The number Anthropic's filing dodges sets the IPO valuation ceiling
The filing sidestepped the one critical figure—October revenue. Whether the $1.5 trillion valuation holds depends entirely on whether that number can support the year-end extrapolation to $1 trillion annualized revenue.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Customer concentration improves, but cash flow conceals the catch
Rob's biggest surprise in the filing was positive: Q3 shows the top five customers account for only ~20% of revenue, with the single largest under 10%—proof Anthropic isn't dependent on a handful of megadeals. But the same document hides a problem. Their ~$1.5 trillion CPU rental agreement with Broadcom is classified as an operating lease, not a finance lease. This accounting move strips that cash outflow from the operating cash flow line entirely, artificially inflating the ‘positive free cash flow’ claim. Account for it correctly, and the actual cash position is worse than disclosed.
— Li ZhaoyuGrowth plummets from 144% to 45%—fine for company, not for market
Q2 revenue was $115.55 billion with a 144% quarter-over-quarter jump. Q3 hit $168 billion, but growth collapsed to 45% QoQ. The company considers 45% fine operationally. The market, though, had been linearly extrapolating outward, expecting ~$1 trillion annualized revenue by year-end. At Q3's rate of growth, year-end ARR would likely land only in the $850–950 billion range—far short of the prior bullish consensus. That gap is why the market took the stock down before October even arrived.
— Li ZhaoyuValuation above $1.5 trillion hinges entirely on October numbers
Initial expectations had IPO valuation at $2 trillion. By mid-October, some investors were circulating a downward revision to $1.5 trillion. To defend the $1.5 trillion mark, the company needs October data that can be extrapolated to ~$1 trillion year-end ARR. If October lands where Q3's extrapolation suggests—$850–950 billion—or lower still, the final price will likely be well below $1.5 trillion. Rob estimates it could drop into the $1.2–1.3 trillion range instead.
— Li ZhaoyuCloud providers take 37%—higher than Apple's tax, but worth it
The cloud provider take-rate works out to roughly 37%, higher than Apple's App Store cut of 30%. Importantly, this goes into R&D expense rather than hitting revenue as a direct reduction. Rob argues it's worth the premium: Anthropic's sales team is tiny and can't yet serve every enterprise customer directly. Cloud providers bring three values at once—the compute capacity itself, customer relationships (enterprises trust vendors they already work with), and capital (Amazon invested directly). The combination justifies the 37% rate at this stage.
— Li ZhaoyuOver $600 billion in compute contracts, but less than half are actually binding
Anthropic's disclosed long-term compute contracts add up to more than $600 billion. But the headline number obscures a crucial detail: not all of it is actually locked in. The SpaceX AI deal tops out at $1.205 trillion but remains non-cancellable only through early 2027; after that either side can exit with 90 days' notice. Nscale's $446 billion agreement can be walked away from entirely. Only the Broadcom contract—roughly $1.612 trillion—is mutual non-cancellable, and Anthropic must post guarantees and equipment collateral. The truly rigid, irreversible contracts total around $3 trillion, not $600 billion.
— Li ZhaoyuWeWork's trap: long-term cost lock, short-term revenue risk
Anthropic's longest compute contracts run 20 years, exposing the company to inflation in electricity and infrastructure costs over the entire life. Yet with many large customers, there's no matching long-term revenue commitment. Rob compares this to WeWork's broken model: cost side is a long-term lease (multi-year, non-cancellable), revenue side is a short-term lease (customers can leave anytime). If downstream demand weakens, revenue cash evaporates quickly while the long-term compute bills keep arriving. Money flowing out exceeds money flowing in.
— Li ZhaoyuCoding's gains are exhausted; the frontier is replacing knowledge workers
The coding vertical is already highly penetrated and fundamentally limited—roughly 1 million developers worldwide is the addressable pool. The real prize is replacing knowledge work. Anthropic's own measurement pegs the global payroll for knowledge workers at $40 trillion. Both Anthropic and OpenAI are now paying data-labeling firms like Mercor and Scale AI to recruit finance professionals, lawyers, doctors, and other domain experts to describe their actual workflows in detail. Those proprietary datasets are being used to train models, laying the groundwork for displacing white-collar work at scale.
— Li ZhaoyuTo spot a bubble, ignore supply—watch how steep the demand curve is
To judge whether AI is in a bubble, the key is not supply-side factors like memory or chip oversupply. It's the steepness of the demand curve. The capex trajectory is fairly locked in and predictable. What's uncertain is how fast revenue demand can rise to meet it. That gap determines how many more years the industry can sustain burning capital before the two curves intersect. Rob sees current market panic as an overreaction to recent slowness. That said, no one can be certain that giants like Amazon's confidence in future growth will actually pan out.
— Li ZhaoyuIn their own words · checked verbatim
To me, probably the biggest upside I didn't know about is that its customer concentration is actually much lower than I expected.
我觉得对我自己来说 可能最大的一个 之前不知道的一个利好的需要 就在于 它的客户集中程度 比我想象中 其实要低很多
Li Zhaoyu4:01
They say our cash position is good, free cash flow is positive. But in their free cash flow calculation, they account for that ~$1.5 trillion CPU rental agreement with Broadcom as an operating lease rather than a finance lease.
他们说我们的现金量很好 自由现金量是证损 但他自由现金量计算里面的话 他把自己跟Blockham签了一个将近150个标元 1500亿美金的这样一个租赁盘的CPU的协议 他把这东西就从会计的方式把它处理成了一个 opling list 而不是一个纯粹的financial list
Li Zhaoyu5:04
From their own perspective, the growth rate is fine. From the market's perspective, it falls far short of expectations.
增速从他自己来说的话 是OK的 从市场的预期来说 是远远不及市场预期的
Li Zhaoyu13:14
I think if they're going to reach $1.5 trillion, they should be happy with that.
我觉得现在他要做到1.5的话 我觉得他应该管理生机 就很开心了
Li Zhaoyu20:16
This is actually higher than the 30% commission many apps pay to Apple's App Store—what some call the Apple tax.
这其实比很多应用 给苹果App Store的 30%的抽成 所谓苹果税还要高了
If you add up all these contracts, there's over $600 billion. But within that, some they don't necessarily have to fulfill, and some they do.
就如果你把他这些合同加总的话 大概有超过6000亿美元 但实际上里面有一些 他并不需要一定去履约 而有一些他是需要去履约的
Li Zhaoyu37:20
The money you pay out is a long-term commitment—say, a 10-year thing or longer. But your revenue is short-term. Your costs are long-term, your revenue is short-term. If economic circumstances shift, that short-term revenue just disappears.
你付的钱是一个长期的一个保定 比如说要10年是更长期的东西 但你的收入是短租 你的成本是长租 你的收入是短租 如果经济收场发生一个变动 你短租更是不租了
Li Zhaoyu41:21
The core issue isn't analyzing your spending curve—that curve is fairly set. The core is analyzing your revenue-demand curve and how steep it actually is.
所以核心核心不在于研究 你的花钱的曲线 因为它曲线是比较确定的 你的核心是研究 你的收入的需求的这个曲线 它这样的协力有多等
Li Zhaoyu53:28
Figures
| Q3 revenue (per media reports) | $168 billion | 13:14 |
| Q2 revenue (S-1 filing) | $115.55 billion | 13:14 |
| Quarterly growth rate change | 144% → 45% | 13:14 |
| Net revenue retention (NRR) | ~900% | 24:17 |
| Cloud provider cut (estimated) | ~37% | 26:18 |
| Compute spend as % of revenue | 250% → 48% | 37:20 |
| Global knowledge-worker payroll | $40 trillion | 44:24 |
Glossary
- NRR
- Net revenue retention—measures how much more existing customers spent year-over-year, reflecting the company's ability to land upsells.
- ARR
- Annual run rate—current monthly or quarterly revenue extrapolated as a full-year annualized figure, commonly used in valuation modeling.
- FDE
- Field deployment engineer—staff a model company deploys on-site at customer locations for deep customization and implementation.
- Operating lease / Finance lease
- Two accounting treatments with different balance-sheet impact: operating leases stay off-balance-sheet and don't capitalize assets, while finance leases do, both affecting reported cash flow.
- Neo Cloud
- A mid-tier compute provider that locks in GPU capacity through long-term contracts and resells to AI companies for arbitrage.
- Bare-metal rental
- Hardware-only compute capacity leasing with no technical support or value-added services—infrastructure treated as commodity.
How to listen
Secondary-market investors and practitioners tracking AI company financial-report quality, compute contract structure, or assessing whether AI capex is overheating.
29:20–32:20: Details on data intermediaries like Snowflake locking in customer relationships. Low value for non-investment professionals—skip.