I am a dabbler, by nature, moving on from topic to topic, but this year, I have spent more time than usual on AI, but that is because it seems to be part of almost every business or economy related story. I tried to value SpaceX ahead of and after the filing of the prospectus for its initial public offering before moving on to the question of whether it and OpenAI and Anthropic should be part of the S&P 500 index. I started my analysis of how AI would affect the rest of us, from doomsday to do-nothing scenarios, largely in response to Citrini and Citadel reports on the topic. in August 2026, I took my first big picture look at AI as a business, which I titled its bar mitzvah moment, with a look at its potential market, business economics and moats. That look was a prelude to the initial public offerings in the works (then) for Anthropic and OpenAI, and I provided a simple way to reverse engineer how much investors would need in revenues from these companies to justify their sky-high pricing. In early October, I looked at how the massive investments in AI are playing out in the aggregate earnings, cashflows and invested capital at US companies, and why it is contributing to holding the market intact, in the face of rising rates. Just last week, I argued that both Anthropic and OpenAI would be priced, rather than valued, for their initial public offerings, using what I called buzzy metrics (like annualized run rates) and promised that I would try to value the company, when it filed its prospectus. On October 26, Anthropic filed its prospectus, and at the risk of being called hopelessly wrong by both bulls and bears, I will try to value the company, using what I see as its most likely pathway, while also leaving the door open for you to disagree with me, and pick your own pathways to valuing the company.
The Prospectus
Before I delve into the details of the S-1 filed by Anthropic, it is worth noting that at 416 pages, it is long and filled with distractions, as has become the case with almost every initial public offering in the last two decades. (I chronicled this trend in my post on the disclosure dilemma, arguing that while filings have become longer, they have become less informative). The Anthropic IPO has useful information that both bulls and bears will spin to justify their priors, and in the section below, I look at key disclosures, broken down by valuation input into revenue growth, profitability, risk and reinvestment categories, from both an optimist (glass half full) and pessimist (glass half empty) perspective:
Revenue Growth and Composition: The good news in the revenue story is thatthe company has grown exponentially over the last year, with revenues in the first two quarters of this year about fourteen times as high as the revenues in the first two quarters of the last year, but the sobering news is that revenue growth slowed down in the third quarter of 2026, as OpenAI increased market share, compute capacity acted as a constraint and some enterprises moved to cheaper open source models. The good news on customer composition is that the number of enterprises that spend more than $1 million a year on Claude increased to 1900, from less than a hundred a year ago, but the bad news is that the company still has concentrated revenues, with its top five customers accounting for 25% of revenues.
Operating and Net Profitability: In bad news, Anthropic reported an operating loss, corrected for non-cash and one-time charges, of over $ 8 billion, for the first six months of the year. The actual loss reported was $80 billion, but $72 billion represented a non-cash charge, expensing the conversion option of the convertible bond. The good news, though, is that not only are the trend lines working in their favor, with the company claiming that it had an operating profit in the most recent quarter. The company is also delivering its revenues more efficiently over time, with all of its expenses decreasing as a percent of revenues across the quarters, with compute costs, its biggest cost,, decreasing from more than 250% of revenues to 48% of revenues, on a quarterly basis.
Reinvestment for growth: AI agents are computing power hogs, and AI companies invest immense amounts in compute power. The financial statements in Anthropic's prospectus actually do not show this investment at the moment, with neither the cap ex in the statement of cash flow nor the debt (and invested capital) in the balance sheet showing substantial investment in compute power. Before you accuse Anthropic of hiding the truth (or worse), the company is open about its future compute power commitments, reporting hundreds of billions in compute commitments in future years. The good news is that Anthropic is now catching up with OpenAI, which had the lead on compute power, but the bad news is that this catching up has not been cheap, in terms of cost commitments in future years. It i s also worth noting that compute power will need to be renewed and reinvested in, and while chip costs (a key component) may come down, data centers, facing opposition from communities, may become more expensive.
Operating and Regulatory risk: Anthropic's argument is that its customers, once they adopt Claude models within their enterprises, will find that they become so integral to operations that they will not abandon them (making revenues sticky and the business less risky). That has not been fully tested yet, and it is entirely possible that spending on AI agents is cyclical and that companies will pull back in bad times. On the financing front, committing to hundreds of billons in compute, purchase and lease commitments exposes the firm to default risk (with a bond rating of BB+), but in response the company will argue that its access to new capital, as a publicly traded large market-cap company, will make this risk recede. In addition, its cash balance, already at $114 billion, and likely to be bulked up with another $100 billion from the IPO offering, will give them more than enough cash to cover commitments in the next few years.
If I were to summarize the information that I obtained in the prospectus that is relevant to valuing the company, I would come up with the following (and neither Anthropic bulls nor bears will be happy with me):
The company's revenues are growing quickly, and for at least of the subset of users that are its core customers, it is delivering on its promise. That said, the revenue growth rate is clearly coming down, as scale, competition and compute power constraints all work against the company, and the fact that Anthropic is so interconnected in both operating and financing practices with other big players in the AI game (the circularity that is often noted by critics) does lead to the potential for double counting and opacity.
The expenses for Anthropic, across categories, are growing slower than revenues and that is helping reduce operating losses and putting the company on a glide path to not just profitability, but significant profitability. That said, notwithstanding one quarter's reported profits, the company is not quite there yet, and will require significant revenue growth to keep ahead of the expenses (from the compute power commitments) that will hit the income statement in future periods.
The immense compute power commitments at Anthropic have not only allowed the company to lock in what it needs to deliver growth but are also largely responsible for the high gross profit margins reported by the company, because they have moved company's biggest expense from the variable to the fixed operating cost column. That has come at a cost, since this not only increases the company's operating leverage, but also will result in further commitments down the road, as compute power needs to be replenished and grown.
Anthropic has built a buffer against risk in the form of a very large cash balance, which will be augmented after the IPO, and it is is possible that Anthropic's customers will become reliant on its products, making its revenue stickier. However, it is exposed to risk on three dimensions - in operations, from high operating leverage, in financing, because of its contractual content cost commitments and on survival, since it is dependent on growth to meet its obligations.
Over the next few days, don't be surprised to see people make cases for and against Anthropic using their spin of the data in the prospectus. Neither side will be convinced, or convincing, and will emerge with their priors about Anthropic intact and reinforced. I think that too much time and effort is being spent on going through this prospectus, as if it contains the answers to every question you may have on AI, when, in truth, it is a snapshot of a young company that is evolving. This prospectus, at least in my view, is analogous to reviewing a report card for a kindergartner (a precocious and talented one perhaps, but still a kindergartner) and trying to determine how much, if any, to set into that child's college fund.
My one piece of advice, no matter which side of the optimist/pessimist divide you may fall on, is that you do not get sucked into conspiracy stories revolving around AI company cabals (where the hyper-scalers, chip maker and LLMs are all working at creating a mirage of AI success to fool investors) or accounting fraud. I agree that self interest is driving many of the actions that you observe in this space and that accountants sometimes use the discretion in accounting rules to make the numbers look more favorable for companies than they really are, but those are common problems across all companies, not just AI and Anthropic, and falling into the conspiracy rabbit holes effectively blocks you from even considering alternative explanations.
The Anthropic Valuation - Story and Numbers
The prospectus, with all of its disclosures, is still a document rooted in the past, and the value of a business lies in your expectations for the future. In this section, I am planning to leave the safety of the historical disclosures and embark on forecasting the future, knowing fully well that I am going to be wrong, and perhaps even massively wrong, in my attempts. At each step of the story, I will offer you a chance to disagree with my take, and provide your own, with the resulting change in inputs. I know that you are busy, but if you can humor me and play alone, we may both benefit from this exercise.
1. The Scaling Up Story
The value of Anthropic, and any of the other AI product and service companies that will come later, ultimately rests on how big you believe the AI products and services market will become, and what portion of that market you company will inhabit. I started on this discussion a couple of months ago in my assessment of AI as a business, and argued that the total market for AI products and services will depend on four assessments -
whether AI is a productivitytool for employees or employee replacement, with the latter creating a larger market for its products and services.
whether that employee displacement will be primarily at the top end (top quintile) of the wage scale or across the board, with the latter pushing up the total market.
whether AI's impact will be concentrated in a few industry groups or across many or most industries, with the market expanding as AI's reach expands
whether the AI disruption will be primarily in the United States and a few other geographies or global, with a global reach leading to a bigger market.
I used the picture below to illustrate the range of total market outcomes:
I believe that AI will have the power to disrupt some industries, and that its most powerful (and highest priced) models will displace white collar (and highly paid) employees, yielding an end market size of $5 trillion or greater for its products. In my view, and that is based on a very early reading of the AI business today, Anthropic is aiming for the premium portion of the market and looks like a leader in that portion of the market. The one emerging add on is Anthropic's ambition (made explicit in the prospectus) to be a competitor in a subset of the businesses that its agents are disrupting, with the pharmaceutical business emerging as a potential play. In effect, Anthropic Pharma, or whatever they choose to call it, will then use its agents to do drug research and testing, displacing the companies that currently provide this service, and reap at least a portion of the revenues from the drugs they develop. With that story in place, I believe that it is plausible that Anthropic's revenues could approach a trillion dollars in a decade. For perspective, there is only no company in the world with more than a trillion in revenues today, with Amazon and Walmart sitting at around $700-$750 billion (albeit in current dollars and time).
Clearly, these are my views, and yours may be very different. If you are willing to play along, I would like you to make your own assessment and make your choice below:
My choice is shaded black, and as you make your choice, pay more attention to the target revenues that you believe Anthropic will have in 2036 than to the compounded annual growth rate between years 2 and 5, which is just an input to get to your target revenues.
2. The Profitability Story
There is clearly movement towards profitability in the prospectus, but where will the profit margins end up after Anthropic has gone through its growth pains? To estimate where margins will converge, it may help to start with the recognition that the AI product and service market, once established, is likely to have a mass market segment where AI agents and models, many of which may be open-sourced and less powerful, will compete for customers primarily on pricing, and a premium segment, where frontier models, many of which will remain closed models, powered by client data, dominate. The operating margins for the former are more likely to resemble those of business services (roughly 10%), whereas the operating margins for the latter will converge on the operating margins of enterprise software (close to 40% for the top quartile), with compute power and R&D becoming the reinvestment needed to deliver these products.
Anthropic is targeting the premium market, as reflected not only in its current revenue breakdown, with almost 80% from enterprises paying based on usage, but also in its focus on using its newly acquired compute power to build next generation models rather than service current clients. I will assume an end-mix of 75% of revenues from premium products, and 25% from the mass market, giving me an unadjusted pre-tax operating margin of 32.5%; that would put them at the ninth decile of US companies in 2026.
Here again, you may find room for disagreement, either on the market mix (mass market versus premium), the margins for each market segment or business economics, as companies scale up. You can make your choice below:
Again, my estimate of the target margin is highlighted in black, but with plenty of room for a different choice.
3. The Reinvestment Story
The reinvestment assumptions you make for Anthropic will revolve around understanding compute power, in terms of both how it is created as well as how much frontier models will need that power, as they become more powerful and scale up. For the moment, Anthropic, with its compute commitments in excess of $600 billion and greater, has sent signals that it views compute power as a necessity to power its growth ambitions as well as to maintain an edge over competitors (current and prospective).
The nature of compute power, and especially with the chips that represent a large percentage of costs, is that investments will age quickly and will need replacement. This is an indication that if Anthropic's growth ambitions come to fruition, there will be more reinvestment needed in future years, to both replenish and grow compute power. It is true that technology may make chips cheaper, but that may be more than offset by new models demanding proportionately more power, and if you add in the real estate and power needs of data centers, there is room for diverge on how much additional investment will be needed in future years. I do believe that the compute power crunch will stay in place for the near term, but will ease over time, and that belief translates into my assumption that Anthropic will need to invest about $1 for every $3 in additional revenues, going forward. That would be in line with what technology companies are generating as revenues per dollar of invested capital today, but with the leap in revenues that Anthropic is forecasting, it will translate into large dollar reinvestments each year.
As with revenue growth and profitability, I will leave the door open for disagreement, and you can make your choice below:
While I have highlighted my base case assumption in black, note that lower sales to capital ratios will make the business more capital-intensive, and increase the cash drain from reinvestment.
4. The Risk Story
As you start listing out the risks that you will be exposed to as an Anthropic investor, you may be overwhelmed both by the number of risks as well as their diversity. To gain some control in this process, I will divide the risks into three groups:
a. Operating risk: The AI business is still evolving, and while Anthropic's growth and customer stickiness so far are noteworthy, the business will face challenges as it grows, and lives through market and economic turmoil. Thus, for Anthropic's customers spending on AI as a productivity tool, we do not know whether they will cut back their spending, and if so how much, in bad times, and for those who plan to replace employees with AI agents, whether the pace and extent of usage will change with the economy. The number that is designed to convey this risk into the valuation is the cost of capital, and for Anthropic, assuming that it carries the risk of large technology companies with an added burden of the compute commitments, I will estimate a cost of capital of 9.74%. Again, for perspective, note that the median cost of capital for US firms in September 2026 was 9.16% and that the third quartile for US firms is 9.88%. If you disagree, here are your choices:
My advice, and feel free to ignore it, is that you should not spend too much time finessing the cost of capital, since of the many assumptions that drive this valuation, it is among the least critical. While I highlight my base case cost of capital, remember that when investing in a publicly traded company, what you would like to use as a cost of capital has to be constrained by what the market is offering.
b. Financing risk: The main talking point for many Anthropic bears is the compute commitment that runs into hundreds of billions, creating fixed costs that will be a drag on profitability, especially if revenue growth falls short of expectations, as well as the risk of distress. While bondholders are right to be trepidatious and will point to the current bond rating of BB+ for the company as backing for their concerns, there are two reasons why failure risk will be suppressed. The first is that the company has the cash balance to sustain itself through the near terms and meet its contractual obligations. The second is that, as a high-profile trillion-dollar plus publicly traded company, it will have access to capital. I do believe, though, that if revenues grow at anemic rates, both these benefits will dissipate, as cash gets burned through and the market cap collapses, and I will estimate a probability of failure of 10% for the company, with a loss of half its value if that happen. If you want to disagree, go into the input sheet worksheet of the spreadsheet, and if you scroll down to cells B52 and B53, you can put in your own views:
c. Regulatory risk: With AI companies, the question with regulatory risk is not whether it will show up, but when it does, what form it will take. Anthropic and OpenAI have both invited the government to restrict how this business will grow, and while the cynical rationale for these pleas is that it may keep out competition (especially from China), the regulations and restrictions are coming. As with the large social media companies, the effect of these regulations will not be a cessation of business but an increase in operating expenses and loweer profit margins. I estimate that the regulatory effects are likely to be mild, at least on a relative basis, and that it will shave my initial operating margin estimate of 32.5%, for Anthropic, down to 27.5% (a reduction of 5%). You can make your own call on the regulatory effects on margins below:
You may disagree with me both on the likelihood of regulation and its effects, and while my base case is highlighted in black, you can diverge in your choice.
The Valuation
With my storyline and inputs for Anthropic in place, I valued the firm, and you can check out the spreadsheet at this link. Rather than drown you in cash flow line items, I will present my Anthropic valuation, with a trillion in revenues by 2036, a target operating margin of 27.5% (with the target margin of 32.5% adjusted for regulatory drag), reinvestment that delivers three dollars of revenues for every dollar of capital invested over time and a cost of capital of 9.74%:
With my upbeat story for Anthropic, I estimate a value for the equity of $1.475 trillion, at the lower end of the offer pricing range ($1.5 to $2 trillion). Note that my debt for the company is close to half a billion and includes my estimate of the present value of their compute commitments, and that my cash balance includes an estimated $100 billion in proceeds from the offering. If you are bullish, you will undoubtedly find ways to tweak my numbers, and find that a slightly higher operating margin and a lower cost of capital or a longer growth period, will get you to whatever number you want. If you are bearish, you will do a similar exercise in reverse, lowering growth and raising discount rates, to capture your priors. It was precisely to avoid this gaming that I tried to get you to make your own choices on revenue growth, margins, reinvestment and risk along the way, and if you had the time and patience to play along, here is what the range of outcomes would look like.
The worst-case Anthropic story, where its target revenues limp (at least relative to expectations) to $200 billion in 2036, margins converge on the mass market (10%), compute power stays scarce and costly (sales to capital of one), cost of capital rises to the third quartile for US companies (9.88%) and the regulatory effect on margins is drastic yields worthless equity, In fact, the value of equity is -$373 billion, and the company will not survive as a going concern.
The best-case Anthropic story, where target revenues surge to $1.5 trillion in 2036, margins converge on the premium market (40%), compute power becomes cheap (sales to capital of four), cost of capital falls to the first quartile for US companies (7.79%) and the regulatory effect on margins is non-existent yields a value of $4.04 trillion for equity.
While I believe that the best case and worse case scenarios are both possible, but unlikely, you can assess your own story for the company, with the resulting value, applying what I call the 3P test: Is it possible? Is it plausible? Is it probable? Investing is a personal decision, and whether you choose to buy Anthropic, sell it short or sit on the sidelines, you should base it on your own assessment. In fact, just to get a sense of what the crowd judgment on the Anthropic IPO looks like right now, I have created a shared spreadsheet on Google, where you can enter your numbers and see how they compare to others valuing the company. I have entered my name, because I don't see a point in being anonymous, but if you want to use a pseudonym for anonymity, I don't have a problem with that.
Corporate Governance and Lock-up Provisions
With publicly traded companies, corporate governance obviously matters, but in the last few decades, that recognition unfortunately has led us away from a sensible rationale of why it matters. In fact, much of academic research on corporate governance as well as practitioner measures of it have become focused on the minutia and details of governance, with number of board members, length of their tenure and board composition (insiders vs outsiders) all vying for front place. The truth is that corporate governance scores, which claim to measure how strong or weak governance at a company is, have been gamed by companies, making them poor indicators of shareholder power.
So, what exactly is corporate governance? In my view, it is a measure of the power that shareholder (in public) and outside investors (in private businesses) have to change the way businesses are run, if they disagree with the way they are run by incumbent. It is therefore true that this power will be greater at companies where boards are more willing to push back against CEOs and activist investors are involved, and will diminish when you change the rules of engagement to make it more difficult for shareholders to push for change. I have long argued that the institutional investors who sat on their hands, when Google decided to break a precedent, at least in US markets, at its initial public offering by issuing shares with different voting rights. Their success has opened the door for a host of tech companies that have used the same structure to protect founder power at companies and effectively reduced the power that shareholders at these companies have to push for change to selling their shares and moving on.
With Anthropic, these concerns about entrenching founders at companies is magnified by two choices that the company has made:
Voting rights: This company has decided to exhaust the alphabet in terms of share classes, with five classes of shares, with different voting rights.
Note that the class A shares come with standard voting rights, but the voting structure, with a benefit trust picking the board and the founders controlling 50.1% of voting, no matter how many more shares may be issued in the futur, effectively emasculates public investors.
Public benefit corporation: The company has structured itself as a public benefit corporation, and in its prospectus, it states that "we may make decisions that we believe appropriately balance the pecuniary interests of our stockholders, the best interests of those materially affected by our conduct, and the Company Purpose, but that negatively impact the short-, medium-, or long-term performance of our business or the market price". Given the worries that people have about AI being a force for destruction, a concern that Dario Amodei has actively cultivated, this protection against shareholder power may seem to be a good thing, but we need to be open eyed about what this structure does. It does not protect society from AI's evils, but puts power in the hands of the few (the founders and members of the benefit trust that they pic) at the top of Anthropic, hoping that they use that power well. If you, as an investor, have had bad experiences with top management at companies using corporate responsibility and ESG as cover for self-serving activity and hypocrisy, the Anthropic structure is that structure on steroids.
As an investor who holds shares in other companies that have dual classes, with different voting rights, including Alphabet and Meta, I will not draw a line in the sand and assert that I will not invest in companies that have "fixed" the voting process in favor of founders. That said, I think Anthropic has managed to put those companies to shame with its combination of a benefit corporation structure (which gives the company the legal license to do pretty much whatever it wants with your money) and a voting share design that is so deviously self-serving that it would put Kim Jong Un to shame.
In its nineteen IPO rounds, Anthropic has already accumulated an investor base that includes some of the biggest names in venture capital, as well as other companies in the AI game. I know that for some looking at the Anthropic IPO, there is the concern that these investors will cash out at the expense of public market investors, and that concern will probably be magnified by the fact that the company will, unlike in many other IPOs, allow some investors to cash out at the offering, with more shares being unlocked over time as the company trades. On the first trading day, insiders (but not founders and certain major figures) will be permitted to sell up to 10% of their shares. The remaining shares will be unlocked in tranches, across the first 180 days of trading, and founders and key named figures will be allowed to trade after 365 days. To be honest, I think that too much is made of share unlocking, perhaps because it gives investors and journalists something to talk about before unlock dates. With companies like Anthropic, where you have large institutional investors in the game, and plenty of liquidity, the effects of these unlocks is likely to be muted. My view, though, also reflects a belief that the smart money (which presumably includes all of those selling at the unlock) is not that smart and that their selling has far less informational value than you might believe.
As you look at both the corporate governance and lock-up provisions that Anthropic has put into its prospectus, it seems to me that they are daring institutional investors to make a choice of speaking up against these provisions (by not partaking in the IPO) or staying passive, and have already calculated that they will do the latter.
Trading the IPO
I would draw your attention to my last post on Anthropic, where I argued that the company will be priced, and that almost everyone who trades the stock right after the offering will be pricing, not valuing the stock. That is not meant as a slight, but as a reality check for investors, and especially those investors who seem to believe that the market owes them a payoff, just because they care about fundamentals. Markets owe nothing to any of their participants, and as an investor, I have to not only recognize that reality, but act accordingly. Consequently, I will keep an open mind and wait for Anthropic to be priced for its IPO, and if it is priced well above my estimate of $1.48 trillion, I will sit out the frenzy. To the question of why I would not sell short, if it is priced well above my value, my response is that while I have faith in my valuation, albeit with a range around my estimate, I have no reason to believe that the market will or should agree with me, and even if it has made an error, correct within my time horizon.
Once Anthropic starts trading, its story will continue to evolve and change, and given AI's presence in the economy and markets, I can guarantee you that it will be a roller coaster. It is possible that we have underestimated the possible payoff from AI, and that Anthropic continues to rise in market cap, but it is just as possible that major parts of the storyline, from revenue growth to margins to risk, will come under stress, leading the market to reassess what it thinks Anthropic should be priced at. Obviously, those stresses will also cause my estimate of the value to change, but it is not just a possibility but almost a guarantee that there will be a day that Anthropic will be undervalued, and it may be sooner than you think. I valued Facebook at the time of its IPO in June 2012, and while I found it over valued then, at a stock price of $36, it became a bargain just three months later, when the price hit $18, making me a buyer of the stock. The structure of Anthropic, with the public benefit facade and the voting rights abuse, will require a much larger discount on my value estimate, but the market may accommodate that requirement as well.
Foreword: I wrote and posted this piece on August 20, 2026, with the promise that I would listen to feedback and try to respond and incorporate ideas and suggestions. To keep this promise, I will edit this piece over time and reflect the changes I make.
Original post: August 20, 2026
First update: August 21, 2026 - Updated AI TAM table to make explicit implications of AI market size for other businesses, employees and the economy
Second update: August 21, 2026 - Updated to incorporate impact of new businesses that may be born out of AI
In a world where AI enters almost every conversation, it takes effort to remember that its breakout moment was less than four years ago, when, on November 30, 2022, ChatGPT was unveiled to the public. I know! I know! Artificial intelligence has been around a lot longer, with a history tracing back to the birth of the computer age. I am old enough to remember IBM's Deep Blue, a machine powerful enough to evaluate two hundred million chess positions per second, and beat the greatest chess players of its time. On the cultural front, we have seen variants of stories, where machines break free from human control and take over the world, in novels and movies, with Hal (the computer) in 2001: A Space Odyssey retorting "I'm afraid I can't do that" to Dave, his human controller, remaining one of my favorite movie lines of all time.
Notwithstanding its longer history, the effect on AI has been explosive in the last four years, manifesting in multiple developments. The most successful company during this period, in terms of increasing market capitalization, has been Nvidia, the chip maker for the AI revolution. After spending a decade talking about FANGAM, the big tech companies that had become part of our daily lives while carrying equity markets forward over the last decade, it was the Mag Seven that became the stand-in for market dominance, with Tesla and Nvidia replacing Netflix in the mix. The Mag Seven, almost all of which have a stake in AI, have accounted for 45% of the increase in market cap across all US stocks between 2022 and 2025, and have an aggregate market cap on August 16, 2026, of $23.7 trillion. It is not just markets that are besotted with AI, since the massive investments in AI architecture, from data centers to large language models (LLMs) have carried the US economy; it is estimated that these investments accounted for about 1% of the 2.5% in real GDP growth in 2024 and 2025. Almost every conversation of businesses now has an AI component, which if not restrained, can hijack the discussion.
AI's effects were not restricted to business and markets, as people were exposed to its reach in their personal and work lives, with reactions ranging from awe, at its power to do tasks that used to require skilled human labor, not just effectively, but in a fraction of the time, to dread, at the possibility of being made obsolete by an AI agent. As the arc of the AI story has unfolded over the last four years, it seems to me that is has also transitioned in the public consciousness from a mostly positive phenomenon early on to acquiring a negative tinge, perhaps because of concerns that the genie is out of the bottle, and is not benign, and partly because some of its leading spokespeople are so unlikeable. Not surprisingly, speakers at graduation ceremonies in US colleges in 2026, were booed by students, when their speeches centered around AI.
The AI Debate: Off the tracks?
As the AI story has unfolded, there have been reams written about it, for and against, and almost as much said about it on television and podcasts. In spite of being so much in the news, the debate about AI, in my view, has gone off the track with advocates and skeptics often talking past each other, with advocates focusing on its "huge" potential market, and skeptics zeroing in on massive upfront investments as "too large", with each side claiming the high ground and labeling the other side as cultists (AI advocates) or Luddites (AI skeptics). There is a great deal of cherry picking of the data on both sides, with the optimists focusing on usage statistics (level and growth) and the pessimists on capital spending and current profitability (or lack of it).
This post is not about proving one side right and the other wrong, but about closing the loop and making it a discussion of AI as a business, recognizing that it is ground breaking, while also acknowledging that it has to be judged like every other business in history, not on potential usage, but on the prosaic details of converting potential to products and revenues, being able to deliver these products at a cost that generates profits and building moats to keep new entrants and competitors out. In short, the AI optimists may be right about AI usage exploding in the future, but big markets don't always become big businesses, and the skeptics have to concede that spending a lot on capital expenditures raises the ante for businesses, but don't necessarily doom them to value destruction.
I will be writing this as an AI novice, a very light user of ChatGPT (I still have only the free version) and acquainted with Claude only in passing (though my content has found its way into some of its bots). If you are an AI expert and feel that I am missing or wrong about a technical component, forgive my ignorance, and feel free to educate me, and if you work at an AI business and feel that I am in error on a business detail, the same offer stands. To be honest, I am writing this post for an audience of one (me), with the purpose of clarifying for myself how to make sense of this space, and if it does help you make sense of this disruption, it is a side benefit.
The Cycle of Revolutionary Change
Through human existence, revolutionary change has been a constant, and even when that change has been an advance for humanity, it has always come with pain for those that the change renders obsolete and unanticipated side costs. At the risk of overreach, I will argue that every major disruptive change has gone through four phases: a period of hope and hype, where the change is viewed as big, but it is unclear how and in what form it will be delivered, a period of build-up, where a subset of people (with more belief in the change and more willingness to take risks) start investing and building products to make the change happen, a period of business building, where the change is monetized and businesses form, and a recalibration, where the change works its way through the economy and society, in both good ways (increased productivity and welfare, new businesses) and bad ways (displacement and damage).
1. Hope and Hype
In this phase, the true believers and visionaries that see change coming start the ball rolling, but to succeed, they need to sell it to the broader public. Since the story of change, at least at this stage, has nothing tangible at its core (no products or services, let along revenues or profits), it is inevitable that there will be false starts mixed in, as well a dose of scams pushed by charlatans and pretenders. It is also par for the course that there will be many who will dismiss change talk as fairy tales, without even listening to the arguments, either out of cynicism or because they do not understand what is being sold. For change to take root, the visionaries selling the story need to be persuasive enough to get people to buy into their vision, both to get foot soldiers who will work to make change happen and investors to supply them with capital.
2. The Investing Build-up
Once the belief that change is possible gets a foothold, there will be a subset of players, with start-ups or in existing businesses, that will invest and build products and services that they believe will be sought after, if change comes. As pioneers in this space, with trial and error and experimentation characterizing these attempts, but even failures will lead to learning, albeit with costs.
If the change is perceived as revolutionary, with a big market emerging, this is the phase where the big market delusion, a term I coined over a decade ago, is likely to emerge. That delusion has its roots in selection bias, where the people building products for the change to come tend not only to be true believers but also over confident, resulting in a collective over reach by companies and investors pricing these companies, and a correction.
Thus, bubbles are a feature, not a bug, when revolutionary change is a possibility. Finally, if business and investing is a combination of (business) stories with numbers, at this stage of the cycle, where there is little material that has already been accomplished, it is the story that drives growth and investment. Investors with actuarial or accounting mindsets will undoubtedly find these narratives unpersuasive and quickly consign these companies to the overvalued heap. While that impulse is entirely understandable, it is worth remembering that there will be other investors, who are willing to bet on optionality, where they invest in this space, hoping that the entities that they invest in will be the big winners (though they have not won anything yet) in a big market (which does not exist right now).
3. Business Building
Not all change is revolutionary, not all revolutionary change translates into big markets, and not all big markets create valuable businesses, and it is in the phase of business building that the truth starts to emerge. Since this is very much a test of businesses growing up, it represents a bar mitzvah moment for these businesses, in the sense that investors are no longer willing to just price on promise, and start demanding tangible evidence of progress. It is during the business building phase that you start to create the structure of converting products into businesses, with production processes, supply chains, marketing and distribution all taking form. In the process of building business, you will confront the realities that will determine whether you are a mass market or niche company, including unit economics and economies of scale. In the process, they will also discover a harsh truth, which is that many creative and talented product-builders lacking business-building capabilities, and either have to partner with someone who does, sell their products to established companies that already have systems in place (expect acquisitions, partnerships) or get pushed out of their own firms by their capital providers (venture capitalists).
4. Recalibration
As businesses start to succeed, the laws of economics, immutable and powerful, kick in. You should expect to see turnover and consolidation, as new entrants and existing players jockey for position, and business economics determine industry structure from splintered to consolidated to winner-take-all all possibilities. At the same time, the dark side also plays out as those (businesses and individuals) rendered obsolete by the change come under pressure, with some shrinking, some disappearing and some in denial.
It is worth recognizing that there is no steady state, because as businesses recalibrate to the innovation, they now represent the status quo and become targets for the next revolutionary change, Schumpeter's creative destruction in motion. The picture below summarizes the cycle, mapping out the pathway from hype and hope to investing to harvest to building businesses to recalibration:
Where does AI fall in this cycle? ChatGPT, as I noted at the start of this post, may have been low-tech AI, but it got the hype cycle rolling, and social media amplified and accelerated that rollout, and its broad reach meant that almost everyone has seen it at work. The hope that AI's popularity and reach would create a big market built up in parallel, with capital flowing into firms in its sphere (as well as wannabes that latched on to it, as a buzzword), pushing the market capitalizations of the companies building AI's architecture (computer chips, LLMs, power and water companies, cloud) into the trillions of dollars, funded with equity and debt. It was not just financial market participants that saw its allure, as hyper scalers and new entrants invested hundreds of billions into AI cap ex, partly because they believed in its promise, but partly out of a fear of missing out. The graph below looks at cap ex in just six of the largest players in the space, four of them in the Mag Seven (Alphabet, Amazon, Met and Microsoft) and two outside (Oracle and Coreweave):
Source: Cap IQ
Cumulatively, the total investment from just these companies amount to $1.7 trillion, over the last few years, and their guidance suggests that they are not done, with trillions of dollars in AI cap ex commitments in the next three to four years. You can see why I use the analogy of a factory, and argue that AI has built the most expensive factory in history, and done so in hyper speed.
To what end? It is only in the last year or so that you are seeing the beginnings of business building, where companies are generating revenues from selling products made by the AI factory, with Anthropic and OpenAI as the most prominent examples. Those revenues are small for the moment, relative to capital invested, and the profitability is still a reach, but there is a host of experimentation going on on model type (open versus closed), business models (subscription versus usage) and pricing. The seeds of disruption have been sown, and there are signs that AI's rise will make a significant dent in the profitability of some businesses, with technology companies in the software and intermediary segments being the first casualties. The AI story is clearly further advanced than it was a year ago, but it is still early, and there will be changes and challenges that face both the players in the space and the investors in these players, making this AI's bar mitzvah moment.
A Business Framework for AI
If you are an onlooker or undecided on the AI question, I don't blame you, if you find yourself whipsawed by what seem like persuasive arguments on both sides and waylaid by distractions aplenty. That is because there are so many strands to this story that taking any strand in isolation can lead you to a conclusion about AI as a business that is hopelessly of course. The best way to bring all these strands together is by going back to basics, and establishing the drivers of the value of any business (not just AI):
In this structure, there are three broad drivers that will determine how AI as a business will unfold. The first is with an assessment of the size of the total market for AI products and services, the second is the industry economics in that market, which, in turn, will determine how many companies will cater to this market and the profitability of these companies, and the third will be an assessment, or at least a preliminary judgment, on what the moats or competitive advantages will be in this business.
1. Market Size
It is true that the value of a business is tied to how big a market there is for its products and services, but it is also true that this metric, converted into an acronym (TAM) has become a gaming tool in the hands of founders, venture capitalists and bankers. In my SpaceX valuation, where xAI is the primary AI business, I noted that bankers estimated a total addressable market (TAM) of $22 trillion for xAI, which I felt was more hallucination than estimate. That said, any discussion of AI as a business has to start with the total market question, and it is worth starting that discussion with an examination of where we are right now in terms of revenues from AI products and services.
As we head towards the end of the third quarter of 2026, with chatter about Anthropic and OpenAI getting louder, both companies are racing to set up their stories by reporting their updated annualized revenue run rates (ARR), an admittedly self-serving metric (for growth businesses) estimated by taking the most recent period (week, month etc.) and extrapolating to a year. On August 17, Anthropic that its ARR at the end of July 2026 was $65 billion. OpenAI's estimate of its ARR at the end of July was about $40 billion. While both numbers represented quantum leaps from their values just a year ago, adding these estimates to the revenues that SpaceX (from xAI), Microsoft (from its AI offerings) and other players, even with the most generous estimates, generate from selling AI products and services yields a total revenues that is modest:
As you look at these numbers, there are a few truths that are undeniable. The first is that the AI product and service market is not only fast growing, as evidenced by the ARR for the lead LLMs, but unpredictable, with Anthropic's most updated ARR coming in $10-$15 billion below estimates. The second is that even with the most upbeat and optimistic estimates of revenues for AI products and services, the current revenue number caps out at about $250 billion, and that sounds like a big number, until you scale it to the trillions invested in the space. Put simply, the big winners in terms of revenues and operating profits, at least so far in this AI cycle, have been the companies that supply the infrastructure components, with chips (Nvidia) electrical equipment providers and power utilities all sharing in the spoils.
That, of course, is just the existing market and with immense growth built into it, the question becomes about the end game, and that end market size, at least at the moment, seems to be anyone's guess. While the xAI bankers will undoubtedly use this uncertainty as a shield to not have to justify their estimate, there are ways we can start framing our choices, beginning with aggregate measures of what businesses spend as operating expenses, since AI's big sales pitch is that it will lower that spending. In 2025, the aggregate operating expenses at publicly traded companies was about $65 trillion, broken down by sector and geography below:
Source: S&P Cap IQ
These expenses include the costs of raw material and inputs that are immune from AI's efficiency push, since AI cannot replace the rubber you need to make tires, the wheat you need to process to get cereal and the chemicals that go into fertilizer. Consequently, it is the portion of these expenses that took the form of employee compensation, in all of its forms (wages, salaries, bonuses, stock-based compensation) that AI is targeting. While some companies break this portion of expense out explicitly in their financial statements, others do not, but there is macro data on this metric, albeit splintered geographically. Drawing on Federal Reserve data of compensation for all US employees, I get the following numbers:
Source: Federal Reserve
I would argue that AI's total addressable market, in the US, cannot be greater than $12.96 trillion, the total employee compensation in 2025, or an inflation-adjusted variant, if it is in the future, it is roughly twice that amount, if you target global spending on employees. While that number is large enough to set AI optimists' hearts aflutter, a world where every employee is replaced by an AI agent would not just be dystopian, but also an economic basket case. In fact, AI's target market will be smaller, depending on the answers to four questions:
1. Tool or employee replacement: As AI products have become more powerful, the debate about whether AI's future lies primarily as a tool or as replacement for human labor has also raged. In a post earlier this year, I focused on a Citrini report that played out the effects of the latter, highlighting the costs to the economy of laid-off white collar workers (and their income) and the effects on the market. In that post, I did note that notwithstanding public stories of layoffs in software companies, there has been little evidence (so far) of aggregate displacement of labor in any sector, at least so far. The takeaway, at least from this discussion, is that AI's disruptor role will be far greater, as will its total addressable market, if it replaces employees, rather than is used as a tool.
The logic for why AI tools than AI as employee replacement will have a smaller market is simple one, from a business perspectives. Businesses spend money on tools, but that spending will be in addition to what they already spend on employees, and while they rationalize that spending with (promised) improved productivity, it has to be a fraction of employee compensation. You can also why the current AI players (OpenAI, Anthropic) are opting for speedier disruption over a slower one, because it will then increase their odds of winning, albeit with higher displacement costs for society.
2. Pricing of AI products: In the last year. Anthropic and OpenAI have garnered publicity for their most powerful products (Claude, Codex etc.), and while some of them do offer the capabilities that will allow them to replace workers, they are expensive enough that it will make sense to use them only for high-paid labor. In 2023, the US government estimated, based upon personal income statistics, that the highest quintile accounted for 51% of all employee compensation.
Bringing this factor into play with the total employee compensation of $12.96 trillion in the United States, you can argue that only about half of that market (at the most) is open to disruption from AI replacement products,
3. Breadth of use: There are some industries where AI will make more inroads, and do so sooner, than others, and what separates them will be the nature of work in the business. As we noted just a little bit earlier, software and coding have been the easiest entry points for AI products, since the output tends to be more rule driven and easily verifiable An article in the Harvard Business Review, for instance, measured the risk of displacement across different occupations:
I would expect AI to be more successful, both as a tool and employee displacer, in settings where there is less client or personal interaction and more rule-driven than principle-driven jobs. If you look back at operating expenses, broken down by sector, the sectors most exposed to AI disruption (technology and financials) have aggregated operating expenses the amount to less than 20% of the global total, whereas sectors more immune (industrials, materials, real estate utilities) amount to a third of the global total.
4. Geography: If you consider the fact that AI is more likely to displace workers and generate revenues in non-manufacturing companies that have high priced labor, it follows that the disruptive effects of AI will be greatest in the United States and have a smaller footprint elsewhere in the world.
Looking back at the table that breaks down operating expenses geographically, for publicly traded firms, you can see that the US, Europe and China are the three biggest markets for AI, since these are regions of the world where companies spent most on operations in the aggregate
If you consider $26 trillion as your upper limit for AI's total market, in current dollars, your estimate of the size of the AI market will depend on your assessments of whether you fall on each of the four factors, with the largest assessments of TAM emerging from a view of AI as an employee replacement that cuts across industries and geographies, but with a cost for AI agents low enough to replace workers with lower income. At the other extreme, your assessment of the TAM will be much lower, if you view it as a tool, no matter how powerful, with application in select industries and geographies.
Updated on Aug 21, with implications added on
Rather than view different assessments of AI TAM as "he said, she said" disagreements, the debate would be much more grounded if these assessors were explicit about where they stand on the dimensions (AI as tool or employee replacement, target high-priced workers or all employees, useful in a subset of industries or all industries and primarily US-based or global) that drive their estimates.
In my first economics class, I became very fond of the words "ceteris paribus", latin for "all other things being held equal", partly because it does make it easier to focus on what's changing, but mostly because latin makes you sound smarter than you really are. Holding all else constant may be reasonable, when change is small, but when the change is revolutionary, as AI could very well be, it is no longer reasonable to make that assumption. Consequently, I think it behooves anyone making an assessment of AI's total addressable market to think through the implications for the rest of the economy and the world, because there may be constraints there that need to be considered. The most optimistic scenarios for AI investors and businesses, where the market for AI products is largest, are also scenarios where the damage done to other businesses is greatest and the impact on employment and the economy, in the near term, is most negative. Put simply, if the AI market explodes in size and does so quickly, and millions of high-paid workers lose their jobs, the economic damage will be deep, and there may very well be insufficient income to buy the products that the AI revolution creates. It is possible that in the long term, AI's benefits may exceed its costs, but that will occur in a very different economic setting than the one we have right now.
Every revolutionary change gives rise to new businesses, some of which are extraordinarily successful. The internet business was the starting point for online retailing, with Amazon as its most successful player, streaming entertainment, with Netflix and Spotify, as players, and search, with Yahoo! and Google cashing in. The smartphone was the vehicle used by a host by intermediary businesses, with Uber, Airbnb and Doordash, all changing the businesses they entered, as well as social media, with Facebook and Tiktok emerging as big winners. If AI's promise plays out, it is almost certain that it too will be the launching pad for new businesses. If you are tempted to count the revenues of these businesses as part of AI's payoff, there are two reasons for caution:
Net benefit test: All of the companies that I mentioned in the last paragraph were immensely successful, but some or much of their success came at the expense of existing players. Notwithstanding Amazon's astonishing success in online retailing in the last 25 years, retailing as a business has not seen explosive growth, since brick and mortar retailers have largely been left in the dust. The success of entertainment streaming companies has not only come at the expense of traditional entertainment companies, but in the aggregate, it has made the business less profitable. It is true that ride sharing has caused the car service business to become larger, as ride share use changes user behavior, but it has contributed to wounding the auto business. The most telling statistic on the impact of disruption, especially from technology companies in the last four decades, is that real GDP growth has been mostly stagnant in much of the world for that period, suggesting that the net impact on economic activity is modest, at best.
Sharing the pie: Even if you take the gross added value from new businesses (in revenues, earnings or value) for revolutionary change as a given, it is unclear what the companies that were the architects of that change gain as a result. The internet was built by telecom companies (with investments in cables and phone wires and supplemented by equipment companies (like Cisco), but none of them benefited from the rise of internet companies like Amazon. The smartphone was built and popularized by Apple and Samsung, but they don't get a slice of what Uber and Doordash make, when customers use smartphones to access their services. Apple benefits at the margin, as customers are coaxed into a cycle of replacing existing phones with newer versions, but the benefits are tangential.
If you look at the history of revolutionary changes, from the steam engine (factory age) and railroads, all the way to the technological changes of the last four decades, it has generally been the case that while the companies that build the architecture for these changes initially benefit, the big winners often catapult themselves off that architecture. You can make the argument that AI will be different, but for AI architecture companies to gain a slice of the new businesses that emerge from their technologies, they will need different business models.
2. Industry Economics
Accepting the premise that AI products and services will have a big market, in terms of revenues, is just the first step in structuring its business argument, and in this section, I will focus on converting those revenues into profitability, by first looking at the business models that AI producers can consider adopting, with pluses and minuses, as well as unit economics and economies of scale, measuring what it will cost companies to produce the AI products that they are selling.
a. Business Models
In keeping with the trial-and-error process that characterizes young industries seeking workable business models, we have seen AI businesses experiment with different versions on at least two dimensions:
Subscription versus Usage Models: Early in AI's business evolution, subscription models for both individuals and enterprises were the entry point for AI companies, and while those subscription models still bring in revenues, companies have shifted away from them for a simple reason. Unlike businesses like streaming or even software, where the marginal cost of an additional subscriber is zero or close to zero, AI products and services are expensive to generate, in terms of compute costs and data, and both Anthropic and OpenAI have discovered that subscribers, left unchecked, quickly become cost generators rather than profit centers. Thus, it should come as no surprise that Anthropic has shifted to usage-based models, where users (especially at the enterprise level) pay based on how much and how intensively they use AI products. OpenAI is still more dependent on subscription models that Anthropic, but it too is seeing a shift to usage-based models.
Source: Leaks and estimates
While advances may change the economics, it looks like while AI-subscription models will stay available, they will come with strict limits on usage, and that the bulk of revenues in this market will be based on usage. At the same time, there will be differences across AI companies, based on whether they are targeting the premium AI market, where usage-based models will dominate, or the mass market, where subscription models will continue to be offered.
Open versus Closed Models: There is an open and vigorous debate going on in AI circles as to whether the AI businesses should offer clients and customers open models (where clients can modify, adapt and build on the models) or closed models (where they are not allowed to do so). This debate is complicated because there are multiple forces that come into play including how power in this space is concentrated (with closed models giving its makers more power), how much privacy they offer (where the argument is that open models require less sharing of private data) and how safe they are (where it is posited that open models can be more easily hacked and turned into disinformation). All that said, there is clearly a business economics twist to this debate. Closed models give the companies that sell them more pricing power (higher margins), and perhaps are stickier (making it difficult to switch away), because they are customized, but they require more resources to build and maintain (higher costs) and may work only with premium products. As with subscription versus usage models, you are likely to see divergence, with companies targeting the premium market more likely to stay with closed models and those building more workhorse applications trying their hand at open models.
As I noted at the start, it is still early in the game, and as technology, regulation and cost structures shift, not only will there be more twists and turns involved, but it is likely that in steady state, we will have different choices for different segments of the AI markets, more subscription-based and open models in mass markets and more usage-based and closed models in premium markets.
b. Unit Economics
Through much of its existence, technology, as a business, has largely benefited from beneficial unit economics. It costs a software company almost nothing to sell its next unit, and for most platform companies (Netflix, Uber, Airbnb), the cost of adding a user or subscriber, once the platform is constructed is negligible. As a consequence, being the largest player or a first mover can put you on a pathway to industry dominance, with large market share and high profit margins thrown in. Early in its life, AI was thrown in by some into the technology pile, and it was assumed that it too would have the same characteristics, i.e., that it would cost little or nothing to produce the additional unit and that high margins and dominant market shares will follow.
With the caveat that things can change quickly, the evolution of the AI product and service market has turned out to be different, and some of those differences look like they are baked in. Beginning with the fact that AI products require more capital intensity, in terms of data center build and cap ex prior to operation, they are also proving costly to produce, at least in their most powerful forms. One way to see the dual forces driving unit economics in AI at play is to focus in on AI tokens, the currency of AI production. The good news, in terms of unit economics, is that the cost of producing a token has dropped dramatically, as AI infrastructure gets built out, but the bad news is that the tokens used to create AI products has surged almost as dramatically, as these products become more powerful. As a result, the price of accessing frontier AI models has increased over time:
These two trend lines point to a divergence that is coming to the AI products and services market. In mass market AI, where AI tools are basic and don't need power boosts, you should expect costs for AI products to go down. In the premium market, where you are building AI products either as tools on high-end tasks or as replacement for highly priced labor, the costs will be tougher to reduce, if each upgrade in product power puts demands on the inputs - more data to process, more power to run data centers and more powerful chips in the data centers - that causes these input costs to rise.
c. Moats and Competitive Advantages
Allowing unit economics and economies of scale to play out in the AI business, you still have a final piece of the business puzzle to consider to make the leap to profitability, where you bring, as you would in any business, the moats and competitive advantages that will separate the winners from the wannabes. Since this is a question that we ask about every business, it makes sense to start this discussion by looking at potential competitive advantages in any business:
Focusing in on AI, the question that we first face is which of these moats is most likely to be the defining one in AI, and I believe that the answer depends on which segment of the AI market you are looking at:
In mass market AI, where products and services are standardized and basic, you should expect competitive advantages to flow from unit costs being lower at a company than at its competitors, either because of scale (with bigger companies having an advantage) or proprietary access to data.
In premium AI, where products are customized, powerful and pricey, you should expect the winners to be companies that have the technological know-how to craft these products, while bringing the costs of delivering power under control. In addition, products that are built around client data, especially if the client is protective of that data, will become stickier and more difficult to displace, giving companies that make them more pricing power for a longer period.
Does having the people perceived as the smartest in the AI space working for you give you a competitive advantage? At the moment, the answer seems to be yes, and you saw this phenomenon at play when Jeff Dean, chief scientist at Google DeepMind (its AI entity), left the company in 2026 to create his own AI start-up, and the market reacted by knocking Alphabet's market cap down by 5.4% (more than $100 billion). In the same vein, the AI firms (especially Anthropic and OpenAI) have been raiding universities for their computer science and technology talent, with some AI-focused economists thrown into the mix and even a few philosophers. I think that the attention paid to these people hires and departures are indicative of how young this industry is, and how much its success will depend on building products right and marketing them to the right customers, and as it matures, I expect this factor to fade in prominence.
Is there a potential brand name advantage? Put simply, would you be willing to pay a premium price for a Claude agent over an AI agent crafted by a different company? The answer is still being worked out, because in addition to all of the business elements of this choice (product power, reliability), trust is a factor, since client companies are giving AI products access to secrets and data. It should come as no surprise then that AI companies are all competing in the virtue space, where each one puts itself out as more trustworthy and caring about public good than the next one. It may be cynical of me, but when I hear Dario Amodei or Sam Altman wax eloquent about how they plan to protect the world from the dark side of AI, I feel the urge to quote Shakespeare, and say "thou doth protest too much". Ultimately, actions speak louder than words, and these companies will be judged based more on how they behave, when confronted with ethical questions, than on what the write about themselves.
3. Constraints and Limits
While much of this post has been about AI's business prospects and evolution, there are parallel discussions that are occurring about AI's impact on society. There are four reasons why AI's social and cultural effects are being so widely debated:
Real estate footprint and resource usage: AI, in terms of the investment footprint it is creating, is closer to the railroads in their early years than it is to any technology company. Like railroads, AI requires data centers that sprawl over huge areas, and unlike railroads, many of these areas have people living in them, whose lives will be altered by the presence of these centers. While the proponents of data centers have sold them on the basis of the economic benefits they bring, and these benefits can be real, it is quite clear that for many people who live in the vicinity, the upending of their lives is not worth the benefits. As a result, the backlash against data centers is real, showing up in politics at not only the local level, but also at the national level; it is quite clear that at least in the United States, it will be a lead topic, perhaps even a wedge issue, in the next presidential election. Another reason that data centers lead to resentment is their disproportionate use of power and water, and even if that cost is pushed back to AI companies, the costs to the planet are still being totaled.
Data privacy and power: Earlier in this post, I pointed to privacy as one of the dividing lines in the choice between open and closed AI models, but that is a small part of a broader question, which is about data being accumulated by and mined at AI companies. After two decades of seeing social media companies step across the privacy line in their use of private data, it is understandable that there is wariness about granting access to even vaster amounts of data to the likes of Anthropic and OpenAI.
People displacement: Going back to the discussion of the total market for AI, I noted that the best case scenarios for AI, i.e., the scenarios where the market for AI products and services will be greatest, are also scenarios where there will be not just be significant job loss, but losses in high-paying jobs. While tech advocates will point to new jobs that will be created to replace the ones lost, that transition, even if it happens, takes time and will come with pain. Looking back at the disruption of blue collar jobs in the US and Europe, from China, that disruption created pain, albeit unevenly shared, and has led to political and economic aftershocks that are still playing out across the world. If AI's disruption plays out on a broad front, the resulting displacement will be much larger, in terms of economic impact, and perhaps much more painful.
Income equality and fairness: For much of this century, one of the recurring themes in both political and economic circles has been the growth of the divide between the super rich and the rest of society. There is a suspicion that AI will make this divide wider, and that suspicion will only intensify with each AI company that goes public. SpaceX, when it went public at a market capitalization of close to $2 trillion created a host of centi-millionaires (worth more than $100 million) and the same phenomenon will play out when Anthropic and OpenAI hit the market. I am not a fan of setting economic policy based on greed and envy, but you can see the pushback against inequality playing out in the political arena.
If you are tempted to ignore these discussions, because you are an investor or interested only in the business aspects of AI, it is only a matter of time before these spill over into economic consequences, and then into the metrics that drive business value.
Data centers will get more expensive to build, and power and water will be more tightly rationed, leading to companies having to spend both more on their upfront and capital expenditures, and as costs in generating AI products.
As concerns about data mount, there will inevitably be scandals around the misuse of data, and those scandals will lead, as they did at social media companies, to tighter restrictions on the use of data and higher costs is acquiring and protecting that data.
The worries about being replaced by AI agents will create counter movements, starting with system requirements that preserve jobs for humans (even if AI makes what they do obsolete) and in some countries, requirements that their employers continue to pay displaced employees for extended periods, in the event of layoffs.
If AI creates its own cache of billionaires and centi-millionaires, the push towards wealth taxes, no matter what you may think about their effectiveness or fairness, will intensify, as will the creation of new tiers in the tax table for higher incomes, and perhaps for AI income.
It has always been difficult to start and build businesses in new industries, and ti becomes doubly so when the rest of the world consigns you to the dark side. If you are an investor in this space, especially one excited about the size of the potential market and pathways to profitability, you need to be realistic in incorporating the constraints that are already cropping up, and will become more binding, into your valuation.
As AI's story line unfolds in ways that invite public pushback, it is worth noting the possibility that this storyline can still change, especially if AI delivers on its promise to change people's lives for the better. In the last year, I have seen stories, some already happening and some far fetched, of AI's capacity to solve mathematical challenges that have deterred our greatest mathematicians, to diagnose diseases that the best-trained doctors seem to miss and perhaps even to come up with cures for diseases that have eluded those seeking them. That power comes from tAI's brain capacity, which unlike those of human beings, is expandable and its ability to not just remember everything that is fed into it, but make connections across data in different disciplines. As I noted earlier, the benefits from these advances will, for the most part, flow not just to other businesses, but will benefit society. If the big AI players want a more welcoming environment to grow, it behooves them to make these social investments, perhaps drawing on AT&T's nurturing of Bell Labs for much of the last century as a place for research that benefits humanity as inspiration. It is true that none of the big AI companies is a regulatory monopoly, as AT&T was, but it should be feasible for them to pool resources to fund a Bell Labs like entity. It will be costly, but they will collectively benefit from the social dividends.
From Macro to Micro: Zeroing in on company valuations and investments
This post has focused on AI as a business, and if your interest is in valuing an enterprise in the space, whether publicly traded like SpaceX or Alphabet or privately owned (but heading for an IPO like Anthropic or OpenAI), you may wonder how it helps you in that endeavor. There are two ways you can approach these valuations. In the first, you can start with the market capitalization (actual in the case of publicly traded companies and estimated in the case of private businesses) and examine what you would need in terms of end revenues and profit margins to justify the market capitalization. In the second, you can start with the company that you are valuing, and lay out a roadmap for that company in terms of product choices (premium or mass market, open or closed) and estimate what it will be able to deliver in terms of revenues, profits and cashflows over its lifetime.
1. Reverse Engineering Breakeven Points
In a post, late last year, I looked at the handful of companies, mostly tech, that have trillion dollar market capitalizations, and rather than take the knee-jerk reactions, i.e., that they must be overvalued with that market cap or that they must be great companies, because the market assigns them a lofty price, I looked at what these companies will have to deliver in terms of operating metrics in the future to justify their valuation.
We are effectively reversing the intrinsic value process, and trying to answer the question of how much revenues will have to be in a future year (where you specify when the company or business will be mature or steady state), given your company characteristics in terms of risk, profit margins and reinvestment needs. With Anthropic, for instance, where the rumored pricing for the IPO is $2 trillion, allowing the company premium pricing margins (after-tax operating margin of 30%) and above-average risk (cost of capital of 10%), the company will have to generate close to $1.2 trillion in revenues, if the AI market matures in ten years, and close to $2 trillion, if the wait is 15 years. That should give ammunition to both those bullish about the company, because in their story line, Anthropic products will be premium priced and replace workers across industries and geographies, and to those who are bearish, since that scenario looks unlikely.
In fact, you can consider AI companies in the aggregate, by adding up the market capitalizations of companies that are already public (or at least the portion of the revenues that come from AI) as well as the rumored pricing of companies like OpenAI and Anthropic, waiting to go public, and going through the same exercise. Using an aggregated market pricing of $5 trillion (probably a conservative judgment, given the VC pricing of hundreds of companies in the space) attached to all AI product and service companies, and assigning a blended operating margin of 20% for the industry, the revenues that you would need for the entire business to breakeven would be $5 trillion, with a 10-year wait, and more than $8 trillion, if the wait is 15 years. Looking back at the discussion of the total addressable market in the earlier section, you can see that this would represent quite a reach, a manifestation of the big market delusion.
I will be the first to admit the limitations of this reverse engineering, but when data is still scarce or non-existent, it does provide a framework for reasonableness and a constraint on story telling. In fact, you can use this framework to examine what any investment, whether it be the price you pay as an investor for an AI business or the capital expenditure into AI made by a company, will have to generate to break even as an investment. Thus, if you are questioning whether Microsoft or Meta's AI cap ex is value creating or destroying, you can use this generic breakeven spreadsheet, to make your own judgment.
2. Build up to value
If you follow the AI business script laid out in the class, you also have a process for valuing any AI company that hopes to make money in this space, but to put this process into the play, here are some of the issues that you will have to address to estimate value.
Product choice and market focus: If the AI market splits into premium and mass-market product market, the first step in valuing any company in this space will be to make a judgment on which of these markets the company will target, and how much of its revenues will come each of the segments.
Unit economics and economies of scale at company: The choice of market segment matters because the unit economics and economies of scale you assume for the company will have to be consistent. AI companies that offer mass market products will charge lower prices, with a greater percentage of revenues coming from subscriptions, but will benefit more quickly from improving unit economics, as the cost of AI tokens continues to fall. In contrast, AI companies that target premium markets, will earn higher profit margins and have stronger moats, but struggle more with unit economics, as token usage increases with product power.
Competitive advantages and moats: The types of competitive advantages (moats) that the company you are valuing will seek out, and lock in, if successful, will also vary depending on the targeted market, with cost advantages and scale working in the company's favor, with mass markets, and technological edges and product stickiness being more sought after, with premium products.
Investment needed to deliver growth: While AI companies are more capital intensive than their tech counterparts, the additional reinvestment needed to deliver value can be altered by investments already made by a company. Companies that have built capacity in advance of growth will be worth more than companies that will have to reinvest contemporaneously to deliver growth, and companies that find ways to invest more efficiently will also have higher value. It is interesting that starting with Deepseek, China seems to be trying the latter path to AI dominance, using less expensive (and less powerful) AI chips and not investing as much in mega data centers, and it may very well be the right choice, for much of the AI product and service market.
Regulatory constraints (current and in the future): To the extent that regulators and governments have a great deal at stake, the value of a company can be affected by where it operates geographically and the rules and regulations that govern AI products in that geography. If past behavior is an indicator, the EU will be an inhospitable setting, for AI products and companies, and that may make a difference in how you value Mistral, with a base in France and more European-focused clients.
I did try my hand at this process, when I valued xAI as part of SpaceX, and I learned from doing so, but the company's stakes in space launch and internet service did muddy the waters. As Anthropic and OpenAI move towards their public offerings, I am looking forward to applying the framework developed in this post to those firms, when their prospectuses are made public. In keeping with my belief that it is best to be open about biases, I will confess that the rumored pricing for both companies ($1.5 to $ 2 trillion) looks rich, but I am open to being surprised.
Conclusion
At the start of this post, I noted that I was writing this post for myself, because like many in this space, I was finding myself pulled in a dozen different directions on AI, and desperately in need of a framework for thinking about whether I should be paying attention in the first place, how to reconcile competing viewpoints and what it means to me, as an investor. This is my try at creating a comprehensive framework, and I am sure that there are elements that I have missed and holes in my thinking, but it is a start. I am clear eyed about what this AI framework will not and will do for me; it will not tell me what the TAM for AI products and services will be or whether Anthropic is worth $ 2 trillion, but it will give me bounds for my estimates of TAM, allow me to determine that a $22 trillion TAM for AI is fiction and recognize that having your ARR grow 80% a year last year is not even close to being a rationale for why you should buy Anthropic at a $2 trillion pricing.
As you work your way through the framework, there will be room for significant disagreement on the reach of AI and its value as a business and anyone who claims to have conviction that they know what's coming is being either ignorant or arrogant. That ties in well with my last post, where I examined the swift rise and the even swifter fall of Leo Aschenbrenner, whose entire investment strategy was built around the conviction that AI would decisively and quickly win the disruption war. The problem that I noted was not that his vision was not plausible (it was), but that it was definitely not certain, or at least assured enough to borrow immense amounts to fund it.