anthropic aiming to IPO for $2t

TLDR:

• Business: I hosted a meetup for people doing cool shit

• AI: NVIDIA posts the single greatest quarter ever by a company

• Builder’s notes: the different phases of AI + harnesses are the differentiators

• Differentiated take: the zone of uniqueness

💼 What’s new in the business

This week I held a catch-up with young people that are involved with AI, run companies and just generally doing cool stuff.

For the event every person had to bring a thesis on the future: what he thought would be true over the next 12 months and beyond.

I was thinking of doing a formalised q&a of sorts, but I thought it unnecessary. 

I find when you put a group of really intelligent and ambitious people together in the same room naturally the conversation lends itself to topics akin to the above.

See one of the boys getting some work in late at night (he runs a construction company and is killing it at the moment)

He prioritised getting from Dubbo to Sydney after wrapping up work at 5pm to squeeze in seeing the boys, and jamming in uni work once there.

I always find it great to meet with people doing interesting things.

(Also this image is hilarious)

🤖 What's new in AI

The Anthropic $30 trillion IPO.

Anthropic is telling IPO investors its market is worth more than $30 trillion. US GDP is $32.5 trillion.

The IPO is targeting around $2 trillion. The $30T is their total addressable market.

The IPO is not news. But the TAM of $30t is fkn insane.

For reference the entire US GDP is $32t.

So how the fuck did they get there?

Anthropic forecasts $190 to $200 billion of revenue in 2028. Against $30 trillion that's 0.65%, which reads modest. But that still doesn’t get them there.

They're extrapolating AI usage for replacing labour.

I.e. labour is roughly 52% of global income. Against $126 trillion of world output that's about $66 trillion in wages every year. So Anthropic is claiming something like 45 cents of every dollar the world pays a human.

Very ambitious to say the least.

The interesting thing though is that I think a lot of this value is being dispersed. Specifically that the open-source models are reducing the price of tokens, which is how Anthropic makes most of its money.

The flaw is that wages and tokens are pulling apart. OpenAI cut GPT-5.6 Sol's price over 20% this month. Alibaba and DeepSeek keep undercutting from below.

So the models capture more work while earning less per unit of it. A labour-sized market with a falling price per unit of labour isn't the same as a $30 trillion business.

I will say their current annual revenue run rate is about $70b, with them expecting that to reach $200b by 2028. But I find it hard to believe they will capture literally half the value of the entire world's labour force.

Furthermore, the rate of adoption, I believe is being severely overstated.

Altman himself admits as much.

Twitter tweet

To be completely honest I think anyone with half a brain could have seen this. People generally are unwilling to adopt until it gets to a point where they are literally forced to.

Anthropic's $30 trillion market size estimate is outlandish. That may be the point

SAAS is not dead.

Salesforce stock jumped over 20% on the 27th, its biggest day since 2020.

The reporting mostly credited Agentforce, and Agentforce is genuinely growing: $1.5 billion annualised, up 240%.

(Agentforce is their AI platform for running the ops of a business)

But the earnings beat was $5.90 a share against $3.27 expected, and a big piece of that gap was a $2.6 billion mark-to-market gain on Salesforce's stake in Anthropic.

Agentforce at $1.5B sits inside $46B of total revenue.

I was recently listening to the All-In podcast. Chamath had an interesting take (I wrote about something similar a couple of weeks ago)

Despite the premise that anyone can build software, the opportunity cost for these large enterprises to move resources away from their core competency and reallocate to building their own software is so great.

As such the monolithic incumbents (Salesforce, Slack, etc) are very well positioned, because they do their core competency really really well. And as I've said before, AI is only as good as the data that you give it. If you host the data, you're an essential part of doing business.

It's the middle-class software that largely gets wiped out. I.e. niche vertical softwares that position their value as having workflows specific to a business, that could easily be replicated by Claude Code in 12 months time (if not already).

Salesforce's moat is owning the system of record. The context an agent needs to do the job is already sitting in their database.

The thing nobody else has, that they do, is the data about your customers.

Nvidia's Historic Quarter, SaaS Comeback, Bessent vs Druck

Neutral infrastructure is over.

OpenAI is cutting off Cursor's access to its models on 12 November.

The reason it gave: it can't be confident SpaceX will honour its terms of service, "based on our experience with Elon Musk's companies violating contracts." SpaceX closed its $60 billion acquisition of Cursor on 14 August.

Doesn't really matter too much because ChatGPT usage on Cursor was only about 5%.

But a little middle finger to a rival nonetheless.

This however is a little precursor to where I think the world is going, particularly for the big labs (more on that below)

OpenAI to cut off AI models for SpaceX-owned Cursor

The mystery model had a name after all.

Last edition I wrote about Ox Alpha, the anonymous frontier-class model that showed up free on OpenRouter with a million tokens of context and no lab attached.

On the 26th the model was claimed. Ox Alpha was Z.ai's GLM-5.3-Flash, and they shipped the weights on Hugging Face under an MIT licence. Fun fact: the codename came from a Chinese film, Niu Lai, "Ox Comes."

320 billion parameters, 18 billion active, natively multimodal, a million tokens of context.

It had already processed 62 trillion tokens while nobody knew whose it was.

To the non-technical person not all that much.

But I provide it as pretty cool evidence that open-source is coming for the frontier.

NVIDIA has the single greatest quarter of a company ever.

$70b in net income in a single quarter.

$96.2 billion in a quarter, up 106%. Datacentre alone was $89 billion.

This is the single most net income recorded by a public company in a single quarter ever.

Twitter tweet

Demand also broadened past the big labs into enterprises, sovereign buyers and neoclouds. I spoke about something similar a few weeks ago following Alex Karp’s rant.

Nvidia is increasingly vendor, investor and financier to the same customers, including a reported $105 billion guarantee on one OpenAI datacentre. Commitments more than doubled to $279 billion. CNBC's own write-up was headlined "blowout earnings contained some red flags."

Having said that in the same week, Nvidia is reported to be buying Hugging Face for $12.9 billion.

I think OpenAI poked the bear with this one. They released their own inference chip, Jalapeno, squarely aimed at removing dependency on Nvidia (see previous newsletter for this)

Nvidia has gone fuck it and is trying to own the whole stack themselves, open source style.

I think we'll see a lot of the labs move in this direction.

1/ they have too much money not to.

2/ it's a scary proposition to be reliant on a competitor for power and compute.

A small note on personal AI companions

Instinct raised $250 million at a $2.5 billion valuation, still in private beta, no revenue disclosed.

It's a personal agent you text or call. No new app. Founder Noah Shinn is 23.

Instinct

The speed of the increased valuations is insane.

Roughly $50-100M in April, $500M in early August, $2.5 billion on the 26th. Fivefold in about three weeks.

Btw this is not revenue - these are just the valuations.

You have to apply to be part of the small number of users.

But honestly stuff like this just makes me think like wtaf.

Who is genuinely using tools like these? Like are people really going to pay money to be able to order from uber eats 10 seconds quicker?

I am bearish on this.

And still, the public isn't sold.

Two years of the products getting obviously better, and Western sentiment has gone the other way.

Pew asked in June. 52% of Americans are now more concerned than excited about AI in daily life, up from 37% in 2021. Only 9% are more excited than concerned.

The under-30s moved most, from 31% concerned to 55%.

Now the contrast. Chinese optimism sits around 83% on whether AI does more good than harm. The US number is 39%.

Possibly as a result of things like the AI Olympics in China, which is absolute insanity.

Twitter tweet

Excitement: China 84%, US 38%. Trust: China 72%, US 32%.

Sacks made this his thesis on the most recent All-In episode. His line was that the US leads China in every category except one, which is optimism, and that it's the biggest risk to winning.

I don't fully accept that premise.

High trust in a country with no meaningful way to express distrust isn't quite the same measurement.

But the trend definitely feels real, and it's not for a lack of capability.

Young adults in the US are increasingly wary of AI

👷 Builder's notes

Chamath laid out a decent map of where we are on All-In last week. I wrote about something similar a couple of weeks ago.

Phase one was models. A brain.

Phase two is harnesses. In his words, giving that brain "a pair of eyes and hands and a notebook for memory and a keyboard to type things on." That's agents, and he reckons we're near the end of it.

Phase three is context. Same agent, but you "train it to be a lawyer, or a customer service rep, or a sales agent." Which needs a ton of information about your specific business.

He didn't use the word employee. But that's basically what he's describing.

The harness is the differentiator

Back to Chamath's phase two. The harness is the scaffolding around the model: how it manages context, when it retries, whether it can spawn subagents, how it checks its own work.

Interestingly, when it comes to actual testing, the harness turns out to be a vital component in how well a model performs.

Harness-Bench ran 106 tasks across eight model backends and six different harnesses. The best harness scored 76.2%. The worst scored 52.4%.

Same tasks, same models. 23.8 points of difference, purely from the scaffolding.

They also found something that cuts against the easy version of this argument. Stronger models are less sensitive to the harness. The weaker your model, the more the build around it matters.

On cost it wasn't close. Harness choice moved tokens-per-solved-task by 40x. Upgrading the model moved it between 1.0 and 1.3x.

By simply changing the harness you could see a 40x uplift in efficiency.

I think we’re going to see a real push towards improving harnesses particularly as we may run into LLM capability constraints.

I’ll have more on how you can take advantage of this in the newsletter next week.

Munder Difflin. A free open-source desktop app that runs a fleet of coding agents in parallel and draws them as little avatars at desks in an office. Claude Code, Codex, Gemini, Cursor and about eight others, all at once. Super sick interface.

You can have a look at orchestrating different harnesses in one space with this plugin.

Kind of like VSCode but way cooler.

Twitter tweet

/design in Claude Code. Type /design and it generates editable UI artboards, you pick one, then it implements it in your codebase.

Shipped the week of 17 August. Pro, Max, Team and Enterprise, needs v2.1.233 or newer. Still a research preview.

For those using Claude Design, this could be super helpful.

Claude memory, everywhere. As of the 25th it carries context across chats and Cowork sessions. Free plan included, on by default for individuals.

It was genuinely strange that memory stopped at the edge of a conversation, and now it doesn't.

Note, admins control it on Team and Enterprise.

Polar. An AI-native browser aimed at work rather than search. You hand agents tasks off your open tabs and schedule recurring ones, and it drives a real logged-in session.

macOS only, free tier then $20/month. Made by an ex-Perplexity engineer who worked on Comet.

The argument for browser agents is that most of the web has no API, so driving a real browser is the only way in.

The guys at Lyra invested in them, 3 young guys out of America.

It’s super cool, actually spins up a browser (which you can see yourself) inside of the app.

Best use case I’ve found?

When I’m scraping for anything on the socials, wayyyy better than the claude or chat extension in your browser.

Hey Clicky. A Mac assistant that watches your screen and answers by voice, or spins up a background agent. YC-backed, $10.1M raised. Free tier, $20/month for Pro. macOS only, Windows waitlisted.

Beautiful UI (look how pretty)

The idea is that it it’s able to interpret your actions without you having to actually use a cursor at all.

Slightly dystopian, really really cool if done right.

Canva Magic Studio. One prompt box, describe the design, get something on-brand in seconds. Free tier gives you 10 Magic Design generations a month, unlimited-ish needs Pro at about $15.

This is the og idea that Canva has always been moving to. Super interesting watching their pivot into this at such a large scale, while also trying to manage pricing changes.

🧠 Differentiated take: The zone of uniqueness.

As I've grown older (yes I know I'm only 25) I've met more and more people.

In these experiences I have grown to believe that there exists a rare breed of people who fit the in the below tri-venn diagram.

I call it the zone of uniqueness.

The avatar of this person is someone who, as the diagram suggests, is intellectually capable and ambitious, morally a good person, and is also a good communicator with a higher order vision for life.

This is not to say that one has to conform to all three of these traits in order to be "successful".

Often it is the case that a tech billionaire, for example, does not exhibit all three. They may be fiercely intellectual and ambitious, a good communicator, but generally could be considered not to be a social person nor morally all that good.

Alternatively, there are many people who are sociable and charismatic, great people, but lack fierce ambition.

Then there are also people who exhibit one or none of these qualities.

With all of that said, if you are fortunate enough to find the rare breed of person that exists in this zone of uniqueness,

I believe it is vital to do what you can to increase your exposure to such people.

Not in a transactional sense of "I want something from said person", but simply because I believe you are generally the average of the five people you surround yourself with.

So if you can exist in circles that allow for increased exposure to these people, it should allow for a better quality of life.

Not too much further deliberation for today, but it's something I've increasingly found myself thinking about as I meet more people.

If you want to itch your existential angst further, please find any number of my writings here.

Cheers,

🎶 Carve outs:

• Someone asked Claude Fable to make its own personal website:

Twitter tweet

• Book list I’ve been collating from the recommendations of people I trust. (anyone else has recommendations - lmk pls!)

Until next week,

Alex