The Case Against the AI Bubble
Is the AI boom a bubble? Azeem Azhar analyzes the data, revealing spending, infrastructure, and demand factors. Is it rational exuberance or a crash waiting to happen?
The rapid ascent of artificial intelligence has led to unprecedented valuations for companies in the sector, prompting a critical question: is the current AI landscape a bubble poised to burst? Nicholas Thompson, CEO of The Atlantic, explores this complex issue with Azeem Azhar, founder of Exponential View, who has conducted in-depth research into the economics of AI.
Mapping the AI Economy
Azhar's research aims to provide clarity on the actual spending within the AI economy, a sector where visibility is often obscured by private company valuations and the broad revenue streams of hyperscalers like Microsoft, Amazon, and Google. His methodology involves a form of "forensic accounting" to reconstruct financial accounts and align public disclosures with internal realities.
"We went off and tackled the thing that people don't have visibility on in the AI economy, which is how much is actually being spent," Azhar explained. "So we know lots of things about AI. We know how many chips are being bought. We know how many data centers are being brought online. We know the amount of money companies are raising their private investment rounds because they advertise them all. But the thing that really rounds out the picture is are companies really spending money and are they spending money that they intend to continue to spend next year and the year after that?"
Through this process, Azhar estimates that approximately 126 billion by July 2026. This number is de-duplicated, meaning that if a dollar is spent with OpenAI, which in turn spends money with Microsoft, it is counted only once as value added.
Infrastructure Investment and Depreciation
A significant portion of AI spending is directed towards infrastructure, particularly compute power. Capital expenditures (CapEx) dedicated to AI infrastructure, which were in the low billions in 2021, are projected to exceed $100 billion by 2026. The core question for the sustainability of this investment lies in whether the revenue generated will justify the costs of depreciation and operating expenses.
A key variable in this calculation is the depreciation schedule for AI hardware, such as GPUs. While theoretical models might suggest a four-year depreciation period, Azhar points to real-world data indicating longer lifespans. Executives from Amazon have reported chips in use for six to six and a half years, and Google has noted AI chips in operation for eight to nine years. CoreWeave, a cloud company, has also reported six-year-old chips remaining fully utilized.
However, the value of this extended use is debated. If older chips are used for low-value tasks, a shorter depreciation schedule might be more appropriate. The current market, characterized by high demand and limited compute capacity, has led to long-term contracts commanding higher prices than spot market rates, a situation that could mask underlying economic vulnerabilities.
Demand and Productivity Gains
The sustainability of AI investment hinges on genuine demand, which is closely tied to demonstrable productivity gains. While many companies are adopting AI tools, a significant portion of CTOs report minimal productivity boosts. This raises concerns about whether demand will continue to skyrocket if companies struggle to achieve substantial improvements.
Azhar's conversations with IT leaders reveal a mixed picture. While few report significant productivity gains, there is a persistent belief in AI's potential, with many companies continuing to invest and experiment. The "J-curve" hypothesis, proposed by Erik Brynjolfsson, suggests that initial AI adoption may decrease productivity before leading to massive gains. However, distinguishing this genuine learning curve from performative adoption driven by market and board expectations remains a challenge for investors.
The Rise of Open Source and Competition
The increasing availability and capability of open-source AI models present another dynamic. Companies are beginning to shift towards these models, potentially reducing reliance on expensive proprietary platforms like OpenAI and Anthropic. This trend could exert downward pressure on pricing and impact the revenue streams of major AI providers.
While this shift might not significantly affect hyperscalers, as open-source models can still run on their cloud infrastructure, it poses a challenge for companies like OpenAI and Anthropic. These companies are responding with price cuts and enhanced data privacy measures, such as zero data retention policies, to retain enterprise clients.
The competitive landscape also includes China, which is making significant strides in AI despite having a fraction of the compute power available to American firms. Chinese companies are demonstrating remarkable efficiency in model development, driven partly by restrictions on high-end chip exports. However, the Chinese AI market is perceived as frothier, with higher valuations relative to revenue compared to the U.S. market.
Systemic Risks and Financialization
Beyond technological and demand-side factors, systemic risks associated with financialization are also a concern. The increasing involvement of various capital providers, including pension funds and insurance companies, in funding AI infrastructure through complex financial instruments, mirrors patterns seen in previous market bubbles.
Ray Dalio, an expert on market cycles, highlights the risk of highly leveraged actors and the potential for exotic financial structures to amplify downturns. While Azhar believes the U.S. financial markets are functioning as intended by facilitating access to this theme, he acknowledges that the increasing complexity of capital structures adds to systemic risk.
The transition from private equity to public markets for major AI companies like OpenAI and Anthropic could also introduce volatility. As these companies seek to convert their "value" into "money" through stock offerings, market dynamics and investor sentiment will play a crucial role.
The Future of AI and Human Cognition
Azhar remains cautiously optimistic that a full-blown bubble, akin to the dot-com crash, is not imminent. He emphasizes the importance of real revenue from real customers as the most critical indicator for business sustainability.
Looking ahead, Azhar advocates for investment in research aimed at enhancing human thinking capabilities in the age of AI. He stresses the need to ensure that AI serves as a tool to augment, rather than replace, human cognition, preventing a decline in critical thinking faculties.
The conversation concludes with an acknowledgment that the AI landscape is dynamic and uncertain. While the underlying technology is powerful, its ultimate economic impact and the potential for market corrections remain subjects of ongoing analysis and debate.
Introduction: Is AI in a Bubble?
Nicholas Thompson introduces the topic of a potential AI bubble, likening the current situation to a plane pushed to its limits. He questions whether this is irrational exuberance or a rational response to high demand, setting the stage for a deep dive into the AI economy.
- The AI market is experiencing rapid growth and high valuations.
- There's a debate on whether this growth is sustainable or indicative of a bubble.
- Factors like circular financing, technological progress, and competition are key to understanding the market's stability.
Mapping the AI Economy: Methodology and Spending
Azeem Azhar details his methodology for mapping the AI economy, which involves forensic accounting to track actual spending and value creation, going beyond public disclosures. He estimates that approximately $126 billion was spent on AI services in the 12 months to July 2026, with a focus on de-duplicating spending to avoid double-counting.
- Azhar's research uses forensic accounting to track AI spending.
- The methodology aims to de-duplicate spending across companies.
- Approximately $126 billion was spent on AI services in the 12 months to July 2026 (outside China).
AI Infrastructure Investment and Depreciation
The discussion shifts to capital expenditures (CapEx) for AI infrastructure. While overall CapEx is in the hundreds of billions, the AI-specific component is growing rapidly, projected to exceed $100 billion by 2026. The core question is whether future revenues will justify these investments, considering depreciation and operating expenses.
- AI-specific capital expenditures are projected to exceed $100 billion by 2026.
- The profitability of AI investments depends on future revenue trajectories.
- The lifetime of AI hardware (e.g., chips) and its depreciation schedule are critical factors.
Depreciation Schedules and Real-World Usage
Azhar explains the choice of a six-year depreciation period for AI chips, citing data from companies like Amazon and Google that show chips remaining in use for longer. He acknowledges that a shorter period (e.g., four years) would make the economic model less favorable, but argues that real-world usage data supports the longer timeframe.
- A six-year depreciation period for AI chips is used, based on observed usage.
- Companies like Amazon and Google report chips in use for 6-9 years.
- The economic viability of AI investments is sensitive to depreciation assumptions.
Demand for AI: Productivity Gains and Real-World Impact
The conversation addresses the demand side of AI. While many companies are investing, the actual productivity gains reported by CTOs are often low. Azhar presents a nuanced view, citing examples of significant AI impact (like the Veterans Administration case) alongside the risk that slow adoption or performative AI use could undermine demand.
- Many companies report low productivity boosts from AI adoption.
- There's a risk that demand may not sustain current investment levels if productivity gains are not realized.
- Some specific AI applications, like in government agencies, show significant impact.
Open Source vs. Proprietary AI Models
The role of open-source models versus proprietary ones is explored. While open-source options might put downward pressure on prices, Azhar argues this could expand the market by making AI accessible for more tasks. He believes hyperscalers will still benefit as open-source models run on their infrastructure.
- Open-source AI models may reduce costs and expand market adoption.
- Hyperscalers (cloud providers) are likely to benefit regardless of model type.
- Proprietary AI companies (OpenAI, Anthropic) face pressure to cut prices and adapt.
AI Competition: The Role of China
The discussion turns to China's AI capabilities and competition. Despite facing compute limitations due to export controls, Chinese companies are highly efficient. Azhar notes that while the US AI market appears more sober, China's market exhibits higher valuations and frothiness.
- Chinese AI companies are highly efficient due to compute limitations.
- US export controls have spurred China's domestic semiconductor development.
- The Chinese AI market shows higher valuations and more frothiness compared to the US market.
Financial Fragility and Systemic Risk
The potential for financial fragility and irrational actors is examined. While Azhar believes the financial markets are working as intended by incorporating various capital sources, he acknowledges the increasing risk from complex financial structures. He also touches on the 'too big to fail' aspect of major AI companies.
- Complex financial structures are emerging to fund AI expansion.
- Increased participation from diverse capital sources adds systemic risk.
- The 'too big to fail' nature of major AI players is a consideration.
Value vs. Money: IPOs and Market Stability
The conversation explores the 'value vs. money' concept, particularly concerning the IPO potential of companies like OpenAI and Anthropic. Azhar suggests that while localized market effects might occur, the overall US economy is large enough to absorb such events without systemic collapse, provided AI companies demonstrate real revenue growth.
- The transition from private equity to public markets for AI companies presents a risk.
- The large size of the US economy can absorb localized financial shocks.
- Demonstrable revenue growth is crucial for market stability.
Ethical Concerns: AI Consciousness and Belief Systems
Azhar expresses discomfort with the quasi-religious beliefs held by some AI leaders, particularly regarding consciousness and sentience. He argues for a focus on AI as a tool and suggests that the pursuit of artificial consciousness raises significant ethical and regulatory questions, potentially requiring public intervention.
- Some AI leaders hold quasi-religious beliefs about AI's potential (e.g., consciousness).
- The pursuit of AI consciousness raises ethical and moral concerns.
- Azhar advocates for AI to remain a tool, not a sentient entity.
Key Variables and Final Thoughts on the Bubble
Reflecting on the various hypotheses, Azhar identifies Ray Dalio's concerns about market dynamics as the most thought-provoking. He emphasizes that real revenue from real customers is the most critical variable for assessing the AI economy's health, more so than any single indicator.
- Ray Dalio's market observations are considered highly influential.
- The most important variable to watch is real revenue from actual customers.
- The AI economy is complex, requiring multiple indicators for analysis.
Investing in Human Cognition: A Future Vision
Azhar concludes by stating that if he had unlimited resources, he would invest in research aimed at enhancing human cognitive abilities, ensuring that AI development leads to humans becoming cognitively stronger, not weaker. He envisions a 'cognitive gym' to foster critical thinking alongside AI use.
- Azhar would invest in enhancing human cognitive abilities with unlimited resources.
- The goal is to ensure AI makes humans cognitively stronger, not weaker.
- A 'cognitive gym' concept is proposed to foster critical thinking alongside AI use.