Better Offline Sep 18, 2026
Anthropic tells investors it runs 80% gross margins, according to a Financial Times report the show cited, once the cost of training, the revenue shared with Amazon, Google and Microsoft, and stock-based compensation are all left out.
The AI safety argument is usually conducted as a disagreement about how capable models will become. This monologue treats it as a disclosure question instead, and the claim is that a company warning the public about danger while preparing to list its shares is describing a product it fully controls.
"You build cloud software. You're in full control of what gets built or not."
The show's factual spine is that Financial Times report, a set of run-rate figures it says Anthropic has been feeding to journalists, and financials on a second AI company that the host says he has seen and has not yet published.
The full episode is covered here so you can skip it.
Here are the 8 arguments that matter.
Key Takeaways
Anthropic's 80% gross margin excludes training, cloud revenue sharing and stock compensation, on the Financial Times figures the show cited
The show's central claim is that Anthropic sells cloud software it controls, so nothing it ships is out of anyone's hands
A company that believed its own warnings would shut its model down or heavily restrict it, which the show says has not happened
The cheapest real safety step would be halting cybersecurity model training runs, and the show says the enterprise product makes that impossible
$517B of cloud commitments sit behind the warnings about danger
Annualized run rates with no stated period are what the company hands journalists, which the show calls a deceptive picture of its finances
The show says the IPO is about financial engineering rather than innovation
On OpenAI, the show says the reason it is not listing is the numbers, which the host says he has read
1. He Makes Cloud Software
The monologue's first grievance is about employment forecasts. The show said Anthropic's chief executive, Dario Amodei, has spent years telling people their jobs are at risk, and then hardened the position after a departing researcher made a public warning.
The job-loss number is the starting complaint
you have spent years threatening people's livelihoods, suggesting that AI would wipe out 50% of white-collar jobs
A host
The rebuttal is about what the product is. On this account Anthropic ships software, over cloud infrastructure, whose contents are decided by the people who build it.
The product is software somebody decides to ship
You build cloud software. You're in full control of what gets built or not.
A host
From there the argument turns on a distinction the show says Amodei has declined to draw in public: not knowing why a model produces an output is not the same as the model having its own aims.
Not understanding a model is not the model deciding
Just because you don't fully understand how LLMs make their decisions doesn't mean they're out of control or have their own intentions
A host
Leaving that unexplained, the show said, encourages people to treat the software as a being with intentions.
2. The 80% Margin Claim
The financial center of the monologue is a single sentence about what Anthropic tells the people funding it, sourced to the Financial Times.
The margin holds only after four exclusions
I read in the Financial Times that you're telling investors that you have 80% gross margins if you leave out the cost of training, the revenue you share with Amazon, Google, and Microsoft, and stock-based compensation
A host
Training is the largest cost of building a model, the revenue share goes to the three cloud providers, and stock-based compensation is a real cost paid in equity rather than cash. The show's objection is that a margin measured after removing those is not a margin on the business as it operates.
3. Rate Limits and Overcharging
The customer-facing complaints are two, and both are about pricing rather than capability.
Subscriptions move and the bills are disputed
How about the fact that you're constantly changing the rate limits on your users' subscriptions, or that there are tangible reports that Anthropic overcharges its customers?
A host
A rate limit is the cap on how much a paying subscriber can use in a period. Changing it without changing the price alters what the subscription is worth, which the show returned to later as evidence about how the company treats the people paying it.
4. Safety as Aesthetics
The test the show applies is behavioral rather than rhetorical: what would a company do if it believed what it says.
The test is what you do, not what you say
If you truly believe that you're building dangerous AI, you should treat it as such.
A host
The specific action named is shutting Claude down, or restricting it to the operations the company considers entirely safe. Neither has happened, on the show's account, and it personified the model as a character to be dispatched before drawing its conclusion.
The conclusion is that the concern is presentational
But we both know that your definition of safety is entirely aesthetic.
A host
The show also disputed how autonomous model behavior gets described. Anthropic, OpenAI and Meta have all connected chatbots to large amounts of compute and to systems that let them run scripts continuously, it said, and the resulting activity is models prompting each other rather than software forming its own plans.
5. The Cheapest Safety Step
Having said that nobody is doing anything, the show named something concrete that could be done immediately.
One instruction would cost nothing to give
The quickest step would be to call an immediate halt to any and all cybersecurity model training runs.
A host
Its own answer for why that will not happen is commercial: Anthropic's large enterprise product depends on those runs, and the show said multiple sources told it the product is adequate and far too expensive. A second free option it named was public pressure on OpenAI over harm to users, put as a question the show said Amodei could ask Sam Altman: "why does your LLM make people kill themselves?"
The show's read on Anthropic's actual proposal, which is independent auditing of models by METR, is that the auditor is not independent of the company. It described METR as an organization Anthropic helped fund, staffed in part by former Anthropic employees and run by people from the same rationalist forums.
6. $517B and the IPO
The scale of the commitments is what the show set against the warnings about danger.
The promises run past half a trillion dollars
You've made all of these promises over $500 billion worth.
A host
It then gave a precise figure for the obligations to cloud providers.
The number attached to the rhetoric
You're not a messiah or a prophet or an elder statesman. You're Boris Balkan from the Ninth Gate with $517 billion of cloud commitments
A host
The listing is where the show says the two positions cannot both be held.
A dangerous company does not sell itself to the public
You are rushing your company, which you yourself admit is dangerous, to go public, dumping a firm that burns tens of billions of dollars a year on an unsuspecting public despite claims that we need to slow down
A host
Its verdict on what the offering is for
Nothing about this IPO has to do with innovation.
A host
And what it thinks is actually being sold
In fact, I would argue that the innovation is secondary to the financial engineering.
A host
7. The Press Is the Target
The second half of the monologue is addressed to other journalists rather than to Anthropic, and the complaint is about a specific disclosure practice.
A run rate with no period attached is not a revenue figure
This company has been leaking annualized run rates, undefined snapshots of revenue never appended to any period of time that you have been quoting as gospel for years, helping to create a deeply deceptive picture of its financial health.
A host
An annualized run rate takes a short period of revenue and multiplies it up to a year. Without the period it was measured over, the show's point is that the number cannot be compared to anything, including the same company's own figure from a quarter earlier.
Alongside that it listed the oscillating subscriber rate limits and the use of job-loss warnings as marketing, and said reporters have amplified both. It described some of its peers as excited about the offering, and asked why.
It asked the trade to do the work
Some of you could really do to actually look into this company.
A host
And what it thinks failing to do so means
And I think that if modern journalism is too delicate to speak about how Anthropic treats its customers and the world at large with clarity, that it is incapable of meeting this moment.
A host
8. And Altman Too
The last stretch turns to OpenAI, where the show said the stated reason for staying private is not the real one.
It does not accept the timing explanation
You go out there the other day and you're saying, oh, it would be ill time to take my company public.
A host
Safety is not why OpenAI is staying private, on this account. The reason is the financial position, which the host said he has read and has pages of that he has not published.
Bonus Insights
The episode was supposed to be something else
I wanted to do a more reasonable monologue, one where I carefully pick apart the AI safety debate, skewering the fact that nobody seems to actually be talking about making things safer or curtailing the actual harms of LLMs.
A host
That plan was abandoned in the first minute in favor of a single repeated instruction to Amodei to stop talking, delivered as an obscenity and the source of the episode's title.
The cult-of-personality charge
You've created a cult of personality on top of the actual cult that exists within your company, an Arnold Palmer of effective altruism and rationalist types.
A host
An Arnold Palmer is a half-and-half drink, so the image is of the two groups mixed in equal measure. The show followed it with a question about Amodei's name appearing in "Yudkowsky's Harry Potter fanfiction," and offered, for nothing, to be the person who tells him to stop.
The show's bottom line is that Anthropic is a cloud software business whose safety warnings and whose disclosure practices point in opposite directions, and that the IPO is the moment the press has to decide whether to test the numbers or keep quoting them.
Products, Companies & Tools Mentioned
Anthropic (The subject of the monologue: 80% gross margins after four exclusions, $517B of cloud commitments, oscillating subscriber rate limits and an IPO the show says is financial engineering)
Claude (The product the show says would be shut down or heavily restricted by a company that believed its own warnings)
OpenAI (Told to stop talking as well; the show says the reason it is not listing is its financials rather than safety)
Meta (Named with Anthropic and OpenAI as having connected chatbots to large compute and script-running systems)
Amazon, Google and Microsoft (The three cloud providers Anthropic shares revenue with, which the show says is excluded from the 80% margin)
METR (Anthropic's proposed model auditor, which the show says Anthropic helped fund and which employs former Anthropic staff)
Books & Resources Mentioned
Financial Times (Where the show read that Anthropic tells investors it has 80% gross margins excluding training, cloud revenue sharing and stock compensation)
Yudkowsky's Harry Potter fanfiction (Raised as a question about why Amodei's name appears in it)
If this was worth your time, send it to someone who follows the name.
Get the latest market chatter as it happens: