17 Comments
User's avatar
Kush Shah's avatar

Oh how I've missed this - a little wit, a little learning, a little humour, a whole lot of fun.

SatPost is back!

Trungphan2's avatar

Appreciate you, Kush. Too good of topic to miss!

Pascal's avatar

I’d call this article a perfect example of the “Trung Phan Paradox”: a brilliantly incisive, in-depth piece about AI and its business models that also reminds us how far LLMs still are from producing original, dense, clever, deep work like this, despite all their progress over the past 18 months. Thanks for sharing, as per the other comments I learned so much reading and rereading it. It’s one of the best SatPosts you’ve published, and that’s a high bar.

Trungphan2's avatar

Really appreciate the read and note, Pascal.

I will say that if default Claude output starts dropping jokes every third line and going on random tangents that later circle back to the main thread...gonna be a bit concerned for my writing style!

Adrian Unger's avatar

So what’s the endgame if anthropic wins? If all goes to plan wouldn’t the biggest player want to keep growing and eventually buy out competitors and any compute it doesn’t yet have access to?

At which point “intelligence” will be monopolized by a single entity? Does that feel like a good scenario for humanity?

Everyone’s distracted by if ai goes “rogue” as worst case but what if everything just goes to plan and the concentration of wealth and power becomes astronomical? 1.5T valuation needs to grow right?

Do we really want every Fortune 500, every student, every institution relying on a single service for “intelligence”?

Trungphan2's avatar

Thanks for the read, Adrian.

If any single one player wins the AI race (which I don't think will happen), it would probably get nationalized before the end game. Way too powerful.

The world that Satya Nadella describes -- where companies stay sovereign by protecting their data and being model agnostic -- seems like where we are headed. Which means value will accrue to a lot of different players in the value chain.

Its also in the interest of Nvidia to NOT have one dominant end user. Atm, they control the key input and are doing everything to ensure compute end user market is fragmented (eg. financing neoclouds, investing into open source).

Zac One's avatar

Interesting … the double dash - - spaces with perfect kerning - - … Saw that.

Umar AhmedTijani's avatar

I don’t have a concrete takeaway from this.

I’m just a bit more informed about Anthropic as a business and I’m looking forward to the IPO

Trungphan2's avatar

Thanks for the read, Umar!

Erhan Civelek's avatar

The Ctrl+F spot check is the step I'd keep. The memo now takes a minute, but the checking doesn't shrink with it: someone still has to tie each number back to the page it came from. Your $72B line is a good example. A summary that reports the $81B first-half loss without saying most of it is a non-cash charge on the Amazon convertible note describes a very different company. Faster summaries make the trip back to the source more important, not less.

Daniel Oliveira's avatar

On the $72B Amazon note charge, Amazon's own filings give a second opinion on the same instrument. Its Q2 10-Q values the $6.2 billion of Anthropic notes at $97.9 billion, while the S-1 carries them at $108.6 billion. The rise in that liability is the whole $72.4 billion charge, so a $10.7 billion gap between the two sides of one contract is a sizable number next to the loss it explains.

The other partners show up in their filings too, just unnamed. Alphabet's 10-Q describes "a $40.0 billion investment in a private company" with the same month, amounts and milestone structure as Google's Anthropic agreement.

I lined the S-1 up against the Amazon, Alphabet, Broadcom and Microsoft filings here: https://equibles.com/research/anthropics-s-1-checked-against-amazon-google-microsoft-and-broadcom-filings

Data: SEC filings, processed by Equibles.

Seixas's avatar

The point about people posting "15 takeaways" from a 400-page S-1 after feeding it to an LLM is fair, and it shows how much of the discussion rests on summaries. The revenue comparison with Goldman or Pfizer is striking, but it also sets a fast-growing, capital-intensive business next to mature ones. I am curious how you read the first operating profit alongside the large paper loss when judging the underlying unit economics. Which numbers in the S-1 do you find most useful for judging how durable this growth is?

Trungphan2's avatar

the irony of this comment is incredible

Umar AhmedTijani's avatar

What’s the irony? That it was probably generated by an LLM?

Zac One's avatar

But is it wise - - to always rely & trust an LLM to generate a summary? A quick down & dirty like TP uses here as part of a article arc yeah.

Is it more wise to create ones own summarisation skill with the Claude to get the nuances. Just bloody look at the daily driver out of the box summarising outputs of Gemini & specialist apps. Then run again against other factors of say ambiguity or contradiction or cross referenced - - yadda yadda.

Seixas's avatar

Good point, Zac. I find it useful to check a summary against the source, especially where assumptions or contradictions can disappear. What cross-check do you find most useful?