Anthropic's Q3 slowdown was compute-constrained, not demand-constrained
The S-1 reveals Q3 had to fill half its compute purchases with a $500M high-cost spot market—long-term capacity was fully booked, not because customer demand for tokens dried up.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Existential risks in passive voice: hedging for both audiences
Zvi notes that Anthropic's S-1 discussion of existential risk uses only passive voice and glosses over it—attempting to tell two incompatible stories at once: ‘we're a highly profitable compute supercomputer company’ to investors, while elsewhere claiming ‘what we built could kill everyone.’ Lawyers instinctively minimize disclosure since S-1 filings become evidence in future litigation. But Zvi argues the silence is itself revealing: if you genuinely believed your stated existential risks, choosing to stay quiet in the document where you should be most transparent says something about who you actually are.
— Zvi MowshowitzThree founders can maintain company control indefinitely if they stay
Anthropic created a Class F shareholder class pooling founder votes with 50.1% voting control on broad corporate matters. The company has roughly seven formal founders, but this super-voting block persists as long as just three remain employed—sufficient, by Zvi's estimate, to outlast ‘the point when the dust settles on the world's fate.’ The company also chartered a long-term giving trust nominally controlling four board seats: the board selects three, the trust selects three, and a seventh seat requires founder approval or else goes to Class E shareholders (employees who've donated stock to charity), who are assumed to align with the founders.
— Zvi MowshowitzUncapped 3:1 giving match may be capitalism's costliest employee perk
The founders committed to donating 80% of their personal shares to charity, coupled with an employee giving-match program offering 3:1 matching without limit—unprecedented in capitalism's history. Standard companies offer 1:1 matching capped at a few thousand dollars. Anthropic employees can donate $20M to charity and trigger a $60M company match. With thousands of employees holding nine-figure equity stakes, this may become one of capitalism's single most expensive unilateral decisions. The Q3 non-cash charge of ~$24B in unattributed equity compensation hints at what's coming, and Zvi expects it won't be the last.
— Patrick McKenzieQ3's slowdown was compute-constrained, not demand-constrained
Analysts widely interpreted Anthropic's Q3 revenue slowdown as evidence demand has plateaued. Patrick's retrospective reading is opposite: the S-1 makes plain the company faces chronic compute constraints, with procurement requiring 12–18 months lead time—meaning they had to commit in early 2025 to compute volumes for models not yet released. The smoking gun: they spent $5B on spot-market compute that quarter, representing fully half their compute spending, well above normal—signaling long-term capacity was fully booked and customer token demand remained queued.
— Patrick McKenzie$141B raised, ~$10B burned, yet carrying ~$800B in compute commitments
Anthropic raised $141B cumulatively and retains $131B in cash and investments; actual cash consumption totals roughly $10B. Hundreds of billions in reported losses stem mainly from mark-to-market losses on convertible debt as stock price climbed—Zvi frames this as pure accounting convention, not real cash hemorrhage. The material number is the ~$800B in forward compute commitments on the balance sheet—effectively spreading years of massive future spending across current-quarter reports, structurally similar to a massive real-estate lease obligation.
— Zvi MowshowitzWall Street and tech circles price the same company on completely different books
Wall Street values Anthropic as ‘a very impressive software company.’ Tech insiders want to shake Wall Street awake and say: Google's entire commercial history might be merely this business's bootloader. Google would exhaust its entire free cash flow, max out its credit lines, just to own this franchise—not just to defend search from disintermediation, but because this business exceeds every business in capitalist history. Investment banks haven't yet priced that thesis into their DCF models. Zvi thinks that omission is precisely where a credible bear case finds its foundation.
— Patrick McKenzieNo one taught it Japanese—the model figured it out on its own
Patrick witnessed a research lab demo an unreleased model that had explicitly not been trained on Japanese, with zero in-house Japanese training data involved. Yet the researchers suspected the model had reverse-engineered Japanese grammar and structure from raw English-language internet content alone. Patrick tested it on a CNN news segment the training data couldn't have seen, and—as someone with 20 years' experience as a professional translator—said the model's Japanese might surpass his own. This exemplifies the Bitter Lesson: scale up the model and unexpected capabilities emerge that nobody trained it to have.
— Patrick McKenzieBinance lost $5B to hackers; the industry's security debt runs $500B
Hours before recording, North Korean hackers breached Binance for over $5B. Patrick argues that the AI-enabled hacking wave won't primarily wield research-lab models themselves, but rather open-source weights distilled from those labs. He estimates the industry's total security technical debt at roughly $500B (priced at pre-AI rates), a sum payable only by betting everything on AI-powered defense. Zvi's outlook is darker: this isn't a one-time invoice because any operator can now parallelize attack probes; vulnerability discovery will compound.
— Patrick McKenzieIn their own words · checked verbatim
my view essentially is that if we do too much of that, that there's a very good chance that, in fact, AI will kill everyone, either philanthropic or otherwise.
Zvi Mowshowitz6:21
if you choose not to make a costly signal, you've told me who you are. You've told me you're not willing to make a sacrifice that's meaningful.
Zvi Mowshowitz7:24
But their matching program for employees might be one of the most expensive single corporate decisions ever made in the history of capitalism.
Patrick McKenzie11:33
Their Q3 revenue softening isn't a demand story, it's a supply story. They're running all the GPUs they have, and it just isn't enough.
Patrick McKenzie23:26
Like, it's possible that the entire business history of Google to this point was a bootloader for this business.
Patrick McKenzie50:08
So after, you know, 20 years of devoting a significant portion of all my cycles to getting good at that, you achieved it by accident. Wonderful.
Patrick McKenzie1:01:49
We have already passed the portion at which AI is more than 5% of the GDP of a nation. It's more than 5% of the GDP of North Korea by enabling them to innovate in their business model of stealing money from crypto people.
Patrick McKenzie1:13:09
Figures
| Cumulative funding | $141B | 31:56 |
| Cash and investments on balance sheet | $131B | 31:56 |
| Actual cash burn | ~$10B | 31:56 |
| Forward compute commitments | ~$800B | 31:56 |
| Q3 spot-market compute spending | $5B (50% of quarterly spend) | 22:25 |
| First-half revenue growth rate | ~15x | 18:29 |
| Unattributed equity compensation | ~$24B | 14:18 |
| Binance hack loss | $5B+ | 1:12:02 |
| Industry security tech debt estimate | ~$500B | 1:12:02 |
Glossary
- S-1
- SEC registration form filed by US companies before IPO, disclosing financials and material risks
- closing the loop
- business term for recursive self-improvement: AI autonomously refining AI systems
- pacing the frontier
- industry self-regulation principle: all labs advance at a coordinated speed rather than uncontrolled racing
- cohort economics
- tracking spending curves by customer cohort entry date to gauge whether growth has plateaued
- net revenue retention
- metric measuring how much existing customers increase spending over time; above 100% means growing spend per customer
- bootloader
- metaphor: an apparently independent legacy business that merely bootstraps a vastly larger new one
How to listen
Investors, founders, and engineers tracking Anthropic's valuation, compute supply chains, and AI safety governance who want to read signals from S-1 details rather than news headlines.
The opening disclaimer about the host's personal investment positions.