Anthropic's largest S-1 misreading: convertibles worth billions aren't actual debt
The $108 billion in convertible notes on the balance sheet is merely accounting revaluation of Google and Amazon's equity stakes as the company's valuation soared—not actual debt.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Anthropic achieves profitability at IPO, nearly unprecedented in tech history
Anthropic's S-1 shows 85% gross margins, over 20% free cash flow margins, and 8% GAAP operating profit margins. What's striking: Anthropic has raised $141 billion cumulatively (including equity, converts, and SAFEs) yet retains $134 billion in cash, meaning it has burned only $7 billion while generating $16.8 billion in quarterly revenue. Jamin Ball calls this an ‘almost accidentally profitable’ business—software companies almost never achieve positive free cash flow at IPO stage, yet Anthropic has.
— Jamin BallInference efficiency jumped sevenfold: revenue grew 14 times, compute costs just 2.3 times
Anthropic splits compute between training/internal R&D and model inference for customers; currently only 27% goes to the latter. Comparing year-over-year: inference compute spending grew 2.3x while revenue from inference grew 14x, meaning inference efficiency improved roughly sevenfold. Jamin Ball sees this as evidence of the ‘almost accidentally profitable’ business model—not driven by price increases, but the opposite: they were lowering prices in the same period.
— Jamin Ball80% of revenue comes from models launched in the past six months
In the three months ending September, 80% of revenue came from models released less than six months earlier. This exceeds expectations, showing customers have almost entirely migrated to the latest, most capable models rather than staying on older versions. This explains why staying at the frontier is nearly a matter of survival for revenue: once a model falls behind stronger competition, revenue migrates quickly to rivals. Jamin Ball expects this to persist: as open-source models capture share of total tokens consumed, proprietary frontier models capture the dollar share.
— Jamin BallThe gap between frontier and open-source models is wider than it appears
To assess the true gap between open-source and frontier models, you cannot compare only published versions, because OpenAI and Anthropic both hold unreleased, stronger models. Jamin Ball cites a chart Mark Chen from OpenAI once showed: on a famous Navier-Stokes math problem, OpenAI's unreleased internal model far surpasses the publicly released Astra. This means the true frontier-open-source gap is much wider than observers see from published products—the opposite of the common narrative that it's narrowing.
— Jamin BallAnthropic's revenue fell when others took the lead, recovered with Opus 5.5
Third-party data shows Anthropic experienced a marked ‘summer decline’: OpenAI's Astra model briefly led, and Kimi K3, GLM series, and DeepSeek collectively gained strength, diverting demand. Revenue from non-Opus 5.5 / Sonnet 5.5 models halved or worse in two to three months. Only the launch of Opus 5.5 and Sonnet 5.5 reignited revenue acceleration. Jamin Ball sees this as proof that staying at the frontier is the direct switch controlling revenue trajectory, not a nice-to-have.
— Jamin BallThe $108 billion in convertibles on the balance sheet is not actual debt
The roughly $108 billion in convertible notes payable on the balance sheet is not borrowed money but convertible notes that Google and Amazon bought years ago at prices far below today's valuation, revalued each quarter at fair value. They chose converts over direct equity to likely avoid triggering third-party ownership consolidation thresholds under public company accounting rules. As Anthropic's valuation surged, the book value of these liabilities surged with it, but they represent only about $11 billion in actual cash invested, of which roughly $5 billion has already converted to equity, leaving about $6 billion in unconverted principal on the books.
— Jamin BallThe $80 billion net loss is merely an accounting revaluation side effect
First half results show a net loss of about $8–8.1 billion, but this is not an operating loss. The balance sheet must stay in balance: when convertible notes' fair value surges, liabilities rise, so retained earnings must fall by the same amount. The change in retained earnings flows directly through the income statement as net income or loss. About $7.2 billion of this stems from valuation increases in the convertibles held by Google and Amazon, another roughly $0.9 billion from other notes and SAFEs—all of it is just the accounting reflection of equity appreciation, not actual cash losses. Had they issued equity directly instead of converts, this ‘loss’ would never appear.
— Jamin BallScaling to 30 gigawatts each in five years means $5 trillion in buildout
Jamin Ball estimates on air: if OpenAI and Anthropic each scaled from near zero to 30 gigawatts of capacity in five years, and the two together represent about 40% of total industry capacity, the industry would need about 100 gigawatts of new capacity in five years. At roughly $50 billion per gigawatt of construction cost, five years of buildout would cost about $5 trillion total, or about $1 trillion per year—roughly 5% of US GDP. That scale approaches real-world infrastructure buildout levels, contingent on credit and debt markets being deep and open enough to support it.
— Jamin BallIn their own words · checked verbatim
Maybe my favorite stat of the whole S1 was $141 billion of capital raised to date. That's inclusive of equity, converts, and safes, and a $134 billion cashback today.
Jamin Ball0:00
their compute that went towards serving models has gone up by 2.3x their revenue over the same period of time has gone up 14x so embedded in that is basically a 7x efficiency gain in serving inference
Jamin Ball17:46
The other interesting stat that Anthropic had in their S1 was for the three months ending September, 80% of their revenue was driven by models that were six months old or younger.
Jamin Ball27:09
the real way to think about or to answer that question is not the the gap between the best leading open weight models today and the frontier models today because the reality is is open and anthropic they have far more powerful models that they've developed that they just haven't released yet
Jamin Ball37:42
That debate is just misguided. There is no debt on the balance sheet. And the $11 billion of converts is inclusive in the $141 billion of total capital raised.
Jamin Ball1:12:49
And so it's kind of one of these like, technically, from Gap Accounting, they do show an $80 billion net loss. Do I view that as an accurate representation of the business? No, because the reality is, is 100% of that loss is just the appreciation of their equity, which everyone would say is a good thing.
Jamin Ball1:14:55
That's like 5% plus or minus of GDP. Like that is a massive, bonkers, massive build out, you know, rivals, the real world build out.
Jamin Ball1:29:15
Figures
| Cumulative funding raised (equity, converts, SAFEs) | $141 billion | 15:42 |
| Cash reserves remaining | $134 billion | 15:42 |
| Gross profit margin (S-1) | 85% | 14:37 |
| GAAP operating profit margin | 8% | 14:37 |
| Inference compute cost increase YoY | 2.3x | 17:46 |
| Inference revenue increase YoY | 14x | 17:46 |
| Quarterly revenue from models less than 6 months old | 80% | 27:09 |
| Convertible notes book value vs. cash invested | $108 billion vs. $11 billion | 1:12:49 |
Glossary
- S-1
- The prospectus a U.S. company files with the SEC before going public.
- Convertible notes
- Debt that automatically converts to stock if specified conditions are met.
- Jevons Paradox
- The phenomenon where efficiency gains paradoxically increase total resource consumption.
- Offtake agreement
- A long-term contract where a buyer commits to purchase a producer's output.
- Neo-cloud provider
- Emerging cloud providers that rent GPU compute specifically to AI companies.
- Compute margin
- Revenue minus the cost of compute infrastructure and inference, expressed as a margin.
How to listen
Investors, founders, and engineers trying to understand Anthropic's financial details and assess whether AI model companies actually turn a profit.
You can skip the closing conversation about Brad Gerstner and his children's investment accounts.