I. Introduction

Nearly every week, we see front-page news of $100 billion and $500 billion AI deals, and AI investment being the core contributor to why we are not in a recession (“How Market”). These valuations and deals are predicated on very high expectations of growth in the next 5 years (Statista), with companies aiming to build intelligence that outperform humans at most economically valuable work (Ghosh). But what if those expectations aren’t met? An analysis of the industry’s structure suggests that AI valuations are highly fragile to the failure to meet these immense performance and cost expectations.

This fragility comes from “circular revenue,” where major players effectively finance their own customers. For instance, chipmaker Nvidia invests billions into datacenter provider CoreWeave, which uses that same money to buy Nvidia’s chips (Weil). This creates a loop where investments look like real revenue, which artificially inflate the industry’s growth. Beyond this financial entanglement, the industry is also structurally vulnerable because it forms a dependent stack: Nvidia at the top, model providers like OpenAI in the middle, and the application layer––companies like Sierra that turn these models into specialized services––at the bottom. The main risk is that model providers may not deliver models advanced or cheap enough to justify the immense investment it takes to hold up this stack. This could cause a cascading failure –– as even profitable downstream companies are exposed to these systemic risks.

This risk is not abstract. Analysts at Morgan Stanley, for instance, point to a “$1.5 trillion gap” that the industry will need to fill using debt markets and other financing. The spending pace has also been compared to funding “a new Apollo program… every 10 months” (Thompson). This immense investment is happening while core companies like OpenAI may still be losing money on their core product (Zitron), widening the gap between current revenue and future expectations.

II. Literature Review

Writing in The Exponential View, Azhar and Warren use five gauges to identify an AI bubble, defined as a period of rapid investment followed by a 50% drop. Historically, bubbles form when infrastructure investment grows much faster than the value it creates––like the billions spent laying unused internet cables during the telecom bubble. Similarly, Azhar and Warren find the AI industry is spending six times more on datacenter infrastructure than the revenue that infrastructure generates. This gap is even more alarming when you consider it measures revenue, not profit. Since companies are losing money on every dollar they earn (Zitron), the true deficit is massive, making it even more difficult to recoup their costs.

Beyond high spending, the industry’s structure also presents risks. The Wall Street Journal shares concerns that AI companies have numerous entangled relationships. Out of an abundance of examples, “Nvidia owns about 5% of CoreWeave and sells chips to CoreWeave [which manages datacenters]. Nvidia also committed to purchase any unsold cloud-computing capacity from CoreWeave through 2032, effectively backstopping its customer.” This means if CoreWeave can’t find customers to buy their chip services, Nvidia will buy it from them. This entanglement may give investors a false impression of chip demand. In addition, this circular flow could cause a cascading collapse if a company can’t pay back its loan, exits the agreement, or fails to find enough demand.

This leads to complicated financial arrangements like vendor financing, where a seller lends money to a buyer just so the buyer can purchase their goods. As Peter Wildeford explains, this helps sellers “move inventory that might otherwise sit unsold” and lets buyers buy chips they “couldn’t otherwise afford.” However, this falls apart if the buyer can’t pay back the money and the seller loses both the loaned money and the product itself. This is particularly dangerous for Nvidia. Since most of its business comes from AI datacenters, financing its own customers increases its exposure to AI, which makes its survival even more dependent on AI’s success.

Noah Smith suggests that while these circular investments might look like diversification among AI, they are all bets on the same outcome: that AI will meet very high expectations. Because the industry is so entangled, Smith warns, “If AI has a big crash, every single company in the circular deal diagram is going to be either toast or in huge trouble, whether or not they have circular deals in place.”

III. Methods

To evaluate the structural fragility of the AI industry, this research uses narrative analysis to examine the expectations driving current valuations. Since AI stock prices are driven by future expectations, analyzing the information diet of the market is crucial. Thus, the secondary research focuses on the dominant narratives in front-page reporting from financial publications, which represent the consensus view and public information available to investors.

I contrasted this mainstream reporting with independent analysis from technical and economic writers on platforms like Substack. These sources were chosen because of their technical counterarguments that mainstream general reporting doesn’t include. By comparing the broad market sentiment in headline news with the structural mechanics detailed by independent analysts, this research aims to identify the gap between investor behavior and the underlying fundamentals.

Finally, to provide an on-the-ground view of companies trying to create value downstream of AI providers, this research includes primary data from an email interview. The subject, an employee at Sierra (a company building customer service agents using OpenAI models), provided details about their profit margins, dependence on model providers, and the decision to build vs buy models. This serves as a case study to examine the “dependent stack” hypothesis from a company in the applications layer.

IV. Results and Discussion

Proponents of the AI boom argue that megacap companies like Google or Meta have strong cash flows from their existing ads business, contrasting them from the poor startups during the dot-com boom. However, these big companies are spending enormous capital funded by debt. Reports indicate that integrating AI into businesses is proving less useful than expected (Lohr). The actual value businesses are getting is mismatched with the expectations of how much value AI will create.

Furthermore, people argue that chips and datacenter infrastructure have real value even if the bubble pops, just like how railroads and fiber optics were later used. However, chips have a lifespan of 1-3 years rather than the 5-6 years that many believe (Kshirsagar). Since electricity and space is a major cost of running a datacenter, datacenters are unlikely to hold on to older chips that are less performant and energy efficient. Thus, it is expensive to keep older chips running when there are chips with better performance per dollar to put in the datacenter instead. If older chips are not used after three years, then those chips stop generating revenue.

This depreciation risk is exacerbated by datacenters not having enough electricity. While chip demand may seem high, the electricity required to power them is not enough. Microsoft CEO Satya Nadella recently admitted that the company has “a bunch of chips sitting in inventory” that they “can’t plug in” because they lack the power capacity (Gerstner). This confirms that the real bottleneck is getting electricity to datacenters, not the chips. As a result, chips are sitting in warehouses depreciating instead of generating revenue.

The fragility of this stack is most apparent in the application layer, as this is where most of the value is expected to be created. In my interview with a Sierra employee, he said customers are already seeing value, but “a loss of inference providers and an increase in cost does directly hurt Sierra’s bottom line.” He further noted that “if the cost of inference goes up a lot, Sierra would likely explore… control[ling] inference directly” in anticipation of limitations at OpenAI. Here, inference costs means the cost it takes to run an AI model.

These answers confirm the concern of the “dependent stack,” as Sierra’s profitability is directly linked to OpenAI’s business model, and chained to the capital expenditure of Nvidia chips. Compared to traditional Software as a Service companies, adding a new customer costs nothing. But for Sierra, every new customer adds inference costs that Sierra must pay to OpenAI. Sierra’s profit margins also decrease if OpenAI raises inference prices. So a price hike at the top (chips) gets passed to the middle (OpenAI), which then gets passed to the bottom (Sierra), who must either absorb the cost, destroying their margins, or pass it to customers who may not be willing to pay.

The employee’s response that Sierra would “explore… control[ling] inference directly,” suggests that the application layer may abandon the model provider if costs are too high. By hedging against a provider’s collapse or unbearable price increases, companies are questioning the long-term stability of the current paradigm. This undermines the valuation of the entire stack: if Sierra doesn’t get enough value from OpenAI, then the revenue growth expected for OpenAI, and thus Nvidia, may never come.

The market’s reaction to technical developments supports this volatility. In January 2025, Chinese model provider DeepSeek released a technical paper on how they made their newest model with less compute and capital than US equivalents. Initially, there was no market reaction. However, weeks later, after DeepSeek hit #1 in the App Store, investors panicked and realized the amount of compute and capital needed might be lower than anticipated (Reinicke). The main reason why the valuations of OpenAI and Google are so high is because it is expected that more capital expenditures correspond to better models. But if a Chinese company can build a model for way less, then US companies’ valuations are not justified.

The market not dropping until weeks after the information was publicly known suggests investors are not analyzing the technology for what it is, but more on the current narrative and sentiment. If valuations are tied to sentiment rather than reality, the industry is susceptible to crashes.

V. Conclusion

The research indicates that AI valuations are highly fragile because they are based on exceptionally high expectations. The literature review showed this fragility comes from massive, speculative spending and a deeply entangled, circular financial structure. While the technology may well be transformative, the economic scaffolding supporting it is brittle. The primary research confirms that even a profitable downstream company is vulnerable to rising costs and loss of industry confidence because of its reliance on upstream model providers.

Ultimately, the circular revenue and high expectations make the industry very correlated, so one misstep in an area could collapse the entire stack.

Works Cited