Irrational Analysis is heavily invested in the semiconductor industry.
Positions will change over time and are regularly updated.
Opinions are authors own and do not represent past, present, and/or future employers.
All content published on this newsletter is based on public information and independent research conducted since 2011.
This newsletter is not financial advice and readers should always do their own research before investing in any security.
Feel free to contact me via email at: irrational_analysis@proton.me
I like to think of the current state of the AI trade as this scene from Pirates of the Caribbean.
Sure there is a public negotiation on “slowing down AI” but these are pirates.
Why should any of us trust what pirates say?
If any of you think Dario and Sam, two sociopaths who despise each other, are going to coordinate effectively… lol.
https://www.reddit.com/r/OpenAI/comments/1r8spw5/sam_and_dario_didnt_hold_hands_at_new_delhi_ai/
Contents:
Delaying twice is not a good thing.
GPT 6 Astra is obliterating Fabel 5.1 in coding.
Coding is BY FAR the most important market.
Sam vs Dario // Sanjay vs Lee Chaebol
VC ARR mental gymnastics.
Profitable at 80% gross margins, before costs…?
Density, sparsity, and economics.
[1] Delaying twice is not a good thing.
Anthropic has delayed their IPO twice. Original plan was late September, then a delay to October, and now a second delay to “after the November elections”.
Everyone initially expected the S-1 to become public after labor day then a 2-3 week roadshow before late September IPO. Leopold spent $300M in call option premium right before labor day weekend presumably in part because of this anticipation.
God bless that heroic degenerate.
Anyway, WSJ yesterday reported of the second delay.
https://www.wsj.com/tech/ai/anthropic-shifts-planned-ipo-to-november-8874dffc
This is not a good thing. Ignore the safety bullshit theater. A basic awareness of what has been going on the last 3-5 months explains what is happening.
[2] GPT 6 Astra is obliterating Fabel 5.1 in coding.
There are only two times where I have witnessed an engineering community UNIVERSALLY agree on something. Normally there is at least a 5-20% sized faction who takes the opposite side.
The first example, which I discussed before a few times, is the Global Foundries “fusion”/monolithic node, 45CLO. That node is universally ridiculed and derided from both the electrical and photonic community.
Horrific thermal crosstalk issues.
Painful 6+ month cycle times because this abomination node has like 120+ layers.
Terrible power consumption, particularly leakage.
Serious reliability issues.
Bad/low ring structure Q-factor.
Bad ring modulator bandwidth and linearity.
Only recently did someone tell me Ranovos apparently likes GloFo 45 CLO. My response was Ranovos are incompetent losers so who gives a shit what they think.
Ayar Labs and Lightmatter both fled GloFo 45CLO to TSMC COUPE. Frankly, every product Ayar/Lightmatter made on 45CLO is garbage and needs to be written-off as R&D. They got to learn how to make DWDM ring modulator CPO/NPO I/O on a economically unviable process node and have since moved to a real process node.
There are people who keep telling me Lightmatter still has hope for the 45CLO interposer product and I am sort of at a loss as to how this level of delusion is possible. If you want to ship the interposer product in 2029, you better re-do everything on Tower hybrid bonded platform.
https://towersemi.com/2025/11/12/11122025/
Anyway the reason I bring up this seemingly unrelated side-tangent is that UNIVERSAL CONSENSUS WITHIN ENGINEERING is extrodinarily rare.
There is one other example and it’s recent.
GPT 6 Astra utterly obliterates Fabel 5.1 in coding.
I am not a programmer. Don’t use any of this stuff so this is more of a “vibes based survey” than first-hand experience.
Fabel 5 had let’s say 80% market share before GPT 5.6 Sol came out.
5.6 Sol took some share but difficult to gauge how much. Remember this is a vibes based survay.
GPT 6 Astra basically went to 100% market share WITHIN CODING almost instantly at launch. It’s kinda unbelievable how seemingly every programmer I know switched so fast and enthusiastically.
[3] Coding is BY FAR the most important market.
Altimeter had this slide at the All-In summit which I completely disagree with.
This is a classic braindead venture-capitalist style argument that “TAM so big its easy to hit crazy future projected numbers.”
AI is an EXCEPTIONAL tool for coding. Really a game-changer. But for other stuff it’s remarkably stupid. Still don’t use any of this crap other than occasional image gen.
Why assume 4% of all knowledge work (30T) is needed to justify the market?
Didn't you VC clowns say software is amazing? Too bad there are now two private companies that exist to obliterate the entire SaaS industry VC has spent the last 15 years investing in. I see a lot of clueless VC people trying to pivot into hardware investing and have literally no idea what they are doing.
Why not 3x coding market to 6T, give AI 20% share, and get to the same 1.2T ARR number?
My personal (vibes-based, no numbers to back it up) opinion is coding is such a massive market that it can justify the AI trade by itself.
Many of you have seen the news about OpenAI and Anthropic solving several previously unsolved math problems.
It’s not because AI is good at math and proofs. You are missing the point.
https://en.wikipedia.org/wiki/Lean_(proof_assistant)
There is this programming language called Lean. The code looks something like this.
Somehow, there exists a programing language that (apparently) defines all of mathematics as a rigidly-defined formal code. Ignore how this Lean thing works. That is not importnat.
Math and code are the same thing because of Lean. AI is good at code, not math.
I assure you none of these “AI mathematics breakthroughs” would be possible without Lean the programming language.
The thing is, code is how computers work. Think of all the productivity computers have brought about. We now have a thing (AI) that is really damn good at coding, in ways that were unimaginable before.
Productivity gains are very difficult to measure, especially when the tool enabling such gains is enabling previously impossible/unthinkable/unimaginable things.
Don’t care how McKinsey and Altimeter clowns came up with that $2T coding TAM. It’s wrong. Way too low.
[4] Sam vs Dario // Sanjay vs Lee Chaebol
In another forced analogy, I want to compare the dynamic between Sam/Dario and Sanjay (Micron CEO) and the Lee Chaebol who control Samsung.
![[Q1 CY24] Micron's HBM Party](https://substackcdn.com/image/fetch/$s_!9Jb_!,w_1300,h_650,c_fill,f_webp,q_auto:good,fl_progressive:steep,g_auto/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fef2114e3-e307-4dbc-91d8-508be7d9fdb2_808x427.png)
I feel like Dario beating the safety drum (again) so loudly a few weeks after Sam obliterated Anthropic’s financials by releasing GPT 6 Astra is a sign of desperation.
Sanjay begs Lee Chaebol for DRAM supply discipline. Dario begs Sam for “safety” discipline.
Zuck had an interesting comment and I shall make him the SK Hynix of this analogy.
As Micron was publicly begging Samsung for wafer start cuts, SK Hynix was sort of not involved and cut their own wafer starts because it was the rational/logical thing to do.
Zuck is sort of pointing out the obvious. I have heard rumors (3 uncorrelated sources) on Dario allegedly asking for liability waivers from USA gov and allegedly being told to fuck off. Bro if you care so much about safety why are you (allegedly) asking for extraordinarily liability waivers?
You can see David Sachs allude to this as well. He probably heard the same rumors I did.
[5] VC ARR mental gymnastics.
Last month, there was a whole kerfuffel in AI/semis/finance circles on Anthropic July ARR. Two numbers were going around. I don’t remember the numbers and frankly it does not matter. You will see.
One ARR number was the traditional “trailing 28 days * 13” number. Personally I hate this venture-capital clown metric but whatever a lot of people use this.
The traditional ARR number was bad and implied deceleration in growth. So the people massively long Anthropic came up with a new ARR number that was July 31st * 365 days.
So Antrhopic longs were already huffing weapons-grade copium in mid-August.
OpenAI rushed the launch of GPT 6 Astra to fuck over Antrhopic’s IPO.
FYI back in April/May people knew OpenAI would crush second half this year, months before GPT 6 Astra was launched. That was the SF rumor-mill.
Given that Astra launched on September 3rd (in a messy way) with real general availability say in mid September, what do you think has happen to Anthropic ARR?
Given the incredible share gains OpenAI has gotten in coding (vibes based 100%) how fucked is Anthropic ARR now?
No amount of goalpost-moving bullshit ARR metric re-definition is going to soften the blow.
Great job Sam. I feel like we are kindred spirits. Agents of chaos.
[6] Profitable at 80% gross margins, before costs…?
There are some ridiculous financial estimates coming from venture capital, Semianalysis, and Anthropic themselves. Remember IPO has been delayed twice so far.
Basically the costs of training future models (CapEx, R&D) is being excluded. So we have a business that is “profitable for two quarters in a row at 80-90% gross margins” that strips out massive costs using bullshit logic.
Let’s go for a quick analogy. Fabless semiconductor companies like Nvidia, AMD, and Qualcomm.
All three of these fabless giants need to ship new chips every 12-18 months. Miss a year of product cycle and the competition will cut your throat ear-to-ear.
Chip technology typically has a 2-10 year total R&D cycle. For example, DWDM CPO technology that Nvidia will ship in 2028 with Feynman had significant relevant R&D (researcher opex, many many test chip tapeout CapEx) likely as far back as 2018.
CPU microarchitecture is very complex. Whatever the Nuvia people were working on in 2025 (before they all left out of disgust and made another startup) will show up in Qualcomm’s 2028/2029 products.
So…. what if all these semiconductor companies decided to update their non-GAPP earnings definition by backing-out all future R&D and capex because “we have high gross margins on the chips we sell today”.
We can take this analogy further. Fabless semiconductor companies are light on CapEx. What if we compare Anthropic to Intel?
Both are capital intensive (datacenters, Fabs) and both have high gross-margin end products.
What if Lip-Bu Tan decided to update non-GAAP earnings to exclude the Fab shells under construction and the entire R&D pathfinding fab in Oregon because that is all for future products.
I have a very poor understanding of accounting and financials, but I can spot complete radioactive bullshit when I see it.
Which brings up an obvious question. Why is Anthropic trying to IPO now? Why not just raise a new private round?
IDK who this danny random is on Twitter but at least 7 financial professionals I know and met at least one in real life and talk to on a semi-regular basis (2 crossover, 2 long/short hedgie, 1 VC, 1 ex private equity, 1 pod) have given me the same answer.
Seems unbelievable to be honest. I thought the stupidity of venture capitalists is infinite. Hell, Situational Awareness LP could probably lead a round for Anthropic and the cultist LPs would happily buy at any valuation after a Leopold seminar at Lighthaven.
But the financial professionals seem to all agree that (for whatever reason) the private markets cannot support Anthropic anymore.
We may have a desperation IPO at the same time as Iran war (still not over…), elections, and rising interest rates.
This is why I am running significantly less leverage than normal. 64% margin equity vs normal 50-55%.
[7] Density, sparsity, and economics.
Few interesting tidbits I want to throw out there. This section is a summary of what smart people have told me. I don’t understand any of this model architecture stuff.
The key advantage of mixture of experts (aka expert sparsity) is that it drastically reduces training FLOPs requirements. However, MoE makes inference much more difficult because routing and such.
Rumors are that Antrhopic models are far less expert sparse than OpenAI models. The gap is so wide that a lot of people who follow this stuff seem genuinely shocked.
So in training compute intensity, Anthropic is at a meaningful disadvantage to OpenAI.
Correlated to this, Anthropic models seem to have a much higher active parameters ratio than OpenAI. This (really over-simplified) means Anthropic inference economics are also at a non-trivial disadvantage compared to OpenAI.
Guesses from a friend…
Astra is maybe 6T param total, 300-600B active.
Mythos 6-8T param total, 800B active.
Fabel (distilled Mythos because they cannot serve Mythos economically lol) 3T param total, 400B active.
Anyway we are going to have a fun time.
How did you achieve the “ less leverage than normal. 64% margin equity vs normal 50-55%.”🤣
"Basically the costs of training future models (CapEx, R&D) is being excluded. So we have a business that is “profitable for two quarters in a row at 80-90% gross margins” that strips out massive costs using bullshit logic."
-> Untrue. Profitable, and 80% gross margin excluding training costs, were two separate statements. They are profitable *including* training costs, but *excluding* equity compensation, which is pretty standard non-GAAP accounting for tech companies.