返回资讯中心
外部精选
软件工程
#开发者实时流

The Anthropic Feedback Loop: Cost of Capital Diversions and Delusions

I knew that my post on Anthropic would draw strong reactions, when I made it public last week. I did mention in the course of the post that I have no interest in selling my estimate of value to the rest of the world, bo…

Hacker News 最新toomuchtodo2026/11/10 · 7 天前30 分钟阅读

以下正文同步自 Hacker News 最新,版权归原站所有,已转换为易读排版。

Aswath Damodaran

Nov 03, 2026

200

34

26

Share

I knew that my post on Anthropic would draw strong reactions, when I made it public last week. I did mention in the course of the post that I have no interest in selling my estimate of value to the rest of the world, both because I don’t own the high ground here and because it is meant to guide my decisions, not yours. That is why I made my model open to changes that you can make to my story and inputs, and while you have to trust my mechanics, I tried to carry your inputs into value. Much of the pushback I have received has been thoughtful and productive, some of it has been emotional (either from people who love Anthropic or from those who would not touch it with a ten-foot pole) and some of it, especially on the cost of capital, reflects a fundamental disconnect with market history, current market pricing and just basic business sense. I plan to play devil’s advocate with my own valuation, looking at my story inputs, and ranking them in order of fragility, partly to stress-test my value estimate and partly to reinforce some corporate finance first principles, and then move on to your feedback and what I take away from it.

Playing Devil’s Advocate: A self-critique of my Anthropic valuation

Since I posted my valuation only a week ago, rather than rehash how I approached the valuation and my end results, you can read the gory details in the post that accompanied it. I will start with a summary of the valuation, though, and then list the parts of my story that I believe are most directly driving my value estimate on both magnitude (how much it is affecting value) and fragility (how uncertain I feel about the part).

I did summarize my valuation of Anthropic in a figure in my original post, but since there was that much to compress into that picture, I will try a different format for the same here:

If you check against my estimates from my earlier post, you will notice that the equity value at $1.45 trillion is slightly lower than my prior estimate of $1.475 trillion, as I adjusted the trailing 12-month revenue and operating income numbers to include the just-leaked third quarter numbers. Note that replacing the third quarter revenues from 2025 with the third quarter revenues of 2026 raises the base revenues, and since I still believe that the revenues for the coming twelve months will stay at my original estimate of about $100 billion, the growth rate in revenues in year 1 has been adjusted down to deliver that result. This reframing allows me to add a few details that I might have glossed over in my last post, as well as correct errors (none of which thankfully changed by value estimate:

Value per share: Though I have a share count, created by adding the shares reported as outstanding in the balance sheet to the restricted shares granted to employees, that share count will change before the public offering, as bankers adjust it to deliver on a per share value that they believe will make the company more “sellable”. I will revisit this number, and update it when the company does (and it has to, before setting an offering price), but it should not change the aggregate value of equity.

Cost of capital choices: I used a 9.74% cost of capital which incorporates risky technology cost of equity and a cost of debt based on a BB+ rating for the company, putting the company at close to the 90th percentile of US companies in 2026. I know that some of you felt this was too low, and I will come back to address what a cost of capital can and cannot do for you in valuation, later in this post, and why I don’t think that this is the input that most moves the valuation.

Compute commitments: I wish that Anthropic had created a table in their prospectus, where they specifically listed total commitments, both contractual and total, by year, similar to the table that they have for purchase commitments. Instead, they have sprayed content commitment numbers in the text, with little detail on timing and contractual standing, which results in differences of opinion in what those commitments are., with estimates ranging from $400 billion to $800 billion, with no clarity on timing or contractual terms. In my valuation, my estimate of contractual commitments results in an aggregated value of just over $450 billion and a present value of about $373 billion. It is possible that these numbers will need revisiting as details emerge of these commitments, but this will remain a long term issue for anyone valuing or pricing the company. To offer an imperfect analogy, I struggle with dealing with content commitments at Netflix, whenever I value the company, and especially so in the early years, when the entire Netflix library was leased or rented from content providers.

Dilution Risk: Anthropic, like many young companies, has used its equity as currency, paying employees with restricted stock and options, and used it as an add-on to borrow money in the form of convertible bonds. In addition, if the company needs to raise capital in future years, it will have to issue additional shares for that purpose, and the larger its ambitions, the greater these issuances will be. When valuing Anthropic, there is no escaping this dilution overhang, and here is how it is incorporated into the value per share. With stock-based compensation, I valued the equity options (at their option value, rather than exercise value) and netted that number ($472 million) from my estimated equity value to get to value of equity in common stock; I also included the restricted shares as part of my share count to compute value per share. With the convertible debt, I have assumed that the marked-to-market number (it was the marking to market that created the $72 billion hit to earnings in 2026) is close to fair value, and subtracting it out will remove the conversion option from the equity. The future share issuances are to fund negative free cashflows, and since those negative cashflows are already weighing down my present value, there is no need to add the shares issued to the share count; adding it will be double counting.

If you are surprised that adding the third quarter numbers did not move my value by much, it is because it was an open secret that Anthropic’s third quarter revenue growth had slowed and I had already incorporated that into my valuation story.

The equity value that I obtain for Anthropic is close to the lower end of the offering price estimates, but as I noted at the time that I posted it, and will reinforce again now, it is a function of not just my upbeat story about the company, but also one about how the AI products and services market will evolve over time. To the question that some posed about whether I was concerned that I could be wrong, of course I am, but as I will argue in the next section, I cannot let that concern paralyze me. I would, of course, rather be right than wrong, but I am trying to forecast the future, and it is a guarantee that I will be wrong. That said, I am more uncertain and exposed on some parts of my story line than others, and here is how I would rank them for Anthropic, from most critical to less so:

Target Revenues: If I were forced to pick one variable around which my valuation revolves the most, it is my estimate of the target revenues for Anthropic, which I estimate to reach a trillion dollars in ten years. Not only would this put them at or close to the top of all companies, in terms of total annual revenues, but they will also be sharing that grouping with companies like Amazon and Walmart that deliver on scale, but at the expense of margins. The revenue assumption is also inextricably linked to not only what you think AI agents and products can do, but also how demand for these will be monetized. In addition, there is a company-specific component to this story, where Anthropic emerges as a winner in the premium portion of the AI market. capturing a dominant market share, with pricing power.

The fragility of the target revenue assumption comes from it being susceptible to any of the parts failing, and if a part fails, revenues will end up below a trillion, and perhaps significantly so. That, in turn, will have ripple effects on the other parts of your story and inputs, with lower revenue growth pushing down margins, creating idle compute power and increasing the risk of failure.

Governance: I did use harsh words to describe the governance structure at Anthropic in my last post, describing the combination of its structure as a public benefit corporation and a voting share concoction that leaves founders in control, no matter how much capital the company raises in equity markets, as a red flag. I did get some pushback from some arguing that this is a good feature, not a bad one, since these are smart people who understand AI better than anyone else in the world. Even if you accept these views I would still argue, perhaps even more forcefully, that concentrating power in the hands of founders is a bad idea, and here is why. Anthropic has a plausible or even probable pathways to business riches, but those pathways come with setbacks (technological, regulatory, competitive), and you need management that can respond quickly and decisively, sometimes abandoning existing policies. For the latter to happen, you often need outsiders, not as immersed or even as informed on the technology front, but with different backgrounds and experiences, to call attention to management missteps and push for change. I don’t know any of the founders of Anthropic, and this may be unfair, but they seem to share a common worldview and operate in an echo chamber. My concern, as an investor, would be that they will not see changes occurring in business economics, and even if they do, be unable to or unwilling to adapt to a changing environment, and for Anthropic, that may be fatal.

Profitability: My story for Anthropic is that it will not only be able to deliver operating profits soon, but do so on a sustained basis, with a target operating margin of 32.5% (before taking the regulatory risk haircut). Implicit in my profit story are assumptions about pricing power, the unit economics of the AI business and economies of scale and each is breakable:

One of the most frequent critiques I received in responses to my Anthropic valuation is that this is a business without moats, and while I agree that moats matter, I am not sure that I go along with the notion that this business will get commoditized. The truth in technology is that unlike consumer products or old economy companies, where the moats take the form of brand names, patents or cost advantages, the companies that win find ways to make their products stickier to their customers, where stickiness implies that customers will not switch to alternatives, even if these alternatives are technically superior and cheaper. I recognize that Anthropic has work to do in creating stickiness, but I can visualize ways in which Claude agents, twinned with client data on customers and employees, can become sticky, allowing the company to preserve its pricing power and market share.

Continuing investment needs: While much of the focus on Anthropic has been on their compute commitments, disclosed (loosely) in their prospectus, my valuation of Anthropic is driven by expectations of what those commitments will look like over time, and especially if my growth story for the company comes into fruition. Delivering a trillion dollars of revenues will require Anthropic to not just maintain and replenish the compute power that is has already locked down, but add to that power over time. Implicit in my story line for the company, where the company is able to add this compute power at lower costs (reflecting declining chip prices) and in time, is the assumption that the compute power story moves along smoothly, working through resource constraints and pushback:

If you have been tracking news stories about data centers,

正文由 探How 从来源站点 RSS 同步,内容未经改写;遇到排版缺失或需要图片、视频时请以原文为准。

打开原文
去社区讨论 →