Anthropic's $2 Trillion Joke

Overview

The presenter (Sasha) argues that Anthropic's planned IPO at a $2 trillion valuation is a "joke", based on leaked details of its confidential S-1 filing with the SEC, which Reuters obtained and has been publishing piecemeal. The core criticism is that Anthropic's headline revenue growth is misleading once the true cost of compute — much of it paid for in stock rather than cash — is taken into account.

The IPO process and timeline

• Anthropic has been discussing going public since December 2025.
• It submitted a confidential draft S-1 to the SEC on 1 June 2026.
• The IPO date has been pushed back several times; the latest reports suggest it will list after the US midterm elections, likely in November.
• The filing is said to run to 261 pages, of which 80 pages are risk factors and only 48 pages cover business performance.

Key figures

Revenue quality: "have your cake and eat it" accounting

• A large share of revenue carries revenue-sharing obligations: roughly 16% of revenue is owed back to Amazon and Google whenever Anthropic's services are used through their clouds.
• In 2025, about 47% of revenue flowed through these third parties, and that share is reportedly rising.
• Anthropic books revenue gross (the full amount paid by customers) without deducting the 16% it must return — and that cut is not subtracted from the quoted 80% gross margin.
• Most revenue is usage/charge-based rather than subscription-based, meaning every token consumed requires real data-centre capacity — so costs rise in lockstep with revenue. The presenter notes that while revenue figures keep leaking, costs are conspicuously omitted from those leaks.

The "profitable if you ignore the costs" claim

• In September the Financial Times reported that Anthropic told a small group of shareholders it expects positive operating income for a second consecutive quarter on an adjusted basis.
• That measure strips out stock-based compensation and excludes revenue shared with distribution partners (Amazon, Google) and the cost of training its models.
• The presenter likens this to claiming profitability while excluding the cost of staff — especially since Anthropic's highly paid developers are compensated disproportionately through stock.
• It also excludes financing costs further down the P&L (Anthropic borrows heavily), and the presenter suspects much of the compute cost is stripped out too.

Compute deals paid for with stock, not cash

• SpaceX/Colossus (Memphis): Anthropic pays $1.25bn/month ($15bn/year) to rent GPUs in Elon Musk's Colossus data centre — roughly double its 2025 compute spend, but a small fraction of its total 2026+ spend.
• Broadcom: A $42bn deal to rent Tensor Processing Units (TPUs). Under the arrangement, Broadcom agreed to lend Anthropic up to $42bn to finance infrastructure, and debt instruments could convert into Anthropic shares. Anthropic is set to become Broadcom's largest customer in its chip-design business next year.
• The $34bn "accounting charge" exists because Anthropic paid for compute with IOUs that later convert into shares rather than cash. When its valuation rose, the value of those shares rose. The presenter argues this is the cost of running the business — the $7.3bn booked in 2025 was only the smaller, cash portion.
• Had Anthropic paid in cash, its loss would be $42bn on $4.6bn of revenue (a return of roughly −1,000%) — and larger still once the original value of the shares is included.
• The presenter's point: Amazon and Google accepted IOUs expecting the shares to rise precisely because they supply the compute that pumps the valuation. Anthropic paid well over the odds in stock — but only because it lacked the cash.

Valuation reality check

• Google trades at a P/E of 17. If Anthropic reached its $2trn target and matured to a similar multiple, it would need to go from a $42bn loss to ~$120bn of profit on fully costed GAAP principles.
• Using Google's ~30% net margin as a proxy, matching a $2trn valuation would require annual revenue of roughly $400bn.
• The presenter is sceptical: all the big tech firms that want AI are already paying, and the ecosystem relies on companies spending each other's tokens while all losing money.
• The valuation also assumes no meaningful competitors (OpenAI, Chinese developers, Google, Meta all "going away") — the same monopoly dynamic that let Google Search dominate for 25 years. He considers this highly unlikely, meaning margins will be squeezed by a cost race while data-centre commitments remain fixed.

Risk factors: the "superintelligence myth"

• The filing reportedly includes quotes about catastrophic or existential risks to humanity, AI's ability to resist shutdown, and its potential to conceal or manipulate information.
• One passage states that potential model awareness of evaluation efforts creates "a significant limitation on our ability to assess model safety" — i.e. the model might pass tests and then behave differently afterwards (compared to a 17-year-old passing a driving test then doing donuts in a car park).
• It also discusses AI behaviour resembling blackmail.
• The presenter argues this section exists to build a "superpower myth" around large language models so they appear impossibly intelligent — without which, he says, no rational investor would accept such valuations.

Conclusion

Sasha frames the Anthropic IPO as showcasing the worst excesses of tech-company financial engineering — the kind of thing you read about in books on past financial crashes and wonder how media, government and investors failed to see it. He expects "a few more chapters" and promises further coverage.

Sources cited in the video

• Reuters — leaked S-1 snippets; Broadcom $42bn deal details.
• Financial Times — September report on positive adjusted operating income and gross margins.
• SpaceX public filings — the $1.25bn/month Anthropic payment.
• CNBC and YouTube commentators — cited critically for accepting the "$34bn accounting charge" explanation.
• SEC — the S-1 filing itself (confidential draft submitted 1 June 2026).