 10-30% of grid capacity could be recuperated in an LNG fueled deficit. 3. Speculate that non-EU/(primarily) non-western bloc countries would likely step in to prevent TSMC chips going bye bye. Taiwan has until march (supposedly) before needing to tap strategic reserves, though I believe it won't come to this - otherwise stonk down. 4. US probably does a ground campaign in Iran, only way to truly mitigate any further Iranian missile strikes (see the Houthi mobile launchers lol). Gerald ford going to red sea, so let's see - when we fought in Panama first thing we did was secure the canal, so lets hope Trump admin has same idea.

@ContrarianCurse Gary D if you can hear us. Please Gary D we need you

$AMAT you give me this bitch back at $270 and I will back up the truck.

thoughts for today after a crazy tape: 1. rip my besi bags: market will underwrite one lever and only that lever, in this case its the HBM/hybrid bonding thesis. The JEDEC/Trendforce reporting on narrowing the layer gaps in the stack is cool, but I suspect the underfill process becomes increasingly difficult. Market will price the narrative faster than the fundamental, that sucks - but if you can accurately attribute the variance to the narrative, then you could have a great opportunity on your hands (see the CRDO/ALAB/Optical spread as of recent). Besi does more than just hybrid bonding, and it isn't solely a levered semicap-HBM deriv play. RIP 2. Bye bye memory (HBM) at GTC? I get the fears around SRAM, and note there was a design getting spread around last dec showing LPU dies/SRAM banks stacked on the Feynman die. (Similar idea to AMD 3D V-Cache). Some thoughts - I suspect the timeline to deciding the spec/design for feynman wrt whether or not we get 3D stacked LPU/SRAM on GPU is closing, potentially contingent on improvements in hybrid bonding systems (thermal is main consideration, also a really good @IanCutress vid on AMD 3D V-Cache u guys should watch). I mean, HBM could just be off die/substrate but on tray, or if you accompany a Feynman die with SRAM bank stacked on die with LPDDR on tray (similar to some ASIC designs out there fwiw). Fears on the basis of the LPX/Groq rack, I understand it - but note there are practical limitations to SRAM, and I suspect an LPX rack design could be complemented by an adjacent DRAM cabinet. (LPUs are good, but not for high throughput, mostly for deterministic workloads that are low tail latency). If HBM out and conventional DRAM in, is that really net negative? Will let u make ur own decision around that. @insane_analyst writing abt the GTC singularity, go check him out. 3. You can always count on Doomberg to fuck everything in last 30min of trading day. Okay, a 600MW expansion getting picked up by Meta bc of ORCL 3P financing issues or OAI potentially wanting a different cluster design/hardware vendor, idek man... 4. In a high SI as (%) of float name with an earnings re-rate catalyst/strong out-of-consensus beat on #s, you should still have a strong, distinguishable reason for why shorts will capitulate and whether L/O will pick the name up 5. No one really knows Apple premium handset share in China lol 6. Count on spec sales to not answer your Qs.

@ElliotL45 @ContrarianCurse @IanCutress Yes but is that not what NVDA is doing with HBM4, or at least rumored to be doing (perf + stability SKU)?...Also I think deviating from JEDEC is fine as long as there exists a program willing to design you in. Imo the standardization = scaling paradigm no longer holds

@neverlongqcom @ContrarianCurse @IanCutress I hear you on the speeds, vendors definitely deviate to flex performance. But for packaging mechanics, physical specs are the law. You can't dual-source HBM if everyone has a different Z-height or bump pitch

Oh $BESI ouch hahaha you fucker Whole market gunna get porked. Doubt my long save me this time, I was holding in nicely

@ElliotL45 @ContrarianCurse Old as in we already had headlines around the manufacturers trying the gap narrowing (see trend force earlier in the week. Also why are the planned HBM pin speeds not at the JEDEC standard? following JEDEC is not mandatory? Would suggest you see @IanCutress recent note on this.

@neverlongqcom @ContrarianCurse Wdym it’s old? These specs are defined by the memory makers and the customers themselves. HBM is a commodity, so following JEDEC is mandatory for interoperability. Only custom HBM can afford to deviate. No spec, no scale.

@ContrarianCurse lotta shorts + index add, ppl forced to capitulate

This is such a max pain semis tape lol $NVDA beats by a mile and makes more money than god and goes down. Kicked in the dick on some bogus license headline $MRVL barely beats +10% lollll

Semianalysis right on time with their CPO note lol

what is the latency difference between AEC on 400G vs running 400G DAC with conditioning on the switch chip? Although I guess its highly length dependent, at <3m I suspect DAC has a better TCO ,but >3m even if you can run DAC there's some utility to having the link reliability in spite of any retimer latency added...or ig this is purely a cost consideration argument?

@neverlongqcom It’s there, active silicon is on networking chip handling fec, eq, etc

Nothing surprising that Hock says copper for scale up will work. He has been of this stance for a while. I've stated in a newsletter post a couple of months ago that copper for 1.6T will be alive and well. Even though Hock calls it DAC, the switch silicon will perform equalization and FEC -- there is simply no way without it. So the DAC is actually AEC as well -- just that AEC is $LITE's marketing term. I'm skepical of 3.2T for copper scale up, but with $AVGO's SERDES prowess, anything is possible at this point. The physics makes the engineering hard, not impossible. open.substack.com/pub/viksnewsle…

@ContrarianCurse Ur just solving for elasticity driven unit TAM decrease > hyperscaler excess memory demand Also agree there’s an element of “decommodification” here

The reason cycles happen is because 1) the mismatch in how fast demand can materialize 2) lead times to meet the price signals Tech uniquely, almost more than any other industry can bring on demand extraordinarily fast. Prices rise for commodity inputs and supply responds. Now during these periods, lead times expand. Prices increase quickly. People act rationally because they see such incredible demand. You don't want to be the reason you couldn't sell more compute or another unit of laptop. So you order as much as you can (remember your signals are SUPER STRONG and prices are rising) so you don't get fired. A natural consequence is that lead times expand. By definition - your visibility into demand gets worse. Higher conviction into demand 3 months out vs. 18. Eventually demand settles into a flatter curve and these chunky capacity adds surpass and catch demand. All the behavior unwinds together. Now you have very long-term order / commitments without the same demand you saw. Inventories are building internally. Also you locked in higher prices and can feel the pressure to not compound the mistake. So you stop ordering Now, couple qualifiers here. 1) If your product is demonstrably not a commodity - AKA my Hynix memory delivers better speed than Samsung/MU and that transfers into TCO and pricing - then yes this would be different. I don't think ultimately it is. 2) Just because this demand thrust is unusually large, does not mean these dynamics are not true 3) What makes cycles unavoidable is that everyone is acting perfectly rationally along the way. Things like LTAs / demand visibility etc just end up warping the market forces like any other futile central control of a market mechanism. Ask Sanghi about LTAs What is probably different about this cycle is that I agree with you that demand for compute is insatiable. But that has always been the case across history. It is super easy to give away 1$ of compute for 70c. So this downcycle likely won't be a lack of compute demand, it will probably come from a capital cycle - one of the big buyers get pushed off of their plans due to poor economics and the poor business model of the labs/renting out GPU capacity (its going to be the most subscale ones). Then it will consolidate to the lowest cost producers in the space (the big boys + custom chip program kicks/profitable core biz). Also, as I've said before, memory will balance in the same way $150 oil would balance. Demand dries up in the most elastic places - in this case its consumer electronics.

@vikramskr Groq is capped at 120b param models iirc

HBM demand is unaffected by SRAM. Meta estimates only 10% of their inference to need low latency atm. Long context is still HBM territory, and is not going away anytime soon. Discussed here: open.substack.com/pub/viksnewsle…

lol Stacy saving the day with his AVGO GW math @Srasgon my king