Topics Pioneers Insight Method Author
Back

185: Anthropic’s IPO Filing | A Conversation with Amont’s Li Zhaoyu on Pricing Logic and Potential Risks

·晚点聊 LateTalk

Rendering the long image…

Topics AI Venture Funding and Bubble Risks Battle for Frontier AI Moats AI CapEx Sustainability and Valuation Cracks AI Inference Cost Optimization Coding Agents and Future Software Engineering Enterprise AI ROI and Cost Optimization Trillion-Dollar AI Capex and Bubble Risks
Pioneers Rob Li(李肇宇)

Summary

  • Li Zhaoyu believes October ARR will have to re-convince the market if Anthropic is to pursue a $1.5T valuation. At the time of recording, media reports put Q3 revenue at $16.8B, with sequential growth slowing from 144% to 45%; extrapolating that trend implied year-end ARR of roughly $85B–$90B, below the earlier bullish expectation of $100B. If October data still looks weak on an annualized basis, he said, “even getting to $1.2T–$1.3T would be a question.” The first version of the S-1 had not disclosed this key figure.
  • Improving customer concentration is a clear positive, but 900% NRR has to be read against its base. The top 5 customers account for about one-fifth of revenue, with no single customer above 10%—healthier than Li had expected. But having more than half of the Fortune 500 as customers does not mean their spending has fully ramped. He stressed that expansion from existing customers is what validates ROI, and “if the base itself is particularly small, that number means nothing.”
  • A roughly 37% cloud-channel take buys customers, compute and financing capacity. Responding to Cheng Manqi’s comparison with the “Apple tax,” Li said the fee was “absolutely worth it”: Anthropic has a small sales team and needs cloud providers to scale quickly, while enterprise customers rely on existing service relationships and data protections. Revenue comparisons also require care: the program says Anthropic reports channel revenue on a gross basis, while OpenAI calculates revenue and ARR net of channel fees.
  • The risk in compute contracts depends on cancellation terms; the headline total of more than $600B is not the right measure on its own. Some contracts can be canceled with 90 days’ notice, while Broadcom’s roughly $161.2B contract is much more binding; Cheng estimates that highly binding commitments still total more than $300B. Customer revenue comes with no equivalent long-term commitment. Li summed up the mismatch with WeWork: “Your costs are long-term leases; your revenue is short-term rentals.”
  • Compute efficiency is improving, but adjusted profitability and positive free cash flow are not yet enough to show the model works. Compute spending has fallen from 250% of revenue to 48%; Li thinks 20%–25% would be healthier. He questioned how Broadcom-related lease cash outflows are presented, calling it “a little accounting trick.” Including spending such as training the next generation of models, he believes the company is still losing money; the key test is whether revenue per unit of compute can keep rising.
  • Growth beyond Coding points toward knowledge work, but there is no ready-made playbook for matching its adoption speed. Cheng asked whether other industries have the same combination of standardization, visible value and user acceptance as programming. Li said the key question is “how much is standardized and how much is non-standard.” Law firms’ adoption of Harvey offers a concrete example, but judgments in areas such as investing are more subjective, making data collection and rollout slower. The $40T in knowledge-worker wages also cannot be equated directly with addressable revenue.
  • Assessing upstream AI demand requires separating a slowdown at one model company from a slowdown in the whole market. Li compared it to Xpeng selling fewer cars without necessarily hurting CATL: customers shifting to other closed-source or open-source models does not mean total compute demand is falling; “the core of the core is on the demand side.” But enterprise CIOs have not even settled how to measure ROI. More useful near-term signals are contemporaneous usage observed by Snowflake, Databricks and cloud providers—not demand forecasts that have already been validated.
  • Li called a verdict that the bubble had burst “clearly an overreaction,” while explicitly refusing to guarantee that AI will succeed. Amazon’s confidence in 40% cloud-business growth in both 2027 and 2028 still has to be validated by end-customer spending; Li’s patience as an investor rests on not using leverage. If demand collapses, he expects compute and chip companies with weak moats and reliance on private credit to be hit hard. But the banking system’s limited exposure and the industry’s share of GDP mean he does not see a direct parallel with 2008.

Deep dive

1. Growth remains fast, but it has already fallen behind the market’s growth script

  • Cheng Manqi first clarified the timing: the episode was recorded last Thursday, before Anthropic updated its S-1 with Q3 revenue and long-term compute-spending figures. The discussion below uses the disclosures and media reports available at the time of recording; its estimates should not be treated as updated formal figures.
  • Media reports at the time put Q3 revenue at $16.8B, against $11.555B disclosed in the Q2 prospectus, with sequential growth slowing from 144% to 45%. Li Zhaoyu’s distinction was: “From the company’s own perspective, it’s OK; against market expectations, it’s far below.” The issue was that the market had expected year-end ARR to exceed $100B, while the contemporaneous run-rate implied only $85B–$90B.
  • Cheng suggested model-release timing as one explanation: OpenAI released GPT-6 Extra and subsequent models in early September, while Claude was not updated until the end of the quarter. Li added that Anthropic says roughly 5%–7% of revenue comes from distillation, so cleaning up those accounts would have a one-off impact on growth. He thinks the practice should be addressed over time, but short-term analysis cannot ignore revenue the company is actively cutting. He also noted that the company said distillation would not be a factor in October; meanwhile, using token consumption in the 28 days through December 31 to annualize could make the Christmas holiday period make growth look weaker than the underlying business.
  • The first S-1 did not include October ARR, which Li sees as its biggest omission. Cheng asked whether the company would have to disclose it ahead of an IPO expected in mid-November. He said, “No,” but if the company already has the number by the roadshow, it would naturally want to share it; withholding it would lead the market to infer that “your numbers aren’t good.”

2. A $1.5T valuation needs current data; the 2028 vision cannot set the price

  • Li believes October data would need to support $100B in ARR by the end of 2026 for the company to have more confidence in pursuing a $1.5T valuation or higher. If it cannot even reach $85B–$90B, the valuation could be far below $1.5T: “Even getting to $1.2T–$1.3T would be a question.”
  • Using $65B ARR at the end of July, Q3 revenue of $16.8B and the company’s ARR calculation, Cheng estimated that October ARR would need to reach $90.4B for year-end ARR to have a chance of hitting $100B, assuming the previous sequential growth rate held. She also cautioned that Thanksgiving and Christmas would affect usage; Li noted that annualizing consumption over the final 28 days of the year could understate normal operating conditions.
  • Cheng estimated that if actual 2026 revenue were around $57.5B, or close to $60B, the company would need a two-year CAGR above 85% to reach the $190B–$200B in 2028 revenue it had previously presented to investors. Li considers those long-range figures a vision and guidance, not a basis for valuation: “We definitely can’t use that as the basis for pricing it.”
  • If the company ultimately lists at around $1.2T, he expects sentiment to take an inevitable short-term hit: traders tend to equate one company missing expectations with the entire market being in trouble. He did not quantify the impact and does not trade short term; he suggested looking at the market’s reaction when the $85B–$90B year-end ARR expectation was previously reported.

3. A broader customer base is real progress, but expansion needs to be tested against a low base

  • After reading the S-1, the clearest positive change for Li was customer concentration: in the latest quarter through September, the top 5 customers accounted for about one-fifth of revenue, and no single customer contributed more than 10%. “The customer mix is relatively healthy”—something he had not known before.
  • Cheng asked whether more than 50% of the Fortune 500 becoming customers paying over $100,000 a year meant penetration was nearly complete. Li said what matters is whether customers spend $100,000 or $100M; growing from $100,000 to $10M is 100x growth in itself. There are also about 250 customers paying more than $10M a year, but it cannot be assumed they are the same customers as that half of the Fortune 500.
  • Li described the roughly 900% net revenue retention rate, or NRR, at the end of September as “same-store growth”: after existing customers add or remove modules and features, does total spending keep expanding? That shows whether the product is delivering real value. But the base has not been disclosed, and multiplying a small trial several times is not as strong a proof point as customers continuing to expand after a major investment.
  • Cheng cited the 120%–130% NRR common at strong SaaS companies and asked whether a fall to 200%–300% would count as a blowup. Li gave no absolute threshold, emphasizing instead the pace of decline: if NRR reached that range in just a year, the market would see what had looked like a vast new market quickly taking on the growth profile of a mature company, with a major hit to sentiment.

4. High channel fees buy the growth rate needed for an IPO

  • Cheng’s calculation of revenue and costs put Anthropic’s cloud-provider revenue share at roughly 37%, above the App Store’s 30%. Li’s response was that the company has a small sales team and no time to build one quickly; to reach the revenue scale needed to support a major IPO in the near term, “you can only rely on channels.”
  • Cloud providers also supply compute, customers and capital. Alongside direct investment, they can offer financing arrangements such as guarantees and leases, enabling Anthropic to procure compute faster. For Li, those capabilities together explain why paying a high commission is still worthwhile.
  • Cheng noted that Anthropic first includes the money paid to cloud providers in gross revenue; according to prior comments from OpenAI executives, OpenAI reports revenue and ARR net of channel fees. Asked whether IPOs require a common standard, Li said companies choose based on their own circumstances: “There isn’t one fixed way it has to be reported.”

5. Customer relationships still sit with cloud and data platforms as model companies push for direct access

  • On why Anthropic also has to bear the cost of cloud-provider receivables, Li said, “Cloud providers are, frankly, very powerful.” Customers such as oil companies and Walmart can buy tokens directly, but buying through a cloud provider or Snowflake costs about the same and preserves a service relationship built over years. Contracted teams can handle API configuration and other issues.
  • Data security is another reason to buy through a platform. Enterprises worry that model companies may train on their data; using Amazon provides an additional layer of protection and a clear party to hold accountable if something goes wrong. Li framed this in particular as protection against personal career risk for CIOs, not just a technical choice.
  • He used a specific example to explain the value of data platforms: Snowflake understands customers’ internal data better, so the same task that uses 100 tokens when connected directly to a model might take only 80 through the platform. “The customer relationship still belongs to the cloud,” or to platforms such as Snowflake and Databricks.
  • Cheng asked whether Anthropic and OpenAI’s joint ventures with private equity firms to provide FDEs were intended to change that division of labor. Li said they have to pursue direct customer relationships, but sales requires knocking on doors one by one, making growth linear and slow. In response to questions about whether a research culture is suited to enterprise sales, he pointed to Google Cloud, which still trails Microsoft and Amazon in sales after many years.

6. Compute commitments should be split by cancellation rights; Broadcom’s leverage comes from delivery certainty

  • On the more than $600B in long-term compute contracts tallied by the market, Li said the first question is whether the company “has to pay that money, or whether it can choose whether to pay.” Cheng estimated that highly binding contracts still total more than $300B; the headline total and the obligations that cannot be avoided are not the same number.
  • Cheng added that the disclosed cap on the SpaceX AI contract is $120.5B, with non-cancellable periods ending in January and March 2027, respectively; after that, either party can cancel on 90 days’ notice. Li’s example: if demand never materializes, the company can give 3 months’ notice that it no longer needs the capacity, leaving the provider to bear the subsequent costs.
  • Another supplier, Nscale, has a contract capped at $44.6B, which can go unfulfilled in some circumstances. Li stressed that with smaller suppliers, it is also important to distinguish between committed capacity and paying only for actual usage. If demand falls, paying only for what is consumed is very different from having to pay the full contract value.
  • Broadcom’s roughly $161.2B contract cannot be canceled by either side except in specified breach cases, and Anthropic provides guarantees and equipment collateral. Li called Broadcom “one of Google’s gloves”: it has been involved in TPU production for a long time and can secure TSM capacity and deliver actual hardware. Whether other suppliers can obtain memory and HBM is less certain, so they cannot demand commitments of the same strength.

7. Compute spending has fallen to 48% of revenue, but still has a way to go

  • Cheng noted that compute spending has fallen from 250% of revenue last year to 48%. Li thinks 20%–25% would be a healthier level; the trend is right, but “it still has a long way to go.”
  • He cares more about revenue per unit of compute: two years ago, the company spent $1 on compute and generated less than $1 in revenue; by the company’s calculation, it now generates more than $2. Reaching $3, $4 or even $5 in the future would show that the model is working sustainably. The key is not to count on supply costs plunging, but for customers to see ROI, spend more and keep expanding the customer base. He said memory prices have not been able to come down in recent years.

8. Enterprises have not settled how to measure ROI; data platforms can offer only contemporaneous signals

  • Cheng asked whether there was an earlier indicator than NRR, which only shows up after customers expand. Li acknowledged, “We haven’t seen a particularly good metric yet.” Asked how much they would spend over the next 2 years, enterprise CIOs often say, “We don’t know ourselves,” because usage is still low and they have not figured out the returns.
  • The challenge is not even knowing how to measure it: after spending $1M on tokens, companies have yet to work out how much labor or time they saved, or how much additional revenue they generated. Li thinks that even asking management or the CFO may not get an answer; unformed budget intentions cannot be treated as clear forecasts.
  • He sees Snowflake and Databricks as useful contemporaneous indicators. Most enterprises do not currently use tokens directly through Anthropic or OpenAI; they go through platforms such as Snowflake, then through a router to Amazon. Those platforms can see how many tokens their customers use each day in real time. Cloud-provider AI growth is also relevant: the program cited Google Cloud growth of 103%, still below market expectations of roughly 114%–115%, showing that high absolute growth and meeting expectations are different things.

9. Long compute contracts and short customer orders create a WeWork-style maturity mismatch

  • Cheng noted that compute contracts can run as long as 20 years and that companies may also bear increases in power and infrastructure costs, while major customers have no corresponding long-term purchasing commitments. Li agreed this is “a real risk,” drawing a comparison with WeWork: “The money you pay is tied up long term,” while “your costs are long-term leases; your revenue is short-term rentals” (你的成本是长租,你的收入是短租).
  • The risk plays out in a concrete way: short-term tenants stop renting and leave, but long-term liabilities remain, so incoming cash can fall far short of required payments. That is why Li sees the central question for all AI companies as whether demand can grow in a real, healthy and sustained way. Debating whether capacity expansion will create oversupply does not answer whether revenue can cover these commitments.

10. Growth beyond Coding has to reach knowledge work, but standardization determines the pace of adoption

  • Li sees replacing white-collar knowledge work as the next major opportunity. By his estimate, the US has about 1 million programmers and Coding penetration is already high, limiting the market’s capacity. Robotics points toward blue-collar work, but involves physical processes and will scale far more slowly than work that can be done online.
  • Cheng noted that the prospectus uses $40T in wages for knowledge workers globally to illustrate the opportunity, but the value of work that can be replaced cannot simply be equated with those wages. Li suggested watching recruiting efforts by data companies such as Mercor and Scale AI: financial, legal and medical professionals are invited to provide workflows so models can learn private-domain knowledge beyond the public internet.
  • Cheng’s question was about demand: programmers are relatively homogeneous, but knowledge work varies widely across fields, and even people in the same profession may work differently. Will users adopt it quickly? Li acknowledged that investing is especially non-standard: for the same question, Buffett may conclude A while another investor concludes B. Collecting enough data takes substantial time and labor.
  • Law is the more concrete example he offered: major law firms he works with are already using vertical AI such as Harvey, which still runs on frontier models. Junior lawyers’ document work is relatively standardized and predictable. But he agreed that conditions differ by industry, and “getting it to scale quickly” may be difficult.

11. Adjusted profitability has to be tested against cash; model companies also lack traditional software’s marginal costs

  • Cheng noted that the company reported positive adjusted operating profit in Q2, with the Q3 figure reported by the media also positive. On the quality of those earnings, Li said adjustments can be extensive, “because no one is using GAAP”; his approach is to look at operating and financing cash flows together and ultimately see how much cash comes in and goes out.
  • He called the accounting treatment of roughly $150B in Broadcom leases “a little accounting trick.” In his reading, the company classifies them as operating leases and puts the associated payments in financing cash flow rather than operating cash flow, making reported free cash flow positive; he believes including those payments would make the picture worse.
  • The company has not broken out the cost of training next-generation models within R&D in detail. Li believes it is still losing money if all spending is included. He does not expect profitability immediately at this stage, but wants to see whether cash can support the next phase. Cheng added that cash, cash equivalents and short-term investments exceed $130B, providing ample resources, while future compute commitments are also substantial.
  • He said there are no particularly good public-company comparables. Once Facebook is built, the incremental cost of serving new users is close to zero; for model companies, the cost of serving the 1,000th customer is fairly similar to that of the first. Their capital appetite is therefore inherently larger, making them closer to capital-intensive industries. Cloud has some similarities, but is not a perfect comparison.

12. Google’s decline and Anthropic’s slowdown both require separating company-specific issues from industry demand

  • Li viewed the rise in AI stocks after the October 26 release of the S-1 as “positive, but fairly muted”—a normal reaction. The disclosed information revealed no shocking hole and nothing especially impressive; the market is still waiting for October ARR, which short-term moves cannot replace.
  • Cheng noted that Google shares fell after its October 28 earnings, with free cash flow at negative $9.07B for a second consecutive quarter. Li said the more immediate issue was 12% advertising growth, below expectations of 14%–15%; the market is also concerned that AI could cannibalize advertising. The pressure on upstream names was tied to quarterly capex coming in below expectations, but Google said spending happens project by project and is “lumpy” across quarters, with no difference over the full year. It gave no specific figure for next year, saying only that most 2027 capex would go to AI and would be significantly higher than in 2026; the market was then expecting growth of about 50%.
  • To explain the relationship between model companies and upstream suppliers, Li used electric vehicles: if Xpeng expects to sell 100 cars but sells 90, and the missing 10 go to BYD, Li Auto and Nio, CATL may not be hurt; it would be different if consumers shifted back to gasoline cars as a whole. Likewise, Anthropic losing share to other models does not mean total token demand is falling. Infrastructure investors have to watch the overall market.

13. The slope of the revenue curve is still unknown; patience comes from not using leverage

  • Li thinks the spending curve is relatively easy to see; what remains unknown is how steep the revenue curve driven by demand will be. That determines whether the two lines meet in 2029, 2030 or another year, and how long capital markets need to support the business. “The core is to study the curve on the revenue side, the demand side,” rather than repeatedly calculating spending that is already relatively clear.
  • The program cited Amazon’s confidence in 40% growth in its cloud business in both 2027 and 2028. If delivered, Li thinks the curves can cross, but ultimately customers will decide. He called a verdict that the bubble had burst “clearly an overreaction,” while stressing there is no evidence that this confidence is 100% reliable; it is also too early to say the company is lying.
  • Asked why he is calm, he said he is used to investing with a horizon of more than 10 years and does not use leverage; otherwise, “even if your logic is ultimately right, you could die in the middle.” He visits industrial and consumer companies and tracks ROI and token spending, but does not think anyone has a clear answer yet: “We are investors, not forecasters” (我们做投资工作不是预测家). Whether his conviction ultimately matches reality remains to be seen.

14. If demand collapses, suppliers with weak moats and reliance on external financing will be hit first

  • On circular deals and borrowing, Li first distinguished businesses by their competitive moats. He described some new cloud providers’ bare-metal businesses as real estate, not technology businesses. By contrast, he prefers memory businesses that “only 3 companies in the world can do” and that have ample cash; he expects the supply-demand gap to widen next year and the year after.
  • If AI demand really collapses, suppliers such as Nscale will be hit hard. Some new chip companies cannot yet mass-produce, or are ramping very slowly, while relying on private credit to fund them. Under those conditions, Li was blunt: “They’ll die quickly.” He does not see the entire upstream sector as facing the same risk.
  • On comparisons with the 2008 subprime crisis, he gave 2 differences: by his estimate, AI accounts for more than 2% but less than 3% of US GDP, far below real estate’s share at the time; and financing comes mainly from private credit, PE and other private markets, with comparatively little bank involvement. So even if the industry runs into trouble, he expects a limited impact on the banking system—and if the banking system does not collapse, he thinks the economy is less likely to repeat 2008–2009.