Anthropic's IPO Filing Reveals the Absurd Economics of Frontier AI

The numbers are genuinely insane

We pulled the S-1 this morning and the top line jumped out: Anthropic lost $2.7 billion in the first half of 2025 alone. Revenue hit $1.4 billion for the same period, which sounds great until you realize they spent $4.1 billion to get it. That's not a burn rate , that's a bonfire. For context, Uber at its wildest never had a revenue-to-loss ratio this inverted this far into its lifecycle. The prospectus projects $34.5 billion in revenue by 2027, which would require roughly 10x growth in two years. We've seen hockey stick forecasts before. Usually from companies selling software, not compute.

The existential risk warning isn't performative

Page 47 of the filing contains a risk factor that reads like a sci-fi novel: their models could develop capabilities that pose existential risk to humanity. This isn't marketing fluff , it's a legal liability shield. If Claude 4 or 5 does something catastrophic, Anthropic can point to page 47 and say we told you so. The clever part? They're using the same warning to justify their corporate structure. The Long-Term Benefit Trust, which can fire the board if safety is compromised, gets framed as a governance feature rather than a control mechanism. We've never seen a public company put 'might end civilization' in the risk factors section. That alone changes how regulators will treat this IPO.

Amazon and Google own the cap table

Here's what the headline misses: Amazon holds 14.7% and Google 10.2% after pouring in $8 billion combined. Neither has board seats. Both have preferred shares with liquidation preferences that pay out before common shareholders. When Anthropic goes public, those two cloud giants effectively control the float. They're not just investors , they're the compute supply chain. Every training run flows through AWS or GCP. The prospectus discloses $6.2 billion in committed cloud spend through 2027. That's not capex. That's rent paid to your biggest shareholders. We've seen this vertical integration play before with OpenAI and Microsoft, but never this explicitly documented in a public filing.

The compute moat is evaporating in real time

Anthropic's entire thesis rests on model superiority. But the prospectus admits something remarkable: they expect open-source models to reach parity with Claude 3.5 Sonnet within 18 months. That's not us reading between lines , that's their own competitive risk disclosure. If Meta releases Llama 4 at GPT-4o level next quarter, the API business becomes a commodity race to zero. The filing shows 73% of revenue comes from API calls. The other 27% is enterprise contracts, but those are mostly proof-of-concepts with one-year terms. We're watching a company price its IPO on a moat they've already told investors is draining.

Dario Amodei's voting control survives the IPO

The dual-class structure gives Amodei 20-to-1 voting power on his shares. Post-IPO, he'll control roughly 60% of votes while owning maybe 12% of equity. The Long-Term Benefit Trust gets another 15% of voting power and can appoint three of nine directors. This means public shareholders get zero say on safety decisions, model releases, or whether the company pivots to military applications. We saw this governance theater with Meta and Google. The difference? Those companies were profitable before they went public. Anthropic is asking public markets to fund a controlled company with no profit horizon and an explicit existential risk warning. The hubris is almost impressive.

The revenue concentration should terrify anyone buying the float

Three customers account for 41% of API revenue. The prospectus redacts their names but the spend patterns match Notion, Replit, and a major financial services firm we can't confirm. If any of those three churns , and they will, because they're all building their own models , revenue drops 13% overnight. The enterprise contracts have 90-day termination clauses. No minimums. No penalties. This isn't SaaS recurring revenue. It's compute reselling with a model wrapper. We've analyzed enough S-1s to know: when a company redacts customer names but discloses concentration this high, the churn has already started.

The safety research budget is a rounding error

Buried in the R&D breakdown: $280 million allocated to 'alignment and safety research' for all of 2025. That's 4.6% of total burn. The rest goes to training runs, cluster buildout, and the sales org needed to hit that $34.5 billion target. The Trust gets veto power on safety, but they have no budget authority. When the board pushes for faster releases to hit quarterly numbers , and they will, because public markets demand it , the safety team has no leverage except resignation. We've watched this movie at OpenAI. The difference is Anthropic baked the ending into the prospectus. They've literally disclosed that profit pressure will conflict with safety. Then they went public anyway.

This IPO prices the AI bubble's peak

If this prices above $30 billion , and the banks are targeting $40 billion , it sets a valuation ceiling for every other frontier lab. OpenAI's next tender will anchor to it. xAI's Grok revenue multiple will benchmark against it. But the comps don't work. Anthropic trades at 28x forward revenue with negative 190% margins. Snowflake at IPO was 60x with positive 30% margins. Datadog was 40x with breakeven. There is no public comp for 'might end humanity, definitely loses billions, compute costs controlled by shareholders.' We're not saying don't buy. We're saying: read page 47 again. Then ask yourself what happens when the next model doesn't align.