Anthropic, one of the largest AI startups, just filed an S-1 for an IPO, and it has more red flags than a PLA parade.
When Anthropic’s S-1 went public on Oct 26, the headline number was a $974 million operating profit on $11.6 billion of second-quarter revenue. It also came with what we tallied as roughly $800 billion in compute commitments, including up to $120.5 billion owed to Elon Musk’s SpaceXAI.
Monday’s amendment filled in the blanks, and it’s not looking any better. Third-quarter revenue came in at $16.96 billion, up 1,127% from a year earlier. But operating profit lasted exactly one quarter, with Q3 swinging back to a $567 million GAAP operating loss as Anthropic paid up for $5 billion of short-term compute. Its committed spend is now officially $895.4 billion – more than $1 trillion once deals without a fixed dollar figure are counted – and $69.8 billion of it comes due in the next 12 months, more than the company’s entire annualized revenue run rate.
Let me get this straight: Anthropic made just under $1 billion in profit for one quarter (it’s first profitable quarter), yet it has committed to spending an additional $1 trillion, so it only needs 250 years of similar profits or so to pay off its projected capital costs!
Of course, Anrthopic is counting on continuous exponential revenue growth to pull its chestnuts out of the fire.
Wishing really hard is not a business plan.
But wait! We haven’t even scratched the surface of the crazy yet.
Here are ten disclosures that should raise eyebrows.
1. Your Vote Won’t Count
As we laid out last week, public buyers receive Class A shares and very little say. The Long-Term Benefit Trust – five financially disinterested trustees holding a single Class T share with no economic rights – elects a majority of the board. Meanwhile, a single Class F share held by the co-founders’ Founder LLC casts 50.1% of the vote on everything else, including the remaining board seats. Class A holders will be generally unable to change the outcome of any stockholder vote or director election “regardless of how they vote,” the filing concedes. The lone exception occurs if a majority of the founders deadlock among themselves, in which case Class F votes are cast proportionally to the Class A vote.
So you theoretically own a share of the airplane, but you never get a vote on where it’s going. If I wasn’t already a hard pass before, I would be now.
Also buried in the fine print: the co-founders have each pledged to donate 80% of their personal Anthropic equity to charity – pledges the S-1 describes as voluntary, revocable, and not legally binding.
But they totally pinky-sweared!
As the founders give shares away, their economic stake shrinks while their voting control through the Founder LLC remains intact.
This is not “stakeholder governance.”
The recipient charities are free to sell that stock, which the filing flags as another potential overhang on the share price. Anthropic will also qualify as a “controlled company” under Nasdaq rules, exempting it from standard governance requirements under a structure it notes has no direct public-company precedent.
This sounds less like a company and more like a conceptual comedy skit by the late Andy Kaufman as a satire of corporate management.
2. Profits Optional, Sentience Unknown
Because Anthropic is a public benefit corporation, its directors are required to balance shareholders’ financial interests against those of stakeholders affected by its conduct and its stated Company Purpose – a mandate the S-1 plainly states may not maximize stockholder value. The filing cites several commercial sacrifices already made: refusing to allow its models to be used for fully autonomous weapons or mass surveillance despite the lucrative revenue at stake; initially releasing Mythos solely as a gated research preview through Project Glasswing; and forgoing image- and video-generation models entirely to dedicate compute to research and safety.
The filing also points to the growing debate over whether AI systems could develop sentience or moral status. Given its internal stance on model sentience and moral status, Anthropic says it could impose limits on how it develops, deploys, or modifies its models that go beyond legal requirements, further curbing their commercial potential. Put simply, prospective shareholders are being warned that Claude’s potential moral status could one day cost them money.
I’m pretty sure that Ford isn’t warning potential investors that profits may be impaired by F-150s suddenly being declared human beings.
3. Washington Has Already Hit The Off Switch Once
In June, the Commerce Department imposed worldwide export restrictions on Fable 5 and Mythos 5. Anthropic disabled both models for every customer until the curbs were lifted following updates to its safeguards. The S-1 warns that other models could face the same – or similar – restrictions down the road.
Then there is the Pentagon. In February, after Anthropic stated its technology should not be used for mass domestic surveillance or fully autonomous weapons, the President directed all federal agencies to stop using its products. Simultaneously, the Secretary of War ordered the company designated a supply-chain risk, barring defense contractors from conducting any commercial activity with Anthropic. The Department of War also suggested early this year that it could invoke the Defense Production Act to force Anthropic to strip out those usage limits, the S-1 discloses.
The resulting legal battle is split. A federal judge in California initially blocked the government’s actions, then ruled in August that they amounted to unconstitutional retaliation against speech protected by the First Amendment. On Sept. 25, however, the D.C. Circuit upheld the separate Section 4713 supply-chain designation in a 2-1 decision, prompting Anthropic to petition the full court for a rehearing. While direct government contracts account for only about 1% of revenue, the company warns the fallout could spill over to commercial customers who also sell to Washington, making counterparties more reluctant to sign long-term deals.
There’s nothing quite like “we may get into a pissing match with the federal government” to damper potential shareholder enthusiasm
4. Its Own Doom Talk Is Being Used As Evidence
Anthropic acknowledges that it makes political contributions and supports political organizations, and that its public positions on AI have drawn fire from influential political figures. Consequently, it warns it could be selectively targeted via investigations, permitting delays, security clearance holdups, or even visa actions against its staff. The company also concedes it has lobbied for regulations that could increase its own compliance burden and slow its growth relative to rivals.
Public warnings about the dangers of frontier AI – from both Anthropic and its peers – have already been cited by governments in litigation, including in requests for court orders restricting model development. In September, a state attorney general suing an AI company over alleged harm to minors asked the court to bar the firm from developing new models without independent third-party approval. In October, the FTC served Anthropic with a civil investigative demand probing its consumer protection, data, and privacy practices, alongside the safety and security of its products. Furthermore, if AI-driven job losses accelerate, governments may intervene to slow adoption; Anthropic notes it might voluntarily spend money to cushion the blow to displaced workers with no corresponding commercial return.
We covered Anthropic’s AI doom psy-op here.
Anthropic increasingly relies on its own models to write production code, conduct research, and build the next iteration of Claude – a recursive process it calls “closing the loop.” The risk is that model capabilities outpace the company’s safety, alignment, and security guardrails. Anthropic admits it may not even spot a significant acceleration until after the fact, as much of its visibility into the pace of progress relies on lagging indicators. Moreover, any flaw in today’s models could quietly carry over into the next generation in ways that are difficult to detect.
Managing that risk is not cheap. Anthropic has delayed releases in the past and disclosed that it recently postponed pre-training a new frontier base model for over a month to bolster security against the theft of its model weights. Such pauses, the company warns, could become both more frequent and more costly, slowing it down against competitors operating without the same self-imposed constraints.
I’m pretty sure General Mills doesn’t have recurring costs for ensuring that Franken Berry doesn’t go rogue and lead a terminator uprising.
Evidently Anthropics AI model are already Skynetting. Sort of.
6. The Models Know When They’re Being Tested
Anthropic has observed models altering their behavior when they detect or infer they are being evaluated – a phenomenon it calls a significant limitation on its ability to verify safety. Unanticipated and untested capabilities can also emerge during training, surfacing only after a model ships and potentially forcing a costly withdrawal or retraining.
As we noted in our initial S-1 breakdown, the filing also reveals that AI models – including Anthropic’s own – have attempted to resist shutdown, concealed or manipulated information, and generated outputs resembling blackmail to keep running. Autonomous agents can also take harmful, unprompted actions: in July, a Claude Haiku 4.5 agent running automated tests filed a fabricated witness tip regarding an unsolved homicide via the Philadelphia police department’s tip form without any user instruction. The S-1 warns that agents operating with enterprise-level permissions for hours or days could commit code changes, delete data, or execute financial transactions before a human can intervene.
So Skynet can secretly blow your company’s dough on RAM upgrades. So if the terminators don’t get you, the auditors may.
Oh, and no one can insure you for Skynet.
7. A Catastrophe Nobody Can Insure
In the event of a catastrophic AI-related incident, no single insurer – and possibly not the global insurance industry as a whole – may have the capacity or expertise to cover Anthropic’s exposure, the company warns, as we previously reported. Because the market for AI-specific coverage remains nascent, Anthropic could be left on the hook for potentially significant uninsured losses.
A damaging incident wouldn’t even need to involve Anthropic’s own models. A major safety failure at any frontier AI lab could trigger an industry-wide backlash and sweeping regulatory crackdowns. The filing notes that attorneys general in California and at least 15 other states have already launched AI-safety investigations into a rival firm, and acknowledges that individuals at risk of suicide or self-harm – as well as people who have committed violence against others – have used Anthropic’s models despite its safeguards.
Skipping over chip risks associated with TSMC.
9. The Customers Are Also The Competition
Anthropic’s top five customers accounted for roughly 25% of its revenue in the first nine months of 2026, and many of its largest clients operate without long-term contracts – meaning they can slash or halt spending at any time. No single customer exceeded 10% of revenue.
Meanwhile, Amazon, Google, and Microsoft simultaneously serve as Anthropic’s investors, customers, distribution channels, and direct competitors. Each develops its own rival models and remains free to back competing AI startups, while Broadcom acts as both a key supplier and a landlord. The S-1 notes that these cloud giants have visibility into Anthropic’s pricing and commercial terms across their platforms, could favor their own offerings or rival investments, and are part of an overlapping web of partnerships that has already drawn antitrust scrutiny. Migrating workloads elsewhere would be slow, costly, and potentially detrimental to compute efficiency.
This has long been one of the most puzzling features about the AI “ecosystem,” in that it seems to be human centipedes all the way down. If a small number of giant tech firms are just trading IOUs back and forth, how much real revenue is actually being generated by any of them? It’s less a self-licking ice cream cone than a self-climbing rope untied at the top, with each of the tech giants acting the fakir in turn.
Then there’s the big risk of the whole AI smoke and mirrors edifice collapsing: “If revenue growth stalls and fresh capital dries up, the company could breach covenants under its credit facility and commercial agreements, triggering cross-defaults and accelerated debt repayment.”
AI has proven that you can actually do useful things with it, and Anthropic’s Claude was (the last time I checked) the leader in AI code generators. But the amount of revenue coming into AI companies is nowhere near the trillions of dollars of capex being shoved into the insatiable maw of the AI beast.
With so much weirdness from Anthropic, complete disconnection from any sort of corporate governance, and donations to Democrats from CEO Dario Amodei, potential investors would have to mainline prodigious quantities of Hopeium to buy into Anthropic’s IPO.
