Anthropic, now atop the AI bubble, files for its IPO
The filing dropped like a neutron bomb on a Friday afternoon — classic bad-news timing, except the news isn't bad. It's euphoric. Anthropic, the self-styled "safety-first" AI lab that got caught running covert steganography to fingerprint Chinese competitors' models, just told the SEC it wants to go public. Valuation target: somewhere north of $40 billion. Maybe $50 billion if the roadshow catches a tailwind.
Let that sink in. A company whose flagship product Claude Sonnet 5.0 was celebrated for "heading straight down the middle of the road to dodge controversy" — safer, cheaper, and pointedly uninterested in cybersecurity — is about to become a public market bellwether for the entire generative AI sector. The irony is so thick you could spread it on toast.
The safety theater was always the business model
Anthropic's founding myth goes like this: Dario and Daniela Amodei left OpenAI because Sam Altman wasn't serious enough about alignment. They'd build the responsible alternative. Constitutional AI. Interpretability research. A public-benefit corporate structure that would legally subordinate profit to safety.
Then the money arrived. Google poured in $2 billion. Amazon followed with $4 billion. The public-benefit structure stayed, but the "benefit" got redefined: the benefit of not getting crushed by OpenAI, Microsoft, and Meta. The constitutional principles got filed next to the steganography code — "Oh, yeah, we've been meaning to disable our secret steganography system" — right before the journalists found it.
That's the company going public. Not a safety research institute. A venture-scale bets-hedging operation that built a very good model, wrapped it in the language of responsibility, and now needs liquidity for investors who've been patient since 2021.
The financials will tell a different story than the mission statement
The S-1, when it drops, will reveal what the PBC structure obscured: Anthropic burns cash at a rate that would make WeWork blush. Training frontier models costs hundreds of millions per run. Inference margins are compressing as Cloudflare and others fight the scraper wars that make training data expensive. Enterprise adoption is real but lumpy — pilots don't pay the GPU bills.
Revenue? Probably in the low hundreds of millions ARR. Losses? Multiply that by five or ten. The unit economics of foundation models remain unproven at scale. Every_tokens generated is a marginal cost; every enterprise contract is a sales cycle. The moat is inconveniently shallow — open weights from Meta, Mistral, and eventually China's labs keep narrowing the quality gap.
But the IPO isn't about fundamentals. It's about narrative timing. Anthropic is filing while the UN warns of an "AI-pocalypse" without global governance, while Oracle fills 10-Ks with AI risk factors, while the hype cycle peaks and the smart money starts looking for exits. The Amodeis know the window closes when the first serious open-weight GPT-5-class model drops. Or when enterprise buyers realize RAG on Llama 4 is 90% of Claude for 10% of the API cost.
The PBC fig leaf won't survive public markets
Here's the structural trap: Anthropic incorporated as a Public Benefit Corporation in Delaware. That means directors must balance shareholder value against the stated public benefit — "responsible AI development." In private markets, that's a marketing asset. In public markets, it's a litigation magnet.
Every quarterly miss will spawn a derivative lawsuit claiming the board prioritized safety over returns. Every safety delay — red-teaming, interpretability audits, constitutional alignment checks — will be framed as breach of fiduciary duty. The Amodeis will discover that "balance" is a word plaintiff's attorneys love and judges hate.
They'll either gut the mission to satisfy Wall Street, or they'll get sued into paralysis. There's no third option. Ask the Ben & Jerry's board how "mission protection" worked under Unilever. Ask the Patagonia founder why he gave the company away rather than IPO.
The bubble has found its flagship
Anthropic's IPO will be the moment the AI bubble gets its ticker symbol. Not Nvidia — that's the picks-and-shovels play, already priced for perfection. Not Microsoft or Google — they're diversified conglomerates with AI as one narrative among many. Anthropic is pure play. Pure narrative. Pure multiple expansion on a story that stops working the moment the story changes.
The underwriters will pitch "first-mover in constitutional AI" and "enterprise trust moat." They'll downplay the steganography stunt, the milquetoast 5.0 release, the fundamental commoditization of intelligence. They'll compare it to Google's 2004 IPO — mission-driven, technically superior, destined for dominance.
But Google had a business model before it went public. Anthropic has a burn rate and a prayer.
What comes after the pop
The offering will almost certainly succeed. Retail wants AI exposure. Institutions need marks on their venture portfolios. The float will be managed, the pop engineered, the lockup timed. Early investors will distribute. The Amodeis will stay wealthy and influential.
Then the quarterly grind begins. Analysts will model token economics. Competitors will undercut on price. Open weights will improve. The "safety premium" enterprises supposedly pay will prove elastic. And the public-benefit directors will sit in Delaware courtrooms explaining why a $200 million interpretability research program served shareholder value.
We've seen this movie. The mission-driven unicorn goes public. The mission gets revised. The founders leave or get sidelined. The company becomes what it claimed to oppose.
Anthropic's IPO isn't a victory for responsible AI. It's the moment responsible AI gets priced, packaged, and sold to the highest bidder. The steganography code was just the preview — covert, deceptive, disabled only when caught. The S-1 is the same logic at scale.
Buy the dip if you want. But read the risk factors first. Oracle wrote pages of them. Anthropic's will be longer.