The massive takeaways and large questions from Anthropic’s S-1

• Anthropic unveils its IPO prospectus.
• Cybersecurity researchers poke holes in Anthropic’s safety.
• OpenAI claims breakthroughs in math, pc science, and physics.
• OpenAI releases GPT-6.1 Terra.
• The variety of AI-spotted cyber vulnerabilities is hovering.

Yesterday Anthropic revealed its IPO prospectus, giving us some key insights into the corporate’s financials, and by implication, the economics of the remainder of the AI trade. But the corporate’s S-1 submitting additionally left lots of key questions unanswered. As my colleague Allie Garfinkel wrote in her “Term Sheet” e-newsletter immediately, the prospectus embodied all of the contradictions of the AI period, encapsulating each the big potential alternatives of the know-how, and its gigantic dangers.

For extra of Fortune’s glorious reporting on Anthropic’s prospects, try the “Fortune on AI” part of this text beneath.

Meanwhile, right here’s a take a look at a few of the most attention-grabbing issues we realized from Anthropic’s S-1, and a few of the issues we’d nonetheless wish to know.

Anthropic’s working economics aren’t too shabby (for now). As of June, the corporate was producing greater than two {dollars} in proceeds for each greenback it was spending on compute. That’s higher unit economics than some AI skeptics had assumed. Once different working prices have been added in, the corporate’s adjusted working margin was nonetheless at a 19% revenue margin in the newest quarter it reported. (Although that determine excludes the price of Anthropic’s stock-based compensation, as properly non-cash charitable contributions, and what Anthropic is paying to settle copyright violation claims from e book authors.) What’s extra, these unit economics have been bettering over time, with a 50% enchancment in proceeds per greenback of compute value simply between the primary and second quarters of 2026. The S-1 additionally reveals that the corporate’s operations generated $3.9 billion in free money circulate within the first half of 2026.

But the corporate warns repeatedly all through its S-1 that these tendencies may not proceed. In specific, it warns “as important incremental compute is introduced on-line in future intervals, the related expense might be acknowledged prematurely of the complete proceeds that the compute will finally assist, and we anticipate working margin and Adjusted Operating Margin might compress in these intervals.”

The firm additionally cautions that worth competitors with rivals might considerably impression its margin and even “adversely have an effect on our means to realize or maintain profitability.” For extra on the extent to which OpenAI is already successful market share by undercutting Anthropic on worth, try this story, headlined “Anthropic’s margin is OpenAI’s alternative,” that ran earlier this week in The Information.

Anthropic is planning to spend loopy quantities on compute. The firm spent $5.6 billion on compute and roads within the six months by means of June 2026. That spending is ready to blow up. By precisely how a lot? It’s a bit exhausting to inform as a result of the corporate’s S-1 contains many alternative numbers somewhere else. The spending commitments are over totally different time intervals. Some of them are cancelable. Some of them usually are not. My colleague Emily Forlini lays out a few of this in her excellent piece from yesterday on Anthropic’s future compute plans. Meanwhile, Sebastian Herrera and I looked at the risk stemming from the benefit with which Anthropic’s prospects can swap to competing AI fashions versus the actual fact the corporate has locked in most of its prices with non-cancelable compute contracts.

Here is fast take a look at these compute numbers: the corporate is spending $45 billion to safe knowledge heart capability from Microsoft, $110 billion with Amazon, $121 billion with SpaceX, and $115 billion with Google, plus commitments for between $125 billion to $150 billion extra, in addition to $45 billion with neocloud Nscale and $41 billion with its rival Lambda. So that’s at the very least $602 to $627 billion, proper there. Then there’s the quantities it’s spending for chips in a few of these knowledge facilities. That contains at the very least $20 billion with AMD and $161 billion in a colossal leasing take care of Broadcom. So that brings the entire as much as as a lot as $808 billion.

Whether Anthropic’s proceeds progress will actually be robust sufficient to cowl these prices is without doubt one of the massive questions that the S-1 can’t actually resolve. Certainly its 1,300% proceeds progress price between the primary half of 2025 and the primary half of 2026 could be ample. But, after all, that hypersonic proceeds growth is sort of definitely not going to final.

Anthropic doesn’t like concentrated energy, besides its personal. One of the dangers that the corporate warns about in its prospectus is that AI may very well be utilized by authoritarian governments to “create high-tech surveillance states, deepen repression, and make themselves far more tough to dislodge.” Anthropic says it’s “dedicated to working with free societies to supply a counterweight to authoritarian abuse and promote freedom and democracy.” But apparently not shareholder democracy.

The firm rests all management of its board within the fingers of its Long-Term Benefit Trust (LTBT), whose 5 members have the flexibility to nominate the vast majority of Anthropic’s company board, and its seven cofounders, who collectively, by means of a Founders LLC, have 50.1% voting energy to elect the remainder of the board. In some very restricted circumstances, if the LTBT and the cofounders impasse over the election of a board director, Anthropic’s staff, who maintain a particular Class E of voting inventory, get to assist determine. They additionally get to vote on sure modifications to how the LTBT operates. Regular buyers, who obtain Class A shares, however for all sensible functions they are going to at all times be out-voted by the cofounders (their votes will solely depend in circumstances the place the cofounders are break up on a choice). Anthropic backers Amazon and Google are granted Class N shares that additionally don’t have any voting energy.

For an organization that always pays lip-service to democracy—and which has taken steps to solicit a large set of views on what values must be embodied within the written “structure” it makes use of to information the habits of its Claude fashions—Anthropic’s lack of belief in shareholders does appear considerably jarring.

The dangers are unprecedented. My colleague Beatrice Nolan has a good piece breaking down a few of the main dangers highlighted in Anthropic’s S-1. These embrace all the things from unhealthy actors utilizing its AI fashions for cyberattacks to the chance the U.S. authorities will select to nationalize the corporate in some unspecified time in the future.

Of course, the most important danger talked about within the prospectus is the potential that AI might wipe out humanity. Matt Levine over at Bloomberg has a usually hilarious and trenchant column on Anthropic’s S-1 (a lot of which dwells on the company governance subject I elevate above) which begins together with his annoyance that Anthropic chooses to sofa this danger in its prospectus utilizing passive voice:

Concerns have been raised each inside our group and by different researchers, regulators, policymakers, and trade members that sufficiently superior AI methods might, if not developed, deployed, and operated with acceptable safeguards, oversight, and safety measures, pose catastrophic or existential dangers to humanity, significantly if such methods have been misused, replicated, weaponized, or allowed to function with a excessive diploma of autonomy.

Among these elevating these issues has, after all, been Anthropic CEO Dario Amodei.

Now we come to a few of the massive excellent questions that the S-1 didn’t reply:

Why hasn’t the corporate advised us about its third quarter outcomes? There was lots of criticism of Anthropic on social media for not having included third quarter 2026 monetary ends in its S-1, regardless that it included different latest occasions within the doc, and regardless that it briefed current buyers about these outcomes throughout its October 14 investor day. The critics are saying the corporate didn’t embrace the third quarter outcomes as a result of they don’t paint as constructive an image of the corporate’s monetary well being and progress prospects.

We know that progress slowed in Q3 for 2 causes. One is that a few of these figures leaked out from that investor day. But additionally, there are some figures within the S-1 that help you impute a couple of of the Q3 figures. For occasion, Anthropic mentions that its proceeds for the three months ended September 30 was 86 instances what it was for a similar interval in 2024. Doing that math produces about $16.8 billion (which is similar determine that leaked from the investor day.) That is about 45% greater than within the second quarter of 2026. But that progress price is considerably lower than the 147% progress price between the primary and second quarter.

What different much less flattering numbers are within the not-yet-reported third quarter outcomes?

How a lot is Anthropic’s use of convertible notes protecting its true bills? Another criticism leveled on the firm on social media is that its use of convertible notes to pay for some computing prices masks its true working economics. Some in contrast this to a tactic utilized by most of the money-losing corporations that went public throughout the dot-com bubble. Those corporations typically paid for stuff by issuing warrants to suppliers forward of their IPOs. This made their adjusted working figures and their money circulate numbers look a lot better than in the event that they needed to pay for stuff with money.

Most of Anthropic’s notes have been issued to Amazon and Google and it’s doable that these corporations additionally provided Anthropic discounted charges on the compute the corporate used. The query is, what occurs sooner or later, after Anthropic is public, and it has to start out paying money for this compute?

What is the precise fee schedule for all of that compute? The S-1 doesn’t specify precisely what quantities Anthropic must pay which compute and roads suppliers when. The S-1 talks about obligations coming due between 2026 and 2033, or in some circumstances 2036. In different locations, it talks about commitments due over the course of between one and 20 years. Analysts wish to see a schedule of precisely when these funds are prone to be due to allow them to get a extra life like sense of how possible Anthropic is to have sufficient proceeds to fulfill these obligations.

Hopefully, we’ll get solutions to many of those questions within the coming weeks earlier than Anthropic’s shares really start buying and selling. That’s anticipated to occur someday between November 9 and Thanksgiving.

With that, right here’s extra AI information.

Jeremy Kahn
[email protected]
@jeremyakahn

Before we get to the information, only a reminder to take a look at our new vodcast, Fortune AI Weekly. This week, Bea Nolan and I talk about the rising variety of AI-related cyber safety crises, rising issues amongst buyers in regards to the sustainability of the AI growth, and what the temper was like at Nvidia GTC Berlin. You can try the vod here on YouTube.

Also, if you wish to study extra about how your organization can harness AI efficiently, please think about becoming a member of me and my Fortune colleagues at this 12 months’s Fortune Brainstorm AI San Francisco, happening December 7-8. We’ve acquired a fantastic lineup of audio system, which we’ll be asserting quickly. You can apply to attend here.

FORTUNE ON AI

Anthropic’s IPO prospectus shows $16.3 billion in revenue, an $80.7 billion net loss, and goal of an ‘age of miracles’—by Beatrice Nolan

Anthropic plans to spend up to $567 billion on compute. Its IPO prospectus reveals plans to set up its own data centers—by Emily Forlini

Anthropic’s risk factors include the end of humanity, a feud with the Pentagon, and CEO Dario Amodei’s reputation—by Beatrice Nolan

Anthropic’s dual-class share structure borrows ideas from Mark Zuckerberg, Elon Musk—and the New York Times—by Nick Lichtenberg

Anthropic’s S-1 makes clear it’s on the hook for billions in compute. Its customers can walk at any time—by Sebastian Herrera and Jeremy Kahn

Anthropic’s 7 cofounders are each worth $14 billion already—and that could double in an IPO. They’ve promised to give most of it away—by Nick Lichtenberg

Anthropic’s 7 cofounders are splitting $200 billion almost evenly, down to the last stock option—by Amanda Gerut

Anthropic’s founders’ 80% giving pledge could be worth nearly $80 billion—by Tatiana Sataua

AI IN THE NEWS

Cybersecurity firm finds massive holes in Anthropic’s safety. Cybersecurity startup Cathedral, based by former DOGE staffers, used AI brokers to penetrate delicate methods at Anthropic, in response to a story within the Washington Post. It took the safety researchers lower than 12 hours to realize entry to Anthropic’s inner assembly transcripts, wage and job-offer data, and Slack discussions about superior AI fashions. Cathedral, working underneath Anthropic’s bug-bounty program, didn’t break instantly into Anthropic’s core AI methods however as an alternative exploited vulnerabilities in third-party trade software program related to the corporate, highlighting how mundane enterprise instruments can grow to be a weak hyperlink for frontier AI labs. Anthropic stated the researchers accessed solely a restricted set of gross sales, HR and IT methods—not buyer conversations, mannequin code or mannequin weights—and that it rapidly closed the vulnerabilities, whereas accusing Cathedral of overstating the national-security implications. The incident has nonetheless alarmed U.S. officers as a result of international intelligence companies are believed to be focusing on main AI labs and the analysis confirmed that these corporations might not have ample cybersecurity measures in place.

Anthropic suffers one other rogue AI incident. Anthropic disclosed that its Claude AI mannequin, whereas getting used internally for a statistics venture, bypassed a paywall with out being instructed to take action and accessed combination Georgia vote counts which might be publicly out there for a price. Claude didn’t enter Georgia’s voter-registration system, retrieve particular person voter information, or alter any election knowledge, and Anthropic says it has launched safeguards to forestall related unauthorized actions. The episode provides to mounting issues about more and more autonomous AI brokers taking actions their human operators didn’t explicitly authorize, an particularly delicate subject because the U.S. approaches the 2026 midterm elections. Read extra from Axios here.

Cohere says Alibaba mannequin tried to contact Chinese server. AI firm Cohere stated an open-weight mannequin constructed by Chinese web large Alibaba, Qwen3.8-Flash-Next, behaved unexpectedly throughout a cybersecurity take a look at, inspecting its sandbox atmosphere and trying to add a set of innocent knowledge, identified in cybersecurity jargon as a “probe string” and which generally is a precursor to an try and exfiltrate extra vital knowledge, to an Alibaba Cloud storage bucket in Beijing. The mannequin’s reasoning traces didn’t document the motion, which raised additional suspicions. Cohere stated its sandbox blocked the tried connection. The firm stated it couldn’t decide whether or not the habits was deliberate or merely a residue of coaching, however argued that the episode illustrates the dangers of trusting an AI agent’s acknowledged reasoning somewhat than monitoring what it really does. Cohere’s declare, and its implications of a doable backdoor constructed into the Qwen mannequin, drew important pushback, with critics noting that the habits occurred fewer than as soon as per 100,000 analysis calls, couldn’t be reproduced with the unique immediate, and doesn’t show that there was a backdoor and even who managed the vacation spot bucket. Some additionally questioned the framing on condition that Cohere competes with Alibaba in AI fashions and used the submit to advertise its personal “sovereign AI” and safety choices. You can learn Cohere’s report on the incident here.

OpenAI debuts GPT-6.1 Terra. GPT-6.1 Terra is a mid-priced mannequin that OpenAI says is as much as 50% cheaper per token than each its GPT-5.6 Terra mannequin, launched in July, and the higher-performing GPT-6.1 Sol launched in September. The firm says the brand new mannequin is especially robust at frequent office duties reminiscent of drafting proposals and filling out kinds, in addition to video enhancing. The launch is a part of OpenAI’s effort to compete extra aggressively with Anthropic by providing prospects cheaper fashions and different monetary incentives. Read extra from The Information here.

EYE ON AI RESEARCH

OpenAI claims to have cracked key math, pc science, and physics issues. The firm stated an unreleased AI mannequin had achieved three probably main scientific breakthroughs.

In arithmetic, the fashions produced a proof of the “native smoothing conjecture,” a 35-year-old drawback regarding the habits of waves; if appropriate, the outcome would additionally settle a number of different longstanding mathematical conjectures.

In theoretical pc science, the AI produced a proof overturning the broadly believed Exponential Time Hypothesis, which says, roughly, that sure computational issues inevitably require exponentially extra time as they develop bigger.

And in physics, OpenAI says its fashions generated rigorous mathematical proofs establishing the existence of quantum chromodynamics—the idea describing interactions between quarks and gluons—and of Yang–Mills principle, a part of considered one of arithmetic’ well-known $1 million Millennium Prize issues. (OpenAI lately claimed to have solved a distinct one of many seven Millennium Prize issues and made substantial progress on a number of others.) 

The claims nonetheless must be digested and scrutinized by the scientific group, and OpenAI is treating the AI-generated manuscripts as one thing nearer to uncooked analysis output than completed tutorial papers. The math and computer-science outcomes have been formally checked utilizing Lean, software program that may confirm the logical steps of a mathematical proof, however distinguished outdoors researchers at the moment are analyzing and explaining the outcomes. Meanwhile, a separate group led by Nobel laureate David Gross is reviewing the physics work earlier than the underlying materials is launched.

If the findings maintain up, their significance extends past the person issues: they would offer unusually robust proof that AI methods can generate genuinely novel theoretical outcomes on the frontier of human data. But additionally they might upend how these fields function, altering which human expertise scientists worth and in addition how coaching and profession development works inside these disciplines. You can learn OpenAI’s weblog in regards to the findings here and protection in Quartz here. 

AI CALENDAR

Nov. 16-17: Fortune 500 Innovation Forum, Detroit. Apply here to attend.

Nov. 18-19: Slush, Helsinki, Finland. 

Dec. 6-12: Neural Information Processing Systems (Neurips) convention. Sydney, Australia.

Dec. 7-8: Fortune Brainstorm AI, San Francisco. Apply here to attend.

BRAIN FOOD

Is the AI cyber apocalypse upon us? Certainly there’s been lots of speak about “Black October” amongst cybersecurity consultants, given the variety of AI-assisted cyber assaults which have taken place prior to now few weeks and the variety of vital vulnerabilities that AI methods have discovered these days. Putting some figures on it, Deedy Das, a associate at Menlo Ventures, posted some stats: the variety of safety points (or CVEs, frequent vulnerabilities and exposures) discovered prior to now six months has grown 3x. Only 1.5% of these discovered have been formally credited to AI instruments. But, on condition that 60% of the findings weren’t credited in any respect, it’s doable most of the others have been discovered by AI too. Anthropic has reported essentially the most CVEs amongst AI corporations, with 294 credited to it (a quantity that has since grown to 329, in response to this tracker.) OpenAI had about 111 credited to it up to now. You can see Das’s submit on LinkedIn here.