A few days ago, Anthropic presented a form describing its business, financials and risks in view of its upcoming IPO, which promises to make history in terms of market valuation. The SEC (US stock market regulator) is in the process of reviewing this document, and in the meantime what the company’s CEO, Dario Amodei, said or said, is being looked at through a magnifying glass.
One of his interventions during the Dwarkesh podcast, which happened earlier in the year, has had some coverage on the networks these days. In the interview, Amodei presented examples of how the company could change its business model around tokens. The executive said that not all words produced by his models have the same value, suggesting a new dynamic pricing structure based on the complexity of the work.
Business model. According to figures cited in Anthropic’s S-1 filing, nearly all of AI’s revenue today comes from volume sales, with usage-based pricing accounting for about 80% of its revenue in the first half of 2026, with subscriptions accounting for the remaining 20%. The idea he introduced during the discussion of pricing based on the complexity of the work could change this statistic.
Details. Amodei cites a few examples to illustrate his reasoning: A model that asks a user to restart their Mac has said the same thing millions of times, so that the answer is worth a dollar or less. But if the model tells a pharmaceutical company how to rearrange a molecule it creates, those tokens could be worth millions of dollars, he says.
From there, he says he expects “payment by results” models to emerge. He admits he doesn’t know which business model will win, since it’s a new industry with a lot to test.
Similarly, Amodei also defends the current API model, calling it more sustainable than many think, as it allows developers to continue building new capabilities that future AI models offer.
Company statistics. Anthropic had $16.3 billion in revenue in the first half of 2026, according to the S-1 filing, and has more than $600 billion committed to infrastructure in the future, and could add an additional $150 billion to that number. The company’s total commitments for computing are about $895.4 billion, with $70 billion due in the next 12 months.
Revenue rose nearly 13-fold to $33.2 billion in nine months, yet the company lost $104 billion. According to the Associated Press, it primarily involved $94 billion in accounting charges in financing deals. Given such a commitment, it is unlikely that Anthropic would want to sell cheap tokens, so the dynamic pricing model might fit Anthropic’s account.
Amodei also explained in the interview that, according to him, the underlying economy is profitable and the losses mainly come from miscalculation of demand by purchasing computing capacity in advance.
What next? Anthropologie is expected to list on the Nasdaq later this month, but there are still questions about how the company plans to convert the power into revenue. A suspicion that is well extrapolated to the rest of AI laboratories. For now we’ll have to see how Anthropic plays out.
Cover image Dwarkesh Patel
In Xataka | Google is funding the future of AI in the very business it threatens to destroy: its classic search engine
