Musings on Markets is the personal blog of Aswath Damodaran, professor of finance at NYU Stern and widely known as the "Dean of Valuation." Running since 2008, it publishes hands-on intrinsic-value teardowns of headline companies (SpaceX, Tesla, Nvidia) using his narrative-and-numbers DCF framework, plus periodic market-wide reviews.
This piece says AI has moved from hype to a 'bar mitzvah' moment where it must prove it can make real money. The author, Damodaran, is cautious because AI companies have invested $1.7 trillion but only generate about $250 billion in annual revenue—a huge gap that risks a bubble. Key names: Nvidia (the biggest winner so far, selling AI chips), Anthropic (its annual revenue fell $10-15 billion short of expectations, showing unpredictability), and OpenAI (around $40 billion in revenue but still spending heavily). The takeaway: don't just buy the story; demand evidence of profits.
One-sentence summary: AI has moved from "hope and hype" into the "business validation" coming-of-age moment, but the vast chasm between $1.7 trillion in investment and only about $250 billion in annual revenue exposes the market to valuation bubble risk. [Cautious]
The article notes that since the launch of ChatGPT on November 30, 2022, AI has transitioned in public consciousness from a positive phenomenon to a negative tinge, partly due to concerns that "the genie is out of the bottle" and is not benevolent, as well as the off-putting nature of some key spokespeople. The author recalls that AI's breakthrough moment occurred less than four years ago, although its history dates back to the dawn of the computer age (e.g., IBM's Deep Blue evaluating 200 million chess positions per second). Over the past four years, the AI effect has exploded: the most successful company is chipmaker Nvidia (by market cap growth); the Mag Seven (including Tesla and Nvidia, replacing Netflix) contributed 45% of U.S. stock market cap growth between 2022 and 2025, with a total market cap of $23.7 trillion as of August 16, 2026; AI infrastructure investment accounted for roughly 1% of the 2.5% real GDP growth in 2024–2025. The author states: "As the arc of the AI story has unfolded over the last four years, it seems to me that it has also transitioned in the public consciousness from a mostly positive phenomenon early on to acquiring a negative tinge." At 2026 U.S. college graduation ceremonies, AI-themed speakers were booed by students.
The author argues that the debate over AI has gone off track, with optimists focusing on the "enormous" potential market and pessimists fixating on "excessive" upfront investment, with both sides selectively using data. Optimists highlight usage levels and growth, while pessimists emphasize capital expenditure and current profitability (or lack thereof). The author stresses that this article is not about proving which side is right, but about judging AI as a business—acknowledging its breakthrough nature while demanding that it prove its value, like all other businesses in history, by converting potential into products and revenue, delivering products at costs that generate profits, and building moats to fend off new entrants. The author states: "big markets don't always become big businesses." Pessimists must also acknowledge that massive capital expenditure raises the stakes but does not necessarily doom value destruction.
The author proposes that every major disruptive change goes through four stages: the Hope and Hype Phase, the Investment and Build Phase, the Business Construction Phase, and the Recalibration Phase. The author positions himself as an AI novice (using only the free version of ChatGPT and knowing little about Claude) and states that this article is written to clarify his own thinking.
1. Hope and Hype Phase: True believers and visionaries need to sell the story of change to the public, but at this stage there are no physical products, services, or revenue, inevitably accompanied by false starts and scams, and many people refuse to listen due to skepticism or lack of understanding. For change to take root, visionaries must be persuasive enough to attract supporters and capital.
2. Investment and Build Phase: Once belief in change is established, some players among startups or incumbents will invest and build products and services. If the change is seen as revolutionary and the market as huge, the "big market illusion" (coined by the author over a decade ago) emerges—driven by selection bias, builders tend to be overconfident, leading to collective overexpansion by companies and the investors pricing them, followed by a correction. The author emphasizes: "bubbles are a feature, not a bug, when revolutionary change is a possibility." In this phase, stories drive growth and investment; investors with an actuarial or accounting mindset find these narratives unconvincing, while others are willing to bet on option value, hoping that the entities they back become big winners (though they have yet to win anything) in a large market (which does not yet exist).
The report argues that AI has passed through the "hope and hype" and "capital deployment" phases and is now entering a critical "business building" period — a "bar mitzvah moment" that tests whether companies can transform products into sustainable business models. The author emphasizes that not all changes are revolutionary, and not all revolutionary changes translate into large markets or valuable enterprises. At this stage, investors are no longer willing to price based solely on promises but demand tangible evidence of progress. The author states: "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."
The report's core argument is that AI infrastructure investment has reached $1.7 trillion, while the current upper bound of annual revenue generated by AI products and services is only about $250 billion — a massive gap. The author likens AI investment to building the "most expensive factory in history" at an unprecedented pace. Capital expenditure data for six major players is listed:
| Company | Role |
|---|---|
| Alphabet | Mag Seven member |
| Amazon | Mag Seven member |
| Meta | Mag Seven member |
| Microsoft | Mag Seven member |
| Oracle | External major player |
| Coreweave | External major player |
The author notes: "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."
The report proposes an analytical framework, arguing that assessing AI's commercial value requires examining three dimensions: market size, industry economics, and competitive moats. The author criticizes "total addressable market (TAM)" as a tool often manipulated by founders, venture capitalists, and bankers — for example, in SpaceX's valuation, the banker-provided TAM of $22 trillion for xAI is described as "a hallucination, not an estimate."
The report's core conclusion is that the AI narrative is clearer than a year ago, but it remains in an early stage, with investors facing significant uncertainty. The author explicitly defines the current moment as AI's "bar mitzvah moment," meaning the market will shift from "storytelling" to "performance-based evaluation." For investors, this implies:
The report argues that the ultimate market size of AI depends on its positioning—whether as a "tool" or "employee replacement"—and that the total U.S. employee compensation of $12.96 trillion serves as the hard ceiling for TAM. The author notes that global listed companies' total operating expenses in 2025 are approximately $65 trillion, but AI cannot replace physical costs such as raw materials; the truly addressable portion is employee compensation. According to Federal Reserve data, total U.S. employee compensation in 2025 stands at $12.96 trillion, and on a global scale, this figure is roughly double. However, the author emphasizes that fully replacing all employees is neither realistic nor economical, and the actual TAM depends on four key questions:
1. Tool vs. Employee Replacement: If AI serves as a tool, companies will incur additional tooling costs on top of existing employee expenses, resulting in a smaller market size; if it replaces employees, it directly impacts the compensation pool, leading to a larger market. The author states: "The takeaway...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 author also notes that current AI players (OpenAI, Anthropic) tend to favor faster disruption, as this increases their odds of success, despite higher social costs.
2. AI Product Pricing: Currently, high-end products such as Anthropic's Claude and OpenAI's Codex are expensive and only economically viable for high-wage labor. In 2023, the highest-income quintile in the U.S. accounted for 51% of total employee compensation, so the actual replaceable market is at most half (approximately $6.48 trillion).
3. Application Breadth: AI penetrates faster in rule-driven, low-interpersonal-interaction industries (e.g., technology, finance), which account for less than 20% of global operating expenses; while resilient industries such as industrials, materials, real estate, and utilities account for one-third of the global total. Harvard Business Review research shows that software and coding are the easiest areas for AI to enter.
4. Geographic Distribution: AI is more likely to cause disruption in high-wage, non-manufacturing regions (e.g., the U.S.), with Europe and China being the next largest markets. Global operating expense data indicates that the U.S., Europe, and China are the three major AI markets.
In summary, with $26 trillion (global employee compensation) as the ceiling, TAM valuation depends on judgments across these four dimensions: the most optimistic scenario (AI as employee replacement, covering all industries globally, low-cost replacement of low-wage workers) approaches the ceiling; the most pessimistic scenario (AI only as a tool, limited to specific industries and regions) falls far below it.
The report points out that the AI industry is shifting from a subscription model to usage-based pricing, as the marginal cost of AI products is far higher than that of streaming or traditional software, and subscribers may become cost centers. The author analyzes that early AI companies (e.g., OpenAI, Anthropic) started with subscription models but quickly found that the computing and data costs per additional user were high. Anthropic has taken the lead in shifting to a usage-based pricing model, where enterprise users pay based on usage intensity and frequency; OpenAI still relies more on subscriptions but is also moving in this direction. The author believes this shift reflects the unit economics of AI products—high marginal costs make fixed-fee models unsustainable.
The report argues that AI business models will diverge between subscription-based and usage-based pricing, with premium markets primarily adopting usage-based models and mass markets retaining subscriptions. The author believes that while technological progress may alter economic models, AI subscription models will come with strict usage limits, with the majority of revenue derived from usage. Meanwhile, different AI companies will differentiate based on target markets: premium markets will mainly use usage-based pricing, while mass markets will continue to offer subscriptions. The author states, "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." This means: "It is highly probable that in a steady state, different segments of the AI market will have different options—mass markets will see more subscription-based and open models, while premium markets will lean toward usage-based and closed models."
The author argues that closed models grant companies stronger pricing power and customer stickiness but require more resources to maintain, making them suitable for premium markets; open models are the opposite. The report notes that there is a heated debate in the AI field between open models (which customers can modify, adapt, and build upon) and closed models (which prohibit modifications), involving multiple factors such as power concentration, privacy protection, and security. From a business economics perspective, closed models give selling companies stronger pricing power (higher profit margins) and, due to customization, greater stickiness (difficulty in switching), but they require more resources to build and maintain (higher costs), potentially limiting them to premium products. The author judges that companies in premium markets are more likely to stick with closed models, while those building foundational applications will experiment with open models.
The report emphasizes that AI unit economics differ from traditional technology—while the cost per token has dropped significantly, the demand for tokens in products has surged, making it difficult to reduce costs in premium markets. The author points out that traditional technology (e.g., software, platform companies) has extremely low marginal costs, but AI products require higher capital intensity (data center construction and upfront capital expenditure), and producing high-compute versions is costly. Specifically, the production cost of AI tokens (the currency of AI production) has fallen sharply due to infrastructure buildout, but the number of tokens required to create AI products has surged almost in tandem. The author states, "the cost of producing a token has dropped dramatically... but the bad news is that the tokens used to create AI products has surged almost as dramatically." This means: "The cost of producing a token has dropped significantly... but the bad news is that the number of tokens used to create AI products has surged almost as dramatically." Consequently, AI costs in mass markets will decline, while in premium markets, costs are hard to reduce due to increasing demands for data, electricity, and chips with each upgrade.
The author believes that competitive advantages in the AI market vary by segment—mass markets rely on scale and proprietary data, while premium markets depend on technical capabilities and customer data stickiness. The report notes that in mass-market AI (standardized foundational products), competitive advantages stem from lower unit costs (scale advantages or proprietary data); in premium-market AI (customized, high-priced products), winners will be companies with technical knowledge and cost control capabilities, and products built around customer data will have greater stickiness. The author specifically highlights talent advantages: when Google DeepMind's chief scientist Jeff Dean left in 2026 to found an AI startup, the market reaction caused Alphabet's market cap to drop by 5.4% (over $100 billion). AI companies (especially Anthropic and OpenAI) are poaching computer science and technical talent from universities. The author believes that as the industry matures, the importance of talent will diminish. Additionally, brand trust is a competitive factor—client companies expose AI products to sensitive data and information, so AI firms are vying for a "trustworthy" image, but the author is skeptical, arguing that "actions speak louder than words."
The report points out that the social and cultural impacts of AI are being widely debated, primarily involving four constraining factors. The author lists four reasons (not elaborated in detail in the original text but mentioned as section headings): these discussions run parallel to AI's commercial prospects, including potential effects on employment, privacy, security, and social inequality. The author does not delve into analysis here but implies that these external factors may in turn influence AI's commercial development path.
The core investment implication of the report is that the AI market is not a single track; investors need to distinguish between mass markets (cost competition, scale advantages) and premium markets (technology competition, customer stickiness), and pay attention to the divergence in business models (subscription/usage-based pricing, open/closed models). Institutional perspective bias: As an academic analyst, the author emphasizes business logic over short-term market sentiment, but it should be noted that the analysis is based on the current technological stage; future technological breakthroughs could alter unit economics and the competitive landscape.
The investment model for AI data centers more closely resembles early railroads than tech companies, and their land, electricity, and water consumption have sparked political backlash from local to national levels, set to become a key issue in the next U.S. presidential election. The author states, "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." The author notes that while proponents market data centers based on economic benefits, for nearby residents, the disruption to daily life far outweighs the gains. This resentment also stems from data centers' excessive consumption of electricity and water resources; even if costs are passed on to AI companies, the overall toll on the planet continues to accumulate. The consequence: data center construction will become more expensive, and electricity and water resources will be more strictly rationed, driving up both upfront capital expenditures for companies and the cost of generating AI products.
Following social media companies' abuse of private data, the public is wary of companies like Anthropic and OpenAI accessing even larger datasets. Data misuse scandals will be inevitable, leading to tighter data usage restrictions and higher costs for acquisition and protection. The author argues that data privacy is one of the dividing lines between open and closed AI models, but it is only a small part of a broader issue—the data itself that AI companies accumulate and mine. The author judges that, as with social media companies, data scandals will spur stricter regulation, directly raising compliance costs for AI firms.
The best market scenario for AI (the largest product and service market) is also one where a large number of high-paying jobs are lost. The economic shock will far exceed the previous displacement of blue-collar workers in Europe and the U.S. by Chinese manufacturing, and the political and economic aftershocks will persist. The author points out that while technology advocates will emphasize the creation of new jobs, the transition takes time and comes with pain. Reflecting on the impact of Chinese manufacturing on blue-collar workers in Europe and the U.S.—the pain, though uneven, was real, and the political and economic aftershocks are still playing out globally. If AI's impact unfolds across a broader range, the resulting job displacement will be greater in economic scale and potentially more painful. The author predicts that fears of AI-driven replacement will fuel counter-movements, including systematic mandates to retain human jobs and, in some countries, regulations requiring employers to continue paying displaced workers after layoffs.
AI will create a cohort of billionaires and multi-millionaires, inevitably intensifying societal concerns over the gap between the super-rich and ordinary people, and fueling political pressure for wealth taxes, new high-income tax brackets, and even specific taxes on AI-generated income. The author cites the example of SpaceX's IPO (with a market cap of nearly $2 trillion), which created numerous multi-millionaires with net worths exceeding $100 million, and notes that the same phenomenon will repeat when Anthropic and OpenAI go public. The author admits to not endorsing economic policies based on greed and envy but acknowledges that the backlash against inequality is already playing out on the political stage. The conclusion: regardless of whether wealth taxes are effective or fair, the push to tax wealth created by AI will intensify.
Investors cannot ignore the above socio-political constraints, as they will soon translate into economic consequences and affect business value metrics. The author emphasizes that those investing in AI, especially those excited about potential market size and profitability paths, must realistically incorporate these emerging and increasingly stringent constraints into valuations. The article concludes with a valuation framework: reverse-engineering the breakeven point—using Anthropic's rumored $2 trillion IPO valuation as an example, assuming a premium pricing (30% after-tax operating margin) and above-average risk (10% cost of capital). If the AI market matures in 10 years, the company would need to generate approximately $1.2 trillion in revenue; if it takes 15 years, nearly $2 trillion. For the entire AI industry (with a combined market cap of roughly $5 trillion), assuming an average operating margin of 20%, a 10-year maturity would require $5 trillion in revenue, and 15 years would require over $8 trillion—the author believes this borders on "big market delusion." Institutional perspective bias: As a valuation scholar, the author tends to test market narratives with rigorous data frameworks, and the judgment of "big market delusion" itself represents a stance; readers should note this conservative leaning.
The author acknowledges the limitations of reverse engineering but argues that, in the absence of data, it provides "rationality boundaries" and "constraints on narratives." He offers a generic breakeven calculator that can be used to test whether any AI investment—whether the price paid for an AI company or the company's own AI capital expenditure—creates value. In his own words: "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."
The author argues that to value any AI company, five key questions must be addressed in sequence:
1. Product Selection and Market Positioning: Determine whether the company targets the high-end or mass market, and the proportion of revenue derived from each segment.
2. Unit Economics and Scale Effects: Mass-market products have low pricing and high subscription revenue share, but benefit from faster improvement in unit economics as AI token costs decline; high-end products offer higher margins and stronger moats, but increased token usage as product capabilities improve puts pressure on unit economics.
3. Competitive Advantage and Moat: Mass-market players rely on cost advantages and scale, while high-end players pursue technological leadership and product stickiness.
4. Investment Required for Growth: Companies that have pre-built capacity are more valuable; those with higher investment efficiency are also more valuable. The author specifically notes: "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."
5. Regulatory Constraints (Current and Future): The rules in a company's domicile affect its value. The author judges: "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."
The author previously used this process to value xAI (as part of SpaceX), but found the analysis "muddied" by the company's involvement in space launches and internet services. He looks forward to applying this framework to Anthropic and OpenAI once their prospectuses are released.
The author explicitly states that this framework will not tell him what the TAM for AI is, nor whether Anthropic is worth $2 trillion. However, it will allow him to bound his estimates of TAM, 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." In his own words: "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."
Institutional Perspective Bias: The author admits that he "feels the rumored $1.5–2 trillion valuations for Anthropic and OpenAI are high," but is "willing to be surprised." He also notes that anyone claiming absolute certainty about AI's future is either ignorant or arrogant—echoing his previous critique of Leo Aschenbrenner: the entire investment strategy rests on the belief that "AI will decisively and rapidly win the disruption battle," but the problem is not that the vision is unreasonable, but that it is "absolutely uncertain, at least not enough to borrow huge sums of money to bet on it."
| Ticker | Direction | Author's One-Sentence View | Key Data |
|---|---|---|---|
| Anthropic | Hold & Watch | Rumored IPO valuation of $1.5-2 trillion, but the author considers it too high and needs reverse engineering to verify reasonableness | ARR of ~$65 billion at end of July 2026 ($10-15 billion below expectations); a rumored $2 trillion valuation would require generating ~$1.2 trillion in revenue over 10 years |
| OpenAI | Hold & Watch | Also faces valuation bubble risk; business model is shifting from subscriptions to usage-based billing | ARR of ~$40 billion at end of July 2026 |
| Nvidia | Hold & Watch | Biggest winner in revenue and profit so far in the AI cycle (infrastructure supplier) | Most successful AI company over the past four years by market cap growth |
| Alphabet | Hold & Watch | Mag Seven member, massive AI capex; Jeff Dean's departure caused a 5.4% market cap decline (over $100 billion) | Capex listed among the six major players |
| Amazon | Hold & Watch | Mag Seven member, massive AI capex | Capex listed among the six major players |
| Meta | Hold & Watch | Mag Seven member, massive AI capex | Capex listed among the six major players |
| Microsoft | Hold & Watch | Mag Seven member, massive AI capex | Capex listed among the six major players |
| Oracle | Hold & Watch | Major external AI infrastructure player | Capex listed among the six major players |
| Coreweave | Hold & Watch | Major external AI infrastructure player | Capex listed among the six major players |
| SpaceX (xAI) | Hold & Watch | The author previously valued xAI using a five-step process, but the involvement of space and internet businesses "muddied the waters" | The banker's $22 trillion TAM for xAI is considered a "hallucination" by the author |
| Deepseek | Hold & Watch | Chinese AI company pursuing a low-cost route (cheaper chips, fewer data centers); the author believes this may be the right choice for most of the AI market | Uses cheaper, less powerful AI chips |
| Mistral | Hold & Watch | Headquartered in France, clients are mostly European; the author believes the unfavorable EU regulatory environment will affect valuation | The EU is judged to be an unfriendly environment for AI products |
| Google DeepMind | Hold & Watch | Chief Scientist Jeff Dean left to start an AI startup, triggering a decline in Alphabet's market cap | Departure event caused Alphabet's market cap to fall 5.4% (over $100 billion) |