Simplifying the world of semiconductor investing in the age of AI. Part of the @semianalysis_ gang. all views are personal
- Tweets 12,171
- Following 867
- Followers 33,222
- Likes 25,109
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.
76
338
4,625
1,040,426
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 🧵
31
280
1,433
The goonwars have begun
The entire porn industry is about 100b which can fund a few GW
We made ChatGPT pretty restrictive to make sure we were being careful with mental health issues. We realize this made it less useful/enjoyable to many users who had no mental health problems, but given the seriousness of the issue we wanted to get this right.
Now that we have been able to mitigate the serious mental health issues and have new tools, we are going to be able to safely relax the restrictions in most cases.
In a few weeks, we plan to put out a new version of ChatGPT that allows people to have a personality that behaves more like what people liked about 4o (we hope it will be better!). If you want your ChatGPT to respond in a very human-like way, or use a ton of emoji, or act like a friend, ChatGPT should do it (but only if you want it, not because we are usage-maxxing).
In December, as we roll out age-gating more fully and as part of our “treat adult users like adults” principle, we will allow even more, like erotica for verified adults.
19
32
1,032
179,760
For those without a sub - the big news is that I am joining the mothership of @SemiAnalysis_
I want to be the best, and work with the best, and that means working w/ my long time friend and GOAT of analysts, @dylan522p
There's a LOT coming, happy to be here folks !
48
17
483
52,181
SCOOP: TSMC IS LOSING MONEY ON NEW 2NM CHIP RAMP!!!!!!
THERE ARE NEGATIVE GROSS MARGINS ON A NEW PRODUCT, THERE ARE NO CLEAR PLANS ON HOW TO IMPROVE MARGIN ACCORDING TO DOCUMENTS ACQUIRED BY ME (SEC FILINGS)
36
16
462
82,128
It’s so funny - US and China both lying about AI costs.
US: we need 7 trillion dollars (500b stargate)
China: we need 5.5m for deepseek (clearly used 10s of thousands of GPUs)
14
35
430
54,232
Holy shit remember America this is the competition. Jfc that’s an insane industrial base, just orders of magnitude above the west. They have full stop leap frogged us in EVs and most heavy industrials, it’s just a matter of time.
Wild: When the BYD Zhengzhou factory is fully complete it will be larger than San Francisco (but smaller than the Denver airport!)
~10x larger than the Gigafactory Nevada. Not apples to apples because it's effectively a city - employees live there in dorms.
34
52
421
78,263
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
13
21
401
70,786
Chamath said all semiconductors are short because of LK-99
Chamath clearly doesn't understand what a "semi" conductor is lmao
20
13
384
71,872
A bit of personal news (sadly). Taking some unwanted medical leave this week, unsure if I’ll post by EOW, but been struggling with some medical issues and they haven’t been getting better.
At this point I don’t have a choice but to focus on healing for a bit. HOPING I can get back to half days on Wednesday and onwards. We will see.
Health is wealth and I really cannot stress take care of yourself enough. Doing stuff a 30 year old guy should never do. Happy holidays.
89
411
45,522
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.
fabricatedknowledge.com/p/th…
I’m very proud of this one
38
42
346
68,252
omg 🤦♂️🤦♂️🤦♂️
Oracle lost nearly $100 million from rentals of Nvidia’s Blackwell chips, which arrived this year. That’s partly because there is a period between when Oracle gets its data centers ready for customers and when customers start using and paying for them, the documents show. It’s not clear what causes the gap or how Oracle plans to shorten it.
20
18
333
62,465
I love this shit. Last month we were talking about scaling laws and the models not improving enough, and therefor the ROI was going to be broken and less GPUs
This month deepseek MLA/COT RL/FP8 and now the models are too good, which means less GPUs needed
11
22
316
42,851
The simplistic framing is
Nvidia gets OAI equity so that OAI can buy more chips
AMD needs to give away their equity to OAI so that they can buy their GPUs
Regardless, congrats AMD today’s your day
17
12
318
28,640
fabricatedknowledge.com/p/th…
Wrote something about Intel - i really cannot believe how incompetent the board is. The Intel we knew is like actually dead w/ the firing of Pat. Let's hope the parts can survive
15
59
294
102,302
Jane Street made so much money in Indian options the regulator just barred them specifically from the market 💀
31
12
277
48,312
2020: learn to code
2025: learn to replace a coolant pump for the machine that learned to code
Non-ironically become an HVAC technician
10
21
268
21,486
Going to role play with what to do with Intel to make it survive the most economic way and possibly have the business(es) live into the future.
1) sell Altera for like 25-30b in cash ASAP. MRVL probably would be the buyer but they look cash strapped. MCHP/AVGO/NVDA? possible.
2) get STM / multiple car OEMS to get a very large stake in MBLY asap abv market continue the spin (think
MBLY CEO wants to stand alone) but accelerate the monetization.
3) sell fabless CPU to Broadcom. Hock tan is the man who can do the cuts and the company survives. In that process do a 10 year agreement to fab at IFS + include a WSPM agreement for Broadcom, including TPU/Switch/Custom, contingent on performance milestones. IE if nodes deliver they move volume over.
4) Pat gelsinger stays at Foundry. He’s the right tech ceo for the tech job, but isn’t the right guy to do cuts. Spin the debt to fabless w/ hock, who is good for the debt. Foundry does the one thing it must, make a god damn American process that is the best in the world bar none. IFS can buy the kingdom back if they do their job.
5) CHIPS act / Government is involved. A bit of incremental money, but importantly funnel every single wafer from defense + incentivize fabless to throw wafers at IFS.
Net that’s probably like what 50-100b dollars and as much support, and then IFS crushes 18A / 14A, and has a roadmap with the sexiest chips (Broadcom custom + defense) and the US has a standalone fab that prob won’t implode:
Downsides; everyone except for Hock might lose money lmao. Idk this isn’t a pursuit for the lighthearted.
30
24
256
92,873
Anyone who doesnt think the bots went away doesn't realize how insane and subversive this platform has become lmao
this works so often its scary
10
26
245
23,210
It’s all so tiring. We had the Blackwell delays echo in media like 4x longer than it needed to be, and I think we will have that in the MSFT DC pullback.
Guys it already happened + pretty broadly known. Somehow the PTC broker notes just got out now?
13
17
233
69,220
Why I think ASML's relative advantage has peaked - Peak litho and its time for investors to accept it.
Read more here: fabricatedknowledge.com/p/ea…
3
31
238
48,739
Imagine having a product so good the only alternative is to raise 7 trillion dollars
11
19
218
48,461
Guys they make a 57% FCF margin, has any company ever done that at this scale ever? lol
13
13
211
66,872
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
12
35
210
I really do not think people appreciate this
The s curve is already done, we crossed the chasm. ChatGPT is already as impactful as the internet and will last / be used for the rest of our lives. That is my takeaway.
15
13
210
25,076
*INTEL, AMAZON IN MULTIYEAR FRAMEWORK FOR PRODUCTS, WAFERS
*INTEL TO MAKE AI FABRIC CHIP FOR AMZN WEB SERVICES ON INTEL 18A
*INTEL TO PRODUCE CUSTOM XEON 6 CHIP ON INTEL 3 FOR AWS
*INTEL, AMAZON TO EXPLORE POTENTIAL FOR FURTHER DESIGNS
We are so back... from the brink
12
24
197
51,795
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
12
3
190
69,537
I really do not think people appreciate what this is: there has never been a source of truth for GPU throughout. Specs on paper have never meant anything.
This is IT !!!
All results and such can be accessed at inferencemax.ai
And the code and everything is open sourced here
github.com/InferenceMAX/Infe…
Methodology and explanation of results are here
newsletter.semianalysis.com/…
4
6
197
25,169
Bank of America went ahead and cut every semiconductor companies EPS this year
breath of relief folks
5
28
184
THIS IS VERY SIMPLE, NATIONAL CHAMPION.
Don’t overthink it, there isn’t a 4d chess move, this is checkers
$ASML is set to become the top shareholder of French AI startup Mistral AI
ASML is committing $1.5 billion out of Mistral’s ~$2 billion fundraise and is expected to get a board seat at Mistral - Reuters
9
4
188
39,806
I think it's so funny that every random quality LO in my replies tells me I'm wrong about peak litho. When you talk to formers, current ASML employees, and people who work on leading-edge processes, they are like, yeah, probably peak litho.
It's okay to keep blindly telling me it's a quality company and burnishing your five books that tell you that. Guys, it's technology, and I don't know what else to tell you. It's clear-cut.
21
8
191
34,375
There is no way for Oracle to self fund the capacity they are trying to bring online. Their recent earnings will likely be the most meaningful result of the cycle.
I think we are entering a new phase in AI, and I wrote about it here: fabricatedknowledge.com/p/or…
16
15
185
25,400
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.
19
10
181
44,148
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
14
18
175
151,466
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
6
22
177
22,509
One of the better articles / graphics I've seen about making a chip in awhile.
out.reddit.com/t3_1b5lva9?ur…
4
34
174
22,015
“AWS is pricing Trainium at 25% the price of Nvidia chips (hopper)”
Jensen: after Blackwell you can give away a hopper because Blackwell will be so performant.
You do the math on who wins in total cost of compute
8
10
171
27,080
Nvidia Earnings Thread: $NVDA
Let's see the most important AI company in action.
First: *NVIDIA SEES 2Q REV. $11.00B PLUS OR MINUS 2%, EST. $7.18B
1/n
10
16
176
96,018
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
6
15
166
22,517
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
15
16
155
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.
1
21
156
Pretty much you're a great American company if you have an antitrust case.
Standard Oil - 1911
IBM - 1969
AT&T - 1982
Microsoft - 1998
Google - 2020
Meta - 2021
Amazon - 2023
Apple - 2024
11
8
154
49,211
Lmao
DeepSeek will create a moment where all the AI startups that bought thousands of NVIDIA chips will probably go bankrupt, leading to a huge flow of NVIDIA GPUs into the second-hand market. This is in addition to startups whose model was to operate datacenters (e.g. Coreweave) and rent NVIDIA GPUs to other companies in hopes of generating ROI.
Lastly, the Magnificent 7 will finally start to slow down or cut any future orders from NVIDIA as there are already thousands, if not more, GPUs stored in warehouses and not fully utilized.
Then, everything about NVIDIA will start to unravel.
8
1
145
31,351
MSFT 10-K: We expect capital expenditures to increase in coming years to support growth in our cloud offerings and our investments in AI infrastructure and training
6
15
141
25,176
I just wrote about Nuclear power. Wanted to do a quick thread on Westinghouse, or the greatest American nuclear company that ever was.
1/n
3
12
139
31,465
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
7
5
139
22,890
Guys I know you know this, but this is now a cost optimization not a technological constraint. As sure as I am that the sun will rise in the East, I know we can move cost / energy down. Never bet against human ingenuity
Just absolutely insane to blow out the lights in 3 months
5
8
142
15,771
From 10Q:
"However, if we are unable to secure a significant external customer and meet important customer milestones for Intel 14A, we face the prospect that it will not be economical to develop and manufacture Intel 14A and successor leading-edge nodes on a go-forward basis. In such event, we may pause or discontinue our pursuit of Intel 14A and successor nodes and various of our manufacturing expansion projects. While we continue to evaluate Intel 14A for use in future Intel products and our plan includes an initial product designed to utilize Intel 14A, at present we are maintaining the option to design future Intel products requiring nodes with performance beyond Intel 18A and Intel 18A-P to be produced internally or by an external foundry. If we were to discontinue development of Intel 14A and successor nodes, we expect that a majority of our products would continue to be manufactured in our own facilities utilizing our nodes up to Intel 18A-P through at least 2030."
21
16
140
61,713
So Nvidia will prob sell more networking in 1 quarter than AMD will sell accelerators in a year.
4
8
136
23,971
I did a podcast on Alphasense, think you should listen. We talk about AI capex - and i seriously cooked on this
alpha-sense.com/resources/we…
2
9
138
18,224
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: fabricatedknowledge.com/p/je…
6
20
123
43,537
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
fabricatedknowledge.com/p/le…
9
16
122
93,023
wait, what is the point of SMCI if Nvidia is going to be selling the entire rack, including liquid-cooled designs. What is left to customize?
14
7
132
72,234
So OAI is tweeting about real interest rates, we get a new fed chair that’s whole job is to lower rates to as low as possible
OBBB means no taxes, money becomes really cheap, you gotta invest in America at all costs
Wow it’s so clear
8
3
135
14,734
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
9
19
125
28,777
38-42b Capex for TSMC
*semicap airhorns*
Ai going to double again
*NVDA/MRVL/AVGO airhorns*
4
7
129
12,181
Oh my god
I started building a position in $MU today in the low $130s (so I’m a little biased) but the stock is simply too cheap especially when you compare it to $AMD.
$AMD has a $250B market cap and is expected to do $5.5B of EBITDA over the next 12 months so it’s trading at 45x NTM EV/EBITDA.
$MU has a $150B market cap and is expected to do $20B of EBITDA over the next 12 months so it’s trading at 7.5x NTM EV/EBITDA.
FWIW, I also have a position in $AMD although I wish I didn’t. Thankfully my positions in $NVDA and $SMCI have more than made up for the $AMD drag. It’s probably too late to sell mu $AMD but they better do $8-10B next year in MI-chip sales or the stock will get pounded because it’s wildly overvalued in that scenario.
I know that $MU has always traded with a lower multiple because DRAM/SSD is a cyclical business with pricing troughs but HBM in the context of an AI-revolution means higher growth rates, higher margins, bigger profits, bigger buybacks … all of which should lead to multiple expansion.
I think $MU could be a $250 stock in the next 18 months.
21
1
119
54,204
Quantum computing stocks is the dumbest thing to happen since BNPL 3rd rail bullshit so I know it’s only getting worse
11
3
124
21,796
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
18
5
118
24,255
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.
8
5
122
19,396
There’s another way to think about it. If 1 product is a 2nm and the other is a 3nm yet one has better performance, they call the difference “margin”
6
3
123
36,625
@SemiAnalysis_ made a free public tool that talks about yield, you can explore what Pat means here:
semianalysis.com/die-yield-c…
speaking about yield as a % isn't appropriate. large die will have lower yield, smaller die - high yield percentage. Anyone using % yield as a metric for semiconductor health without defining die size, doesn't understand semiconductor yield. yields are represented as defect densities.
2
12
120
23,923
This is one helluva TAM estimate for transceivers, and if this is right then I think $MRVL is a long
13
12
120
23,816
I think that the killer app so far for LLMs being coding is definition of experts understanding how to use the tool to build vertical solution in practice
Just happens that the AI labs consume one type of knowledge work heavily, so of course have better understanding of the product.
As the heaviest power users understand how to embed LLMs into workflows and generate information work, more adoption will happen. From relatively newer knowledge work to older in no order
Data sci > digital marketing > software development > hardware development > mgmt consulting > advertising and PR > journalism and media > accounting/finance > law > medicine > scholarly research / academia
One vertical at a time, software is just the one right next door:
8
12
120
16,608
You really should read the details on this. Pretty confident that any model of consequence has never been trained on AMD
Our 5-month journey conducting independent analysis & benchmarking of AMD MI300X vs Nvidia H100 + H200
Detailed, open source low-level benchmarks
performance vs TCO
Comprehensive public recommendations
It’s not just immature software, they need to change how they do development
6
5
119
17,230
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)
14
117
13,646
$INTC is rightfully selling off. Every big cloud has their own CPU, what exactly is the role of an intel x86 CPU today? I don't know!
20
8
116
54,469
Literally think tegus does more harm than good now
Phase 3 cooling systems will change the entire design of the data center => vertical data centers become possible to optimize land use. This also enables data center skyscrapers in metropolitan areas. Microsoft data center architect via Tegus:
6
5
121
37,812
nitter.app/Andercot/status/168628…
another replication video
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 magnetized piece of iron or similar 'magnetic material'. A simple magnetic flake would be attracted to one polarity of the strong magnet, and repelled by the other. A diamagnet would be repelled under either orientation, since it resists and expels all fields regardless of the polarity.
Caveats
There is no way to verify the orientation of the strong magnet in this video, also, there are yet to be published experimental measured values of this sample. Diamagnetism is a property of superconductors but without measured and verified data, this is just suggestive of a result.
Take-away
If this synthesis was indeed successful, then this material is easy enough to be made by labs other than the original research team. I would watch carefully for results out of Argonne National Lab, who are reported to be working on their own synthesis of a sample.
This overall corroborates two independent simulation studies that investigated the original Korean authors claim about material and crystal structure, and both studies supported the claims.
Lawrence Berkeley National Lab: arxiv.org/pdf/2307.16892.pdf
Shenyang National Lab: arxiv.org/pdf/2307.16040.pdf
The attached video shows a small flake of their sample responding to an external magnetic field. I scroll through the video to skip to the relevant part.
original video credit to: @altryne
2
4
114
82,841
So many warning signs in cyclical especially.
NKE = Consumer Weakness (and share loss)
FDX = Industrial weakness / b2b weak
LEN = orders weak
MU = lmao
Occam's man
11
4
117
17,530
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"
7
10
117
20,501
Pretty insane moves in Semis, don’t think I’ve ever seen anything like it in terms of magnitude…. I mean it’s always been volatile but come on
20
3
111
20,521
I wrote a very important piece about BSPDN, and why Intel might have a chance.
Seriously it’s worth a read!
fabricatedknowledge.com/p/vl…
8
25
108
18,657
Press F for Intel. INTC will be +/- 100b for a decade as they figure it out - 18a/14a, but this is not an investment anymore, this is an open heart surgery and its public so sadly we watch it blow by blow.
Best of luck Pat, but even a good rider can't ride a dead horse
Wow it really is over at Intel.
THIRD QUARTER FORECAST Sees revenue $12.5 billion to $13.5 billion, estimate $14.38 billion
(Bloomberg Consensus) Sees adjusted loss per share 3.0c, estimate EPS 30c
Sees adjusted gross margin 38%, estimate 45.5%
7
5
115
25,172
