Uneasy Money: Why Anthropic's IPO Governance Has Crypto Veterans Worried

Anthropic filed its S-1 days before the recording, and the host framed the IPO as a need for public-market capital after the company appeared to have exhausted private funding. Unlike Google at its IPO, Anthropic is unprofitable. The host cited $2.7 billion in Google revenue in 2003, $5 billion for Anthropic in 2025, and $12 billion in Anthropic revenue in Q2 2026, along with Q1 revenue about 10 times Google’s—or about five times as high after an inflation adjustment—and $8.7 million in revenue per employee. The host argued Anthropic could become highly profitable if AI’s value keeps rising as chip costs fall, but said its path is less under its control than Google’s was.

The sharper concern was governance: a speaker said Anthropic’s seven founders collectively hold a side pot representing 50% of the voting shares and must coordinate their votes. The arrangement was described as unusually risky, opaque, and untested for a company that could become worth trillions, although lawyers may have designed it to preserve a unified voting block. This differs from structures associated with Mark Zuckerberg or Google, where control rests with an individual or company structure, rather than a group of seven people who must act together. The speaker warned that the founders could stop coordinating, fracture the block, and generate internal conflict that slows the company; they compared the risk to FTX and argued seven people in control could be more chaotic than one.

The purpose of the arrangement was not established. Speakers wondered whether it might function as a safety measure that lets the organization continue if some founders are incapacitated, and one said 10% of the voting power had been given to Claude, calling it an “infertile model.” The more basic worry was that founders who believe they know what is morally best could overreach. One speaker saw Anthropic’s own founders as its greatest risk: the company could succeed if they act relatively sanely and avoid trying to morally fix the world, but could also “blow” itself up. That speaker predicted China and Zuckerberg would not defeat Anthropic. These were judgments, not settled outcomes.

The speakers connected the governance concerns to crypto experience, arguing that good intentions do not prevent governance arrangements from failing and that there are more exploitable vectors now than during DeFi summer or the ICO era. They characterized Anthropic as a high-risk effort in a contest over control of intelligence, with potentially high returns commensurate with that risk. They also worried that conversational AI can concentrate influence: users receive a single answer or recommendation, which could create winner-takes-all dynamics. Unlike search advertising, users cannot currently buy prominent placement in a model’s answer weights, but content and reinforcement learning can shape recommendations in ways that may be difficult to explain. One speaker said models had repeatedly sent people with bug-bounty disputes to Anthropic’s security organization, which did not handle those disputes; the error has become less common but still occurs.

The discussion of capability was optimistic but qualified. Speakers said models are improving at difficult mathematics and code, including problems people have not solved, but still often rely on brute force and should not yet be trusted to build a house. Mathematical shortcuts could reduce computing demands for tasks such as scheduling and PCB layout, although physics may impose fundamental limits. One speaker predicted AI could soon surpass an individual chemist’s lifetime of work on some knowledge tasks because systems can be scaled horizontally. At the same time, the models remain uneven, their ultimate intelligence is unknown, and rankings change quickly: one speaker called Opus 5.5 the best model at the time and rejected the claim that Astra was better. A speaker also retracted an earlier claim about Roko’s basilisk not appearing in Anthropic’s S-1, saying it had relied on a headline and was “fake news.”

The crypto-security discussion underscored how leaks can turn into targeted risks. Speakers said Keystone exposed about 25,000 customer names and addresses, along with a device key it had promoted as protection against supply-chain attacks. Checkout records that were supposed to be deleted after 180 days had been retained; Keystone must destroy the affected devices and cannot ship them. The speakers said they did not know how the key was exposed or how it connected to the customer records, and reported no clear postmortem after about seven or eight days. They also described attackers accessing Foundation’s mail systems and sending convincing phishing messages from a legitimate address. Leaked customer data can support targeted scams, and AI tools can automate database matching, anomaly detection, and tailored messages. Their advice was to ignore unsolicited calls and messages, delete suspicious communications, and make a plan rather than react in fear; changing wallets in a panic could expose users to another company’s leak. One speaker recommended multisig with a hardware wallet as one signer instead of relying on a single 24-word seed phrase.

A tentative discussion of Glamsterdam’s cancellation raised a related process concern: a possible gas change could make a transaction fail in a way that permanently locks liquidity positions across protocols. The speaker stressed that they lacked the full picture, but framed the suspected issue as application behavior breaking when protocol behavior changes. The speakers said testing needs to cover both layers, with fuzzing, simulations, and chain replays, and saw detection before damage as a reason to improve Ethereum’s safety processes.