Semiconductors can be a hard industry to understand, especially for investors without a technical background.
Here are a few of the books that really helped me understand semiconductors better. In order of approachability - a 🧵
LK-99 is real. Simulation papers are dropping. replication by a Chinese scientist with video, wow just wow. We are probably a decade away but congrats everyone we just walked through a keyhole for humanity. Things will speed up now.
The things that have been mentioned that will be impacted if it can be manufactured:
-power transmission
-rail guns
-maglev
-quantum computing? Idk this
-fusion, stronger magnets means we shape it better
-probably a lot of other stuff
yeah this is so exciting and huge
It’s rare I write something I’m so proud of. Todays post is about why there’s going to be a new Moore’s law at the DC level, and energy densification by liquid cooling is a new vector to compete in. And surprise, NVDA is ahead.
I’m very proud of this one
Let's tell a little story about how I came to semicap as an investor. At my old firm, we had a list of qualitative things we wanted to see in biz before we bought
secular growth
oligopolistic+ industry
pricing power
positive FCF and returns that to shareholders
It's been brought to my attention that a Morgan Stanley desk just emailed out my tweet as a reason for SMCI down today.
To which I have to say
░S░U░B░S░T░A░C░K░I░N░B░I░O░
You can search on Baidu but apparently there's been an ability to get H100s from SMCI in China, and that rumor in conjunction with no pre means SMCI gonna have a bad bad time
Maybe you’re sick of hearing it but NVDA has already thought about, and then executed on the problems that everyone is just starting to figure out. Jensen is so fucking ahead, the industry and people who are betting on mean reversion is chasing the after image.
Fun fact - according to most of the semicap players, I believe over 75% of the 1990 era is still with us today.
So yeah this take is wildly wrong. Most old generations of chips never go away, and are the reason why electronics are so cheap.
Oh and tool reuse is huge!
It's wild how bad capex spend is in semiconductors, this machine is basically tossed out as soon as the next gen is developed. Imagine doing this in manufacturing...new CNCs every few years lmao
You can search on Baidu but apparently there's been an ability to get H100s from SMCI in China, and that rumor in conjunction with no pre means SMCI gonna have a bad bad time
This is probably the most scaleable business in the history of the world. Semiconductors is one of the GOAT industries and it's because of it's business structure.
Nvidia grew revenue 18 billion YoY in absolute dollars, and it cost them an incremental 800 million in opex
Does anyone realize how screwed Samsung is? They haven't shipped over 100 layer NAND, they lost QCOM as a fab customer, and they lost their own socket to QCOM. I know we are all focused on Intel's demise, but Samsung is decaying at a much more alarming rate if you ask me
TSMC reported earnings last night and their results shocked most investors. Why? Buckle up and let's talk about one of my favorite companies in the world: TSMC.
I really think that the dell pivot at Nvidia is wild. SuperMicro was winning too hard, Jensen is like we cannot have that. No company is allowed to be but a spec on our event horizon. It's time to pump Dell servers instead.
Jensen does not fuck around lol
Jensen's keynote highlights something universally true in computing since the beginning: it's much cheaper to simulate reality than to observe it. AI is the final rendition. The whole keynote is just permutations of that truth.
Read more here:
I wrote about the Telecom bubble, examining the infrastructure build-up to today's AI infrastructure build-up.
I *think* it's not quite a bubble of that magnitude yet, and I think there's some important takeaways between then and now
Shout out to ARM to updating the graph
It cost 29% / 27% more to design at 3nm/2nm
It's decelerated from the inflection at 7nm, but it's still costs a staggering amount and is growing much faster than any other costs. EDAs rule
I think one of the biggest enablers of Intel's foundry ambitions is that TSMC is doing a ~54% gross margin during the cyclical low in the cycle
Implies full margin TSMC is going to be much higher during the high, and that margin is probably Intel's opportunity
First claimed successful replication of LK-99
Accomplished by a team at the Huazhong University of Science and Technology and posted 30 minutes ago.
Why this is evidence:
The LK-99 flake slightly levitates for both orientations of the magnetic field, meaning it is not simply a
The single sweetest thing happened today, mom read my newsletter about OpenAI and said it was super clearly written and the best thing she’s read on it.
She’s like you got a future in this! And I’m like yeah, I know LMAO.
Thanks mom. (She never understood the semiconductors)
so my read on the MU call was that almost the entire gross margin lift was probably due to HBM, and that should continue to really crank it for the rest of the year.
also 3x wafer intensity? SHEESH!! that is literally better than dreamed, i thought it was just 2x intensity.
I have been writing about how Automotive is in the worst spot in the cycle for the stocks, and $MBLY confirms how true that is today.
"we have become aware of excess inventory at our customers, which we believe to be 6-7 million units of EyeQ® SoCs"
Pretty nuts that ARM is now larger market cap than Intel, Qualcomm, Applied materials, etc. Only companies larger are AMD, ASML, Samsung, TSMC, Broadcom, Nvidia.
For context its prob the only company w/o a billion in revenue on that list.
1/ In the beginning there was Fairchild.
I think that is the best way to start. Makers of the Microchip by Christophe Lécuyer is the best book to begin with IMO.
Wrote about backside power delivery, and how it’s the next leg of transistor scaling past GAA. It’s also kind of the big bet vs TSMC that Intel is taking. Pay attention because this is what will matter in the coming years.
Wow that is not the AMD prints of old. Barely beat, barely guide up, and we mostly know that there is not much upside on the MI300 number. Get fucked I'll see you at the 200D
I don't think we have *ever* seen 52% QoQ growth from a revenue base of ~7 billion. That is completely unprecedented.
For context, Intel's revenue next quarter is going to be like ~11bil. They passed them in 1 quarter. Food for thought, they are guiding to a beat no?
btw I really think that Nvidia should be the most valuable company in the world - 100 billion in EBIT is not that insane in like CY25, and that would be approximately MSFT's EBIT but growing at a muchhhh faster rate
Also it's incredible to me that Nvidia is doing ~75s GMs
Im going to be honest with you, I really don't think that NVDA going to stop growing revenue anytime soon. I think the real test will be next quarter, 1Q with the guide into 2Q.
At that point in time there will be b100 to talk about, but they continue to accelerate results guys
Let me break it down for you:
deflation? yeah that's good for semis
inflation? also good for semis, they outpace it and are commodities
low growth? good for semis, gonna grow faster
high growth? yeah that's gonna be good for semis
All 4 quadrants are buy
I wrote my 2024 outlook at the last possible second but I wanted to discuss two big parts from my piece
First: I'm this is AI's adolescent years. Growing pains and growth spurts are to be expected, and I think election risk is real.
Second: MEMORY IS GOING TO MATTER A LOT
Wrote my first earnings post:
Think TSMC was good, SMCI means NVDA beats, TXN sucks, AWE is cheap and finally did a good job, LRCX crushed it and HBM is on fire, but there’s a lot of great disclosures.
More takes are coming very soon! Stay tuned!
Samsung has made its 400-person task force for HBM (high bandwidth memory) chips a permanent office led by Sang-joon Hwang, directly under the president of Samsung’s memory business, The Elec reports. The new HBM team will focus on stabilizing production yields of HBM3E and
finally reading the Broadcom day and I think that they are playing their hand well. Hey we will let you bring whatever you want, make sure it scales up w/ ethernet. (Android-ish)
meanwhile Nvidia is pursuing system integration and software lock in (Apple)
TWT who wins!
The man the myth the legend. Made the law so to speak, and literally changed the world.
Moore’s law has always been an aspiration of the industry, and the continued improvement has made all of our lives better. R I P.
This is probably the most interesting graph I've seen this earnings season. Memory test has been pretty much range bound for a long time, and it actually might be breaking out. 1B +/- from ~18 > 22. Now 1.3+ ?
This is the sea shift tbh.
bumping this. Managed to have this ready to go at the peak, and SOXX is down ~15% since then.
At some point I'm a massive buyer, but I'll make sure to tell you when. 😘
I wrote about the 3 headed hydra of competition that Nvidia has
And a basic overview of LLMs/ semiconductors that power them (NVDA!)
I really think that the 3 headed hydra is a powerful lense of competition for Nvidia
Everyone is fucking talking about the multiple and no one is forward looking, it feels like I’m taking crazy pills. When earnings like 2x next quarter yoy is everyone going to be surprised pikachu?
By the way - the first episode of Transistor radio w/
@dylan522p
and I came out yesterday. Listen in if you want to hear a conversation about why TSMC's earnings matter for the industry at large.
i get annoyed when people call my newsletter niche - dude semiconductors are not niche. It's in every single electronic device, the ones you stare at for 5 hours a day - yeah
Yeah new semianalysis post confirms what I thought, groq is only useful for a small subset of the market (small models inference) and won’t ever make a lot of sense for big models.
“Is Groq NVDA top?” Is an insane take, but shows you more about everyone’s desires than reality
I cannot believe this, like the actual worst cycle since 08/09, going to prob make a record for memory pricing decline since 90s and we are going to whip out on the other side to a fucking AI bubble
The entire group gross margin has expanded like 500bps and it’s floor prob
Yo this is an amazing graph!!!! I wrote about semiconductor pricing getting a lot better in like 2021, but this is historic. Stringing along multiple years of positive growth is unheard of in price
How much has pricing contributed to semiconductor industry growth in recent years?
Would you be surprised to learn the answer is “More than all of it?”
Our note today discusses this (for institutional clients only):
Semiconductor production has rebounded among major Asian exporters after taking a steep dive in 2022—Japanese production is at record highs, while Taiwanese and Korean production are closing in on prior peaks
It's almost exactly what I thought it would be.
Gaming revenue was $2.04 billion, down 44% sequentially and down 33% from the prior year. Data Center revenue was $3.81 billion, up 1% sequentially and up 61% from the prior year.
you're a recent AI startup with a new LLM model, you think you got a shot. The problem is you need a lot of GPUs, so you go raise a billion in capital and you throw it into GPUs, in the hopes of figuring out a business model to eventually buy more GPUs
this is triple-ordering
I wrote about TSMC. There really is something for everyone this earnings report at TSMC.
AI is only 6% of revenue but should grow at 50% CAGR? WFE spending is moderating? Oh I have a whole diatribe about why EUV intensity is past us and the 2024 EUV orders gap is worrying.
5/ Now I think it's time to get a bit more specific. In particular I think it's time to read Fabless by
@DanielNenni
/
@paulmclellan
Fabless IMO is one of the best books of the bunch, and the last chapter's predictions hold up shockingly well.
Love this quote from AVGO AI day, truly its bundle and rebundle all the way down. Nvidia is pursuing the mainframe, which I think some view as this dinosaur technology, but at the time was the 800 pound gorilla. Which thesis wins? T B D
Before I get started, a couple of things
ARM after hours price is just messy. Pretty much there was a huge short interest into this quarter, and I think its just a float.
DO NOT SHORT LOW FLOAT STOCKS. Think of it this way, its all supply and demand. Almost no supply, some demand. When that changes... do something
Feel weirdly qualified for a take on this.
The second method is how I run this newsletter, and while not perfect it is how to SEND research. I’m going to write a quick thread on da process:
There’s 2 ways people learning/work: (1) research, read first then write. (2) just start writing.
I’ve always been the latter. Found it easiest to just start doing and it’ll expose what I don’t know and help structure thoughts more.