Nvidia sells chips to Anthropic and wants into its IPO

On Friday, September 11, Reuters reported that Nvidia was in talks to invest up to $10 billion in Anthropic's initial public offering. Not as one buyer among many: as an anchor investor, the one who commits before the stock opens and gives the rest of the book something to lean on.

The report rests on anonymous sources. Bloomberg, The Next Web and The Decoder all picked it up citing the same wire: a raise of up to $100 billion. Neither Nvidia nor Anthropic has commented.

The next day, Dario Amodei published a 3,800-word essay on his personal site, "We Must Pace the Frontier." In it he asks the industry to slow down how fast model capabilities improve. Sam Altman and Elon Musk both endorsed it the same day.

Both facts are dated, and they don't cancel each other out: a company can sincerely want its sector to slow down and sincerely want to raise a hundred billion dollars. What can be measured is what that slowdown leaves untouched.

What the essay asks for, and what it doesn't

We read the whole thing. It lays out a three-step plan, and only one is something Anthropic commits to on its own: hosting an external review team on site, permanently. The other two require coordination between governments.

Amodei draws the boundary himself. "Pacing does not mean halting model training or technical progress," he writes. Coordinated pacing, he adds, would give frontier developers time to do the safety work "without sacrificing commercial advantage or the United States' lead in AI." It's a speed limit everyone agrees to respect, as long as they stay out front.

One passage is circulating in mangled form. So that labs can talk to each other about standards, Amodei writes that the US government doesn't need to take part, but does "need to issue a narrow waiver for certain kinds of safety conversations." That's a recommendation in a signed essay, not a formal petition, and it covers safety discussions, not an understanding about production pace.

What the text leaves out counts too. Across its 3,800 words, there is no mention of an IPO, a share, a shareholder, an investor or a market.

The supplier and the shareholder

Nvidia holds two positions at Anthropic at once, and Anthropic announced both itself, on the same page of its November 18, 2025 press release.

The company committed there to buying "$30 billion of Azure compute capacity," with up to a gigawatt of it running on Nvidia systems. And in the same text: "NVIDIA and Microsoft are committing to invest up to $10 billion and up to $5 billion respectively in Anthropic." The equipment maker funds the club it sells the kit to, and the club is the one writing it down.

One thing cuts the other way, and it sits in the essay. Amodei calls for a ban on selling powerful AI chips to China. On that point, the lab is arguing against the order book of the company that might anchor its listing.

The loop, as a central bank describes it

The case isn't isolated. The Bank for International Settlements, the central banks' bank, named the arrangement in its annual report of June 28: "circular financing: chip makers and hyperscalers take equity stakes in AI labs or neocloud providers, who in turn commit to multi-year purchases of chips or computing power." It adds that the terms of such deals are "typically poorly disclosed."

The Nvidia-OpenAI letter of intent from September 2025 wrote the loop into its own sentence: up to $100 billion invested in OpenAI, released progressively as the new Nvidia systems get deployed. The supplier puts up the money, the customer spends it on the supplier's chips, and the line between an investment and a purchase order gets hard to draw. The $12.93 billion Hugging Face acquisition ten days ago runs on different mechanics, but it gives the scale.

Three facts push back on that reading. Huang answered the charge three days ago at the Goldman Sachs conference: "it's not circular because we put a little bit of money in, and a lot of money comes back." He went further: "we put in $1 and $100 comes back in. Is that circular? If that is, let's do more of that." Announced amounts also run ahead of committed ones: we checked in July that OpenAI's $100 billion turned out to be $30 billion. And Anthropic buys elsewhere: AWS remains its primary cloud provider, on top of five gigawatts of Google-Broadcom capacity.

In March, the listing was the reason not to invest

On March 4, 2026, at a Morgan Stanley conference, Huang said that Nvidia's $30 billion investment "might be the last time" it puts money into OpenAI. His reason, as reported by CNBC: "The reason for that is because they're going to go public." He added that the $10 billion committed at Anthropic would probably be the last there too.

Six months on, the figure being discussed is the same and the door swings the other way.

The two moves aren't the same kind of thing: an anchor order sits in the primary market, where March's was a private round. But the argument made back then was that a listing closes the window. It's turning into the window.

What no document says yet

None of the figures going around this week appears in a public filing. The only official number is still the one Anthropic published itself in the spring: a $965 billion valuation after its Series H. The $2 trillion being talked about is a number that went around every meeting room without going past an auditor.

We went and checked the SEC's EDGAR registry: no S-1 filed by Anthropic as of September 13. The same search with no filter returns thirty-five companies, so the engine answers fine; it's the prospectus that's missing. That's consistent with the confidential filing announced on June 1.

Three markers will tell the rest: a $15 billion revolving credit facility still to be finalized, the analyst meetings, then the public filing of the prospectus, which the SEC requires at least fifteen days before any roadshow. That's the day someone outside finally sees audited accounts. The elections fall on November 3, 2026, and Reuters' sources put the listing a few days before that. Nobody involved has said the election factors in, and a calendar moves for twenty ordinary reasons.

Once it's listed, who decides?

One question the plan doesn't settle. The external reviewers Anthropic promises to host will have desks, badges, and the right to publish key findings "without editorial control by Anthropic." The lab keeps a narrow redaction right, covering among other things "commercially sensitive" information.

In a listed company, an outsider publishing freely about model risk produces something other than a report: financial information, the kind that can move a share price. That same weekend, Sam Altman named the tension for his own account. Speaking to Fortune about why OpenAI keeps its hybrid structure, he said the company needs to be able to make decisions "that are not obviously in the interest of our business and our shareholders."

Neither the essay nor Anthropic's press releases say who decides between the transparency promised and what a listed company owes its market. Nothing is pressing yet, as it happens, since no reviewer has signed: Amodei writes that Anthropic "intends to invite" a team, in the future tense.