TL;DR

Anthropic raised Series H at $965B and filed a confidential S-1; autumn is still a press window, not a date. OpenAI filed as well, but Altman is holding for $1T and coverage has pointed to 2027. Bloomberg says Stripe agreed to buy OpenRouter for more than $7B. Dario's August 15 posts are the useful part for buyers: harder tests on frontier models, skip the cancer slogans, and admit the labs have not yet delivered what they keep describing.

I have been reading this stretch of August as a capital file. Anthropic is closer to public markets than any other frontier lab. OpenAI already has a confidential S-1 at the SEC, and Sam Altman has put a $1 trillion floor under a listing. Bloomberg reported on August 16 that Stripe finished a deal to buy OpenRouter for more than $7 billion. Two days earlier, Dario Amodei published a long reply on regulation and how he talks about AI. If you buy models, route them, or take this to a board, those four items sit on the same page.

Key Takeaways
  • Anthropic raised $65B in Series H at $965B on May 28 and filed a confidential S-1 on June 1. Last week Bloomberg had preliminary Q2 revenue above $11.5B and positive adjusted operating income. Autumn, possibly October, is still a press window.
  • OpenAI announced its confidential S-1 on June 8. Last private mark: $852B. Mid-August run rate: about $40B, with enterprise now larger than consumer. Altman has told advisers he will not list below $1T. Coverage since June has leaned 2027.
  • Stripe / OpenRouter: Bloomberg says more than $7B, about 5.4x the May Series B. Stripe has not confirmed.
  • Amodei's August 15 posts: tests that bite the frontier and spare smaller labs; public distrust is older than this cycle; the fair knock is that the large benefits are still unpaid.

Anthropic

On May 28, Anthropic said it raised $65 billion at a $965 billion post-money valuation. That figure includes $15 billion of hyperscaler money already committed, Amazon among the names. Company-stated annualized revenue had crossed $47 billion. On June 1 it submitted a confidential S-1. Goldman Sachs, JPMorgan, and Morgan Stanley are in the banker conversation, according to the usual reporting. Anthropic has not set a date. "Subject to market conditions" is not a calendar.

February's Series G was $380 billion. May more than doubled it. A few backers have talked about a public print above $2 trillion if 2026 run-rate heads toward $100-120 billion. That is hope. What you can underwrite is newer. Documents shown to prospective investors, reported by Bloomberg on August 14, put preliminary Q2 revenue above $11.5 billion, up from $4.73 billion in Q1 and $787 million a year earlier, with positive adjusted operating income. The figures are unaudited and can still move. They sit next to the $47 billion run-rate Anthropic stated in May. CFO Krishna Rao has started early investor meetings this week. The rest of the book is still the same: a decade-scale AWS compute commitment reported above $100 billion, and a $50 billion U.S. infrastructure program with Fluidstack. Anthropic is trying to arrive as a growth company that also wants traffic rules. Those two stories will have to live in the same S-1.

OpenAI

OpenAI announced its confidential S-1 on June 8 and said timing "may be a while," because some of the work is easier as a private company. The March financing marked the company at $852 billion post-money, with $122 billion of committed capital and about $2 billion a month in revenue then. Bloomberg on August 13 had the run rate above $40 billion; CNBC confirmed it the next day after CFO Sarah Friar told investors that enterprise now outruns consumer, two quarters ahead of her own earlier forecast. Altman has told advisers he does not want to list below $1 trillion. After SpaceX and a look at the cash-burn math, June reporting had the company leaning toward 2027. Nothing in the last two weeks has walked that back.

Microsoft remains the primary cloud partner, with a reported ~27% economic stake and a non-exclusive IP license through 2032. SoftBank's follow-on, including a $40 billion bridge due March 2027, puts a clock on liquidity even if product leadership would rather wait. The practical read for a buyer is simple enough: OpenAI is keeping the listing option alive without promising the year. That is a different file from Anthropic's autumn chatter.

Stripe and OpenRouter

On August 16, Bloomberg said Stripe had finalized an agreement to buy OpenRouter for more than $7 billion. Stripe does not comment on rumors. The Wall Street Journal had the talks in July. If the number is right, it is about 5.4x the $1.3 billion mark from OpenRouter's $113 million Series B in May (Sequoia, a16z, Menlo, Capital G). The product is a single gateway to 400-plus models, with centralized billing and an anti-lock-in pitch. Alex Atallah used to call it Stripe for AI. If Bloomberg is right, Stripe decided that sentence was cheaper to buy than to rebuild.

Routing is turning into a settlement problem. The firm that sits between Claude, GPT, Gemini, DeepSeek, and Qwen will meter tokens the way Stripe meters cards. A $7B check is a view that the switchboard is worth more than another fine-tuned chatbot. It also lets the frontier labs stay private a little longer while a payments company absorbs the metering layer.

What Dario actually argued

Amodei published two posts on August 15, answering Gavin. He does not usually live on social media, which is why the length is worth the time. The thread sits under Anthropic's IPO path more than it sits under a comms calendar.

On rules, he rejects the shorthand that regulation equals capture. He also rejects the opposite cartoon, that rules always protect ordinary people. What Anthropic has actually backed is a set of tests that bite the frontier and leave smaller labs alone. California's SB53, which the company supported, does not cover firms below $500 million in revenue or training cost. The testing design pushed at CAISI and the White House is heavier for frontier models than for everyone else. "Pacing the Frontier," in Anthropic's preferred reading, would slow the best models and leave catch-up shops, including open-weights labs, less boxed in. He is blunt that AI still concentrates power because of scaling laws, and that open-weights mostly move that concentration to whoever can buy the chips. He says he supports the administration's reported path: pre-deployment tests for frontier models, and tests for open-weights once they get close. He also likes Demis Hassabis's idea of a FINRA-style body.

On tone, he disputes the charge that he is disproportionately negative. He points to Machines of Loving Grace on the upside and to Policy on the AI Exponential on process. Short clips of risk travel. The benefit essays do not. He thinks public hostility is older than this industry: people do not trust companies, governments, or tech, and they assume the next product is a way to get played. A glossy campaign will not reverse that. The line that will get quoted is his: saying AI will cure cancer is a cliche. Curing it would not be.

The criticism he accepts is the one I would keep on a diligence list. The labs, Anthropic included, have not yet delivered the large real-world benefits they describe. He says that is on them. Biology and medicine are being staffed up. He hopes for early results in months and larger ones in years. Until then he will not make empty promises, and he will keep talking about cyber, bio, and alignment risk. His claim is that honesty costs less trust than pretending those risks are imaginary.

My View on This

I would treat Anthropic as the first public comparable, with the usual private-market haircut. OpenAI is defending a number that 2026 public investors may not grant. If the OpenRouter price is real, routing and billing are consolidating faster than model brands. Amodei's posts are the language a listed Anthropic would have to repeat: more tests, fewer slogans, and an open admission that the biology proof is still in front of them. For a 2026 procurement plan that means two model families, a routing layer you can switch, and a safety policy you can explain to a board without a press team in the room.

Sources