Anthropic's $80 billion loss is accounting fiction; actual cash burn under $10 billion
The $80 billion loss stems from convertible notes issued to Amazon and other infrastructure partners—as the company's valuation rises, these notes generate larger accounting losses on the income statement, but the cash never actually left. Total actual cash burned is only $9–10 billion.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Anthropic's S-1 targets $1.5–2 trillion valuation
The filing shows Anthropic hit $16.3 billion revenue in H1 2026, up 14x year-over-year, with Q2 alone at $11.6 billion—a 30x surge from the prior five quarters. The company aims to raise $100 billion through IPO at a valuation target of $1.5–2 trillion, which if successful would surpass SpaceX as the largest IPO in history. Sacks directly calls it one of capitalism's greatest businesses.
— Sacks$80 billion accounting loss, not real loss
Media long pointed to this loss to claim Anthropic is a cash-burning story, but the filing shows it stems from convertible notes issued to infrastructure partners like Amazon—as company valuation rises, these notes trigger larger accounting losses on the P&L, yet no cash actually leaves. Sacks points out Anthropic raised roughly $110 billion total with actual cash burn of only $9–10 billion, while $100 billion sits on the balance sheet—a company of extreme capital efficiency.
— Sacks$1 of Nvidia spending becomes $18 in revenue
Freeberg maps profit distribution across the entire AI stack: Nvidia spends $1 to make a chip and sells it for $4, with 75% gross margin; compute companies amortize chips over five years and achieve 83% gross margin; Anthropic discloses incremental gross margin above 85%. Chain it together and $1 of Nvidia spending becomes $18 in revenue—nearly every link in the chain is profitable, proving AI creates real productive value, not just capital burn.
— FreebergCompute is the only real fear in the filing
Chamath had the team count word frequency in risk factors, and they overwhelmingly point to compute—data centers, power grids, electricity, chips rank first, with regulation second; the company barely mentions customer churn or employee attrition. He draws a political inference: if Democrats want to engineer recession before 2028, the most efficient lever is blocking power and data center approvals; if Republicans want to win, they must deliver on power and data centers.
— ChamathCash is larger than expected; Chamath walks back forecast
Chamath previously predicted Anthropic would need $200–250 billion in cash reserves and a $1 trillion valuation. But reading the filing, he found the company already held $125 billion in cash by June with total net cash burn of only $9–10 billion—far more cushion than he'd forecast. So he revised: the company needs only $80–120 billion more, and IPO pricing likely lands between $1.3–1.7 trillion, with $1.5 trillion as midpoint.
— ChamathAI's top seven firms' debt now equals 30% of Treasury issuance
Chamath's deeper concern isn't whether Anthropic's business works, but that the entire AI complex now owes the market $340–400 billion, and by year-end 2027, these companies will need to raise capital equal to 30% of concurrent US Treasury issuance. This means AI is now competing with the US government for capital markets, and if competition raises borrowing costs, risk propagates through the entire industry chain—he says this is a dynamic he's only seen in financial crises over the past 15–20 years.
— ChamathAnthropic publicly calls for restraint while pursuing recursive self-improvement
Sacks notes Anthropic publicly urges the industry to decelerate while pushing recursive self-improvement (RSI) more aggressively than any company—using its own models to discover algorithm improvements and compute multipliers, the key to achieving 10x cost efficiency versus peers. In an August blog, Anthropic disclosed 26% of model iteration tasks already use RSI. Sacks suggests the public deceleration letter reads like a pre-written liability waiver—if RSI goes wrong, they can claim others refused to cooperate.
— SacksGoogle erased $500 billion in market value overnight
Google's stock shed roughly $500 billion the day after earnings as search ads disappointed for two straight quarters—search volume is climbing but click-through rates lagged. Cloud revenue doubled year-over-year to $30.7 billion, with operating profit soaring from $35 billion prior-year to over $110 billion. More troubling: a post with 5 million views accused Google of quietly abandoning the Vickrey auction mechanism, systematically overcharging advertisers. Sacks counters that search revenue actually grew 12% year-over-year from $56 billion to $63 billion; cloud just grew faster, compressing search's share below 50%.
— 多人In their own words · checked verbatim
I think this is probably going to be the biggest venture outcome of all time, especially given the time in which this was created.
Sacks5:07
And as Anthropic's valuation increased, those convertible notes got recorded as a loss on the balance sheet and that flowed through into the income statement, but it's not a real loss.
Sacks7:16
So at the end of the day, a dollar that NVIDIA spends turns into $18 that Anthropic is making, selling to customers.
Freeberg13:40
And what this tells you is the single biggest risk to the ant business is compute.
Chamath18:47
There's no single company in the world that is doing more to race towards full RSI than Anthropic. They are hell-bent for leather on getting to full recursive self-improvement.
Sacks43:40
But if now what they're doing is they've abandoned the Vickrey auction, Freeberg's right, you're over-earning.
Chamath1:07:46
So it's true that Google searches relative share decrease, but it's absolute revenue still increased.
Sacks1:11:57
Figures
| Anthropic H1 2026 revenue | $16.3 billion, up 14x year-over-year | 4:04 |
| Anthropic Q2 2026 revenue | $11.6 billion | 4:04 |
| Anthropic IPO fundraising target | $100 billion | 3:01 |
| Anthropic $80 billion accounting loss vs. actual cash burn | $9–10 billion | 7:16 |
| Anthropic cash reserves (as of June) | $125 billion | 30:08 |
| Anthropic sub-million, $100k+ annual spend customers (late September) | 11,900 | 11:37 |
| Google Cloud revenue (doubled year-over-year) | $30.7 billion | 1:01:57 |
| Google Search revenue growth | From $56 billion to $63 billion, up 12% | 1:11:57 |
| Microsoft four-year revenue and headcount change | Revenue from $19.8 billion to $33.2 billion; headcount increase of only 2,000 | 16:46 |
Glossary
- S-1 / S-1 filing
- The registration document US companies file with the SEC prior to going public
- Convertible notes
- Debt instruments convertible to equity under specified conditions; rising valuations increase mark-to-market losses on the P&L
- Recursive self-improvement (RSI)
- The process where AI systems use their own capabilities to improve the next generation of AI models
- Vickrey auction
- Second-price sealed-bid auction mechanism where the highest bidder wins at the second-highest price
- ARR
- Annual recurring revenue; a year-ahead revenue projection based on the current repeat revenue run rate
How to listen
Founders and investors tracking capital flows in AI infrastructure and the intersection of company valuations with macro policy, plus anyone trying to understand the disputed numbers in Anthropic's S-1.
The opening card game banter and closing midterm election predictions are low on information density.