185: Anthropic招股书丨与Amont李肇宇聊定价逻辑和潜在风险

1. Key Themes

The Debate Is About Demand, Not Supply

Rob Li, who runs Amont Partners and holds AI compute-layer positions, repeatedly pushes the audience away from the popular "funding/supply" framing of the AI bubble question. His argument is that the cost curve is knowable and the revenue curve is the unknown. He says: "核心其实不在供给端。我觉得很多人在讲供给供给供给，funding funding funding，核心的核心我觉得还是在需求端" — in English: "the core isn't on the supply side. A lot of people talk supply, supply, supply, funding, funding, funding, but the core of the core is still demand" [00:53:22]. He frames it as two curves: "you're spending curve is quite certain... the core is studying the revenue/demand curve and how steep its slope is" [00:54:09]. The only open question is whether the curves intersect in 2029, 2030, or later.

Customer Concentration Is Healthier Than Expected

Li's biggest positive surprise from the S-1 was how diversified Anthropic's revenue base is. He says: "他的客户集中程度比我想象中是要低很多...面前五大客户已经将近只占他的大概1/5了" ("customer concentration is much lower than I imagined... the top five customers now account for only about one-fifth") [00:04:17]. Average top-five share is under five points each, and the largest customer is under ten points [00:04:27]. This is a key de-risking datapoint, since the model-lab business had been feared as dependent on a handful of coding-tool buyers.

Revenue Growth Is Fine for the Business, Badly Short of Market Expectations

Quarterly revenue growth fell from 144% in Q2 to 45% in Q3 (Q3 $16.8B per media reports vs. Q2 $11.555B in the S-1). Li: "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" ("fine from its own standpoint, far short of market expectations") [00:13:16]. Linear extrapolation from Q3 implies only an $85–90B year-end ARR run-rate vs. the $100B+ the market was hoping for [00:13:44]. The October ARR figure, absent from the first S-1, is the swing factor; Li argues that if management has it and withholds it, "市场的推断...一定是觉得你这个数字不好，所以你才不愿意跟我讲" ("the market will infer the number is bad, which is why they won't tell me") [00:07:38].

The Valuation Haircut: $2T → $1.5T → Possibly $1.2–1.3T

Initial expectations were $2T, then reports of a cut to $1.5T in mid-October. Li's view: "现在他要做到了1.5的话，我觉得他应该管理层已经就很开心了" ("if he gets to 1.5 now, management should already be very happy") [00:20:33]. A strong October number (extrapolating to $100B run-rate) could support $1.5T+; a weak one means "能不能做到1.2到1.3都是一个问题" ("even $1.2–1.3T is a question") [00:22:19]. Host Chen Manqi's math supports the tension: reaching a $100B year-end ARR needs October ARR of about $90.4B at a ~34% monthly pace [00:21:32].

Compute Commitments: Headline Numbers Overstate the Hard Obligations

The S-1 headline compute commitments exceed $600B, but Li argues the key is which are actually binding. SpaceX AI/xAI (up to $120.5B): "你不要用的话，你可以提前三个月，90天通知，你可以不要的" ("if you don't want to use it, you can give 90 days' notice and walk away") [00:34:19]. Neoclouds like Nscale (up to $44.6B) are pay-for-what-you-use. The hard commitment is Broadcom ($161.2B), non-cancellable except for default, with Anthropic providing guarantees and equipment collateral [00:35:33]. Manqi estimates the rigid contracts still total roughly $300B+ [00:37:31]. Li's takeaway: "核心不在于那个数字有多大还是多小，核心在于这个数字是否能够兑现" ("the core isn't how big the number is, it's whether it will actually have to be honored") [00:35:00].

Unit Economics Are Improving, but Compute Spend Needs to Fall Further

Two years ago, $1 of compute produced under $1 of revenue; now it produces over $2 by Anthropic's own calculation [00:08:50]. Compute spend as a share of revenue fell from 250% to 48% [00:37:39]. Li's "healthy" threshold: "至少应该降到1/5，就20到25点之内算能够算比较健康" ("should fall to at least one-fifth, i.e., within 20–25 points") [00:38:01]. He lists his three long-term tracking metrics as customer-concentration mix, same-store (net revenue retention) growth among big spenders, and revenue per unit of compute [00:08:00].

The Customer Relationship Belongs to the Clouds and Data Platforms, Not the Labs

Li, a long-time Snowflake shareholder, argues that enterprise token consumption mostly flows through AWS/Snowflake/Databricks rather than directly through Anthropic or OpenAI. He gives three reasons: (1) Anthropic can't staff integration support for all large clients, (2) data security and CIO career risk ("我可以找亚马逊找说法" — "I can go to Amazon for accountability") [00:30:57], and (3) Snowflake's data context makes token use more efficient: "你要需要烧一百个token，你通过我可能只需要烧80个token" ("if directly you'd burn 100 tokens, through me you might need only 80") [00:31:32]. This explains why Anthropic pays ~37% channel fees to cloud vendors, which Li considers worth it: "他们能够提供客户，能够提供钱，还能够提供基础设施" ("they can provide customers, money, and infrastructure") [00:27:54].

The Duration Mismatch Is a WeWork-Style Risk

Anthropic has compute contracts up to 20 years but customers with no long-term commitments. Li: "跟当年我们说WeWork风险是同样的风险...你的成本是长租，你的收入是短租" ("the same risk as WeWork... your costs are long leases, your revenue is short leases") [00:41:59]. He ties it back to the demand question, not supply.

What Comes After Coding? White-Collar Knowledge Work

Coding is already highly penetrated and the market is limited. Li says the next leg "核心就在于替代白领的工作" ("is fundamentally replacing white-collar work") [00:43:45]. The S-1 itself cites $40T in global knowledge-worker wages. Li sees the lab strategy visible in the data-labeling vendors: "你可以去跟踪...Mercor啊，或者像Scale AI，他们每天在干啥，你就知道现在Anthropic、OpenAI他们的重心在什么地方" ("track what Mercor and Scale AI do daily and you know where Anthropic and OpenAI are focused") [00:44:45]. Standardizable fields (e.g., legal work via Harvey) will move faster than non-standard ones like investing [00:46:21].

Why the AI Bubble Is Not 2008

Li calls the "eve of the burst" talk "很明显是过度恐慌" ("obviously over-panic") [00:54:31]. His comparison to 2008: AI is under ~3% of US GDP (2-and-a-bit points) versus housing at a far larger share, and the financing runs through private credit, sponsors, and PE rather than banks [00:58:30]. "只要美国的银行系统不爆裂来说的话，美国经济很难变成08年、09年这样一个情况" ("as long as the banking system doesn't rupture, the US economy can hardly become a 2008–09 situation") [00:59:11].

2. Contrarian Perspectives

"10万美元以上的500强客户" Is Not Saturation, It's Under-Penetration

Many read "over 50% of the Fortune 500 pay over $100K" as a sign of saturation. Li flips it: "你不知道他所谓超过10万美元，是花了11万美元，还是花了一亿美元...单个客户他从花10万美金变成花1000万美金，本身就是一个100倍的增长" ("you don't know if 'over $100K' means $110K or $100M... a customer going from $100K to $10M is itself a 100x increase") [00:23:34]. If most are near the floor, penetration upside is huge. Manqi adds that the S-1 reports ~250 customers above $10M annually.

900% NRR Is Nearly Meaningless Without the Base

Anthropic's net revenue retention is ~900% at end of September, which looks spectacular versus the 120–130% of strong SaaS firms. Li's contrarian read: "如果这个数字我们不知道它的基数的话，很多时候会变成没有意义" ("if we don't know the base, the number often becomes meaningless") [00:25:09]. Because Anthropic's revenue only exploded from late 2025 to H1 2026, the prior-year base is tiny. The real signal is how fast NRR decays: if it falls from 900% to 200–300% within one year, "大家觉得这个市场本来是星辰大海，后来发现短短的一年时间...已经变成了一个比较成熟的公司一个增速了" ("people thought the market was boundless, then found within a year it had the growth profile of a mature company") [00:26:01].

Anthropic Has No Good Public Comp: It Is Closer to a Heavy-Asset Business

Most investors benchmark frontier labs against software. Li says the marginal-cost structure breaks that: "我是Facebook的话...我新增了一个用户，我服务他的费用基本上是零。而如果我是一个模型厂...第1000个客户和第一个客户...服务的成本其实是比较类似的" ("if I'm Facebook, serving a new user costs basically zero; as a model lab, the 1,000th customer costs about the same as the first") [00:50:30]. Therefore, "它其实更类似于我们说一些重资产的一些领域...跟云有点类似" ("it's actually more like heavy-asset sectors... somewhat like cloud") [00:50:55].

A Missed Number Can Be Neutral or Even Good for Upstream Suppliers

A drop in Anthropic's growth is not automatically bearish for infrastructure. Li's analogy: if XPeng sells 90 cars instead of an expected 100 because BYD, NIO, and Li Auto took share, the EV market (and CATL, the battery supplier) can still be fine. "只要这个盘子的增速还在持续往上的，那中间你们几个个人的市场占有率的起起伏伏，对于上游来说的话，其实不是一个很重要的因素" ("as long as the overall pie keeps growing, share shifts among a few players aren't important to upstream") [00:19:00]. He notes open-source models have already passed half of token burn on OpenRouter [00:18:36].

Accounting for "Positive Free Cash Flow" Is a Trick

Management claims positive free cash flow, but Li says the $150B Broadcom TPU capacity deal was treated as an operating lease rather than a financing lease. "把很大一块钱的现金的支出移了出去...实际上...他实际的自由现金流的情况会比他告诉大家的会有要差一些...管理层玩了一个就会计方面的这样一个小把戏" ("moving a large cash outlay out... the real free cash flow is worse than they tell people... management played an accounting trick") [00:05:14]. Note that Li, a Buffett-style investor, mixes all cash flows together; on that basis, "基本上就还在亏钱的" ("basically still losing money") [00:49:08].

3. Companies Identified

Anthropic

Frontier AI lab behind Claude, filing for IPO via S-1. Discussed as the central subject: customer concentration improved (top five ≈ one-fifth of revenue), unit compute economics doubled, $130B+ in cash and short-term investments. Li: "对我自己来说可能最大的一个之前不知道的一个利好的信息，在于他的客户集中程度比我想象中是要低很多" ("the biggest previously unknown positive is that customer concentration is much lower than imagined") [00:04:17]. Manqi on liquidity: "现金及现金等价物，还有短期投资现在是超过1300亿美元" [00:49:36].

Broadcom

Chip and infrastructure supplier; effectively the route to Google's TPU capacity for Anthropic (~$161.2B contract). Mentioned because it can actually deliver capacity and therefore holds negotiating leverage. Li: "博通你可以把它想象成就是Google一个手套...他是能够实打实的拿到TPU的货，而且他自己能够在TSMC拿到足够多的wafer的一个capacity" ("Broadcom is basically Google's glove... it can actually secure TPU supply and enough TSMC wafer capacity") [00:36:36].

Micron

Memory maker; a long-term position of Amont Partners since around 2022–2023. Li uses it to illustrate drawdown tolerance: "美光从30块钱2022年到现在将近100块钱的话，至少经历过五、六次的就超过三成的...回撤" ("Micron went from $30 in 2022 to nearly $100, with at least five or six drawdowns over 30%") [00:55:30]. He also argues memory is a moat business: "全世界就三个人可以做...这个缺口是在逐渐放大的，明年比今年更大" ("only three players in the world... the gap is widening, bigger next year than this year") [00:57:12].

Semiverse

California company doing contract manufacturing for Google's TPUs; held by Amont since ~2020–2023. Li: "给Google TPU做代工的一个企业，叫做Semiverse" ("a company that does contract manufacturing for Google TPU, called Semiverse") [00:03:17].

Snowflake

Cloud data platform in which Amont has been a shareholder for many years. Mentioned as a leading indicator of real enterprise token burn because customers route AI calls through its router. Li: "如果你是Snowflake的话，你会...实时的看得到你的企业客户到底每天在干嘛？烧得多还烧得少？" ("if you're Snowflake you can see in real time what enterprise customers are doing—burning more or less?") [00:40:25].

Databricks

Data and AI platform, named alongside Snowflake as a coincident indicator and as an owner of enterprise customer relationships. Li: "客户的关系是在他们里面的...比如说Snowflake和Databricks" ("the customer relationships sit with them, e.g., Snowflake and Databricks") [00:29:21].

Amazon (AWS)

Anthropic's major investor and cloud channel; its guidance is the bull-case evidence. Li: "今天亚马逊的财报里面他也谈到，他说27年、28年我们的云这边每年的增速都是要40个点...AI token收入的增速明年、两年都是40个点" ("Amazon says cloud will grow 40 points a year in 2027 and 2028... AI token revenue growing 40 points next year and the year after") [00:42:43].

Harvey

Vertical AI tool for legal work built on frontier-lab models. Li: "用最有名的一个叫做Harvey...基本上大的律所都在使用它" ("the most famous is Harvey... basically big law firms all use it") [00:46:57]. Mentioned as an example of standardizable white-collar work being automated first.

Mercor and Scale AI

Data-labeling companies hiring finance, legal, and medical professionals to generate private-domain training data. Li: "他们现在在招很多有金融、法律、医学...有技能的人，让他们去参与去清洗这样一个数据" ("they're recruiting many people with finance, legal, medical skills to help clean this data") [00:45:08]. Also a window into where labs are heading.

CoreWeave and Nebius

Neoclouds that Li considers to have some moat because of software stacks. "除了前面两个人，他们有一些software stock，还算有壁垒，CoreWeave和Nebius" ("besides these two, which have some software stack and thus some moat, CoreWeave and Nebius") [00:56:59].

Nscale

A bare-metal neocloud preparing to list; Li calls its model a "房地产的生意" ("real estate business") with no technological content [00:56:59]. It has a contract with Anthropic of up to $44.6B that doesn't have to be fully honored.

SpaceX AI / xAI (Elon Musk's compute)

Provider of up to $120.5B in compute to Anthropic, with a 90-day cancellation right after early 2027. Li flags its cancellability as a structure that favors Anthropic [00:34:09].

Google / Google Cloud / TPU

Reported Q3 with negative free cash flow of $9.07B (second consecutive negative quarter). Li attributes its stock fall mainly to ad growth slowing to ~10% vs. expected 14–15% and cloud growth of 103% vs. 114–115% expected [00:10:52; 00:41:09]. Also the ultimate TPU source behind the Broadcom deal. Li notes Google Cloud remains behind Microsoft and Amazon in enterprise sales [00:32:48].

OpenAI

Frontier lab, a competitor; notably recognizes revenue net of cloud-partner fees, unlike Anthropic's gross approach. Manqi: "OpenAI在算收入包括ARR的时候，它是算的净收入" [00:28:10]. Li also notes OpenAI's September model release ("GPT-6/Astra" per Manqi) affects Anthropic's quarterly growth narrative [00:16:20].

Moonshot's Kimi and open-source models

Mentioned by Manqi as Chinese open-source models that approached top-tier performance and shook valuation narratives; Li adds open-source now exceeds half of token burn on OpenRouter [00:18:36].

CATL (宁德时代), XPeng, BYD, NIO, Li Auto

Used in Li's analogy of upstream suppliers vs. downstream share shifts [00:17:31].

WeWork

Used as a cautionary analogy for the long-lease/short-revenue mismatch [00:41:59].

Microsoft

Cited as stronger than Google Cloud in enterprise sales [00:32:48].

Meta and Nvidia

Compared on cash: Anthropic's >$130B cash pile is less than Amazon's, Nvidia's, and Meta's [00:49:36].

4. People Identified

Rob Li (李肇宇)

Managing partner of Amont Partners; founded the fund in summer 2023. Previously a technology investor at a Tiger Cub, then a partner at J.P. Morgan Asset Management's L/S fund, with some activist investing, and also earlier at Morgan Stanley and Slate Path per the host introduction. Buffett-style: undervalued, no leverage, conservative, long holding periods. Mentioned as the guest whose framework (demand over supply, no leverage, ten-year horizon) shapes the discussion. "核心理念的话还是比较偏巴菲特的一个风格，低估和杠杆，相对比较保守" ("the core philosophy leans toward Buffett's style, undervaluation, conservative") [00:03:02].

Dario Amodei

Anthropic CEO. Li uses him to make the point that nobody truly knows long-term demand: "我并不觉得全世界有任何一个人，Dario也好，Sam Altman也好，亚马逊也好...心里面对这个东西有一个很明确的答案...他们可能有一个信仰" ("I don't think anyone, Dario, Sam Altman, or Amazon, has a clear answer... they may have a belief") [00:52:51].

Sam Altman

OpenAI CEO, mentioned in the same remark as someone operating on belief rather than certainty [00:52:51].

Warren Buffett

Cited as Li's stylistic anchor and also as an example of non-standardized investment judgment that AI will struggle to learn: "巴菲特的判断是A，另外一个很有名的投资人...是B" ("Buffett's judgment is A, another famous investor's is B") [00:46:21].

Elon Musk

Referred to as the large compute provider via SpaceX AI/xAI; Li notes the contract is more favorable to Anthropic than its headline suggests (90-day exit) [00:34:09].

程曼祺 (Chen Manqi)

Host from LateTalk, supplied much of the supplementary quantitative work: channel-fee estimate (~37%), October ARR threshold ($90.4B), rigid-vs-flexible compute breakdown, and the revenue-recognition difference between OpenAI and Anthropic [00:21:32; 00:28:10; 00:35:22].

5. Operating Insights

Use Channel Partners as Capital, Distribution, and Capacity, Not Just Sales

Li's explanation for Anthropic's ~37% cloud channel fee is a template for scaling without a sales org: with a small sales team and no time to build one, "你只能靠渠道...第一，算力他们可以给你提供...第二，他们可以提供客户；第三，他们可以给你提供资金" ("you can only rely on channels... they provide compute, customers, and money") [00:27:18]. The fee buys three things at once: infrastructure, distribution, and balance-sheet support (guarantees, leasing structures) [00:27:42].

Sit Where the Enterprise Trust and Context Already Live

Li's Snowflake lesson: enterprise buyers avoid direct vendor relationships because of integration support limits, data-security anxiety, and personal CIO career risk, and prefer a trusted intermediary with shared context. If you are a vendor, ask whether your customer's pain point is "who do I blame if this breaks" rather than price; the intermediary offering contractual support and 20% token efficiency wins [00:30:05; 00:31:09].

Build Contract Optionality Before Committing to Capacity

The SpaceX AI deal's 90-day exit and the Nscale pay-as-used structure show how to take large capacity while preserving the right to walk. Li reads the S-1 for "更多的让他退出的一个机制" ("more mechanisms letting him exit") [00:33:55]. Operators negotiating long-term supply should price in exit rights, not just unit price, since duration mismatch against short-term revenue is the main solvency risk.

Cut Low-Quality Revenue Early, and Disclose It

Anthropic deliberately removed distillation-driven usage, estimated at 5–7% of revenue, accepting a short-term growth hit for long-term defensibility [00:15:11]. Li calls it "一个长期的考虑，这是他应该做的一个事情" ("a long-term consideration, something he should do") [00:16:55]. The tactic: remove revenue that undermines your moat even if it flatters growth, and be ready for seasonality distortion (e.g., trailing-28-day ARR at year-end holiday weeks) [00:15:34].

Instrument Your Customers' ROI Before They Can

Even Fortune 500 CIOs can't quantify token ROI ("我们自己都不知道...不知道如何去测量") [00:39:31]. Operators selling AI tools should build ROI measurement into the product (hours saved, headcount, revenue), because buyers' inability to measure is currently the binding constraint on spend expansion.

6. Overlooked Insights

Anthropic's Hard Compute Obligation Is Dominated by a Single Supplier with Collateral Rights

The $161.2B Broadcom contract is non-cancellable, requires Anthropic guarantees, and equipment collateral, while the larger-looking SpaceX AI and Nscale commitments are escapable. Manqi: "Anthropic在这个合同里他就处于看起来是比较弱势的状态" ("Anthropic in this contract appears to be in a weak position") [00:36:07]. Li's interpretation, that Anthropic accepts this because Broadcom is the only provider that can reliably deliver TPUs and TSMC wafers, implies a hidden dependency: Anthropic's real survival risk is concentrated in Google's TPU supply chain via one counterparty, and the $600B+ headline obscures that. This also hints at the strategic reason the TPU/Broadcom ecosystem (and contract manufacturers such as Semiverse) are the most "real" part of the AI compute stack.

Anthropic Books Revenue Gross, While OpenAI Books Net: Reported Growth Isn't Comparable

Briefly mentioned in the channel-fee discussion: Anthropic includes the ~37% paid to cloud partners inside revenue, whereas OpenAI's executive said revenue and ARR are reported net of cloud payments [00:28:10]. Li confirms each company chooses its own convention [00:28:36]. Any head-to-head comparison of ARR, growth, or margins between the two labs, and any valuation multiple built on those figures, is distorted by up to roughly a third in the gross-vs-net gap. It also means Anthropic's "cost of revenue" and compute-to-revenue ratio (48%) are on a different base than a net-reporting peer's, which is exactly the kind of disclosure difference that will matter once both are public.