The Schrödinger bubble: the main risk of AI stocks may not be where it is sought

Artificial intelligence has become the main investment story of recent years. The growth in demand for computing power has led to a sharp increase in revenue for chip makers, expansion of cloud infrastructure and large-scale investments by major technology companies. Against this backdrop, the current boom is increasingly being called the “AI bubble.” But it seems that this phrase is meaningless. In some American “AI-shares” are already visible signs of strong overheating, but mania of the scale of the late 1990s is not yet visible. To understand this, we need to answer three questions: what expectations are already embedded in stock prices, how much the current public market is similar to the dot-com market, and whether the new infrastructure will provide an acceptable return on invested capital.

The prices of “AI-shares” laid very different expectations

The main difference between the current boom and the late 1990s is the scale and profitability of its largest participants.

In the late 1990s, the market often capitalized on future internet businesses long before there was a steady profit margin. Companies with real sales existed, too, but the gap between the current size of the business and the future reflected in the stock price was huge. Today, the largest market participants have a much smaller gap.

The following examples can be cited: NVIDIA in the second quarter of fiscal year 2027, ended July 26, 2026, increased revenue by 106% year-on-year, to $96.2 billion, with a gross margin of 75%. Broadcom received $29.6 billion in revenue and $13.7 billion in free cash flow in the third quarter of fiscal year 2026. AMD’s data center revenue grew 107%.

Current players have more opportunities to monetize the new technology: Microsoft, Amazon and Alphabet are developing AI on top of existing cloud, advertising and software platforms and immediately access existing customers. Therefore, a simple comparison of today’s multiples with 1999-2000 shows little. It is much more useful to estimate how much growth the current value of companies requires.

To do this, it was decided to use the “reverse” DCF method – calculating from the current value of the company to the business parameters that explain it. The calculation is based on quotations as of September 30, 2026, based on estimates of financial indicators and assumptions about the long-term margin of free cash flow.

The range of expectations was very large. For NVIDIA and Broadcom, the current price suggests revenue growth of about 13% on average and 16% per year through 2035. The market effectively requires both companies to remain businesses of enormous scale and high margins for nearly a decade to come.

For AMD, Palantir and Oracle, the requirements are significantly higher: the current price assumes revenue growth of almost 29%, 37% and 23% per year, respectively. For AMD, the market requires growth above its long-term historical pace, and for Palantir and Oracle, the price-based pace suggests a significant acceleration relative to past years.

In other words, the market puts completely different expectations in the cost of different participants in the AI chain.

This fact plays a big role due to the high concentration of the American market: as of September 30, 2026, the ten largest components of the S&P 500 accounted for just over 39% of the index. Therefore, an error in expectations for several companies with high weight can significantly affect the entire index.

AI boom significantly less speculative than dot-com bubble

The two periods can be roughly compared by which technology companies the public market is willing to finance. In 1999, there were 370 technology IPOs in the United States, and another 261 in 2000. According to IPO market researcher Jay Ritter, there were 31 technology IPOs in 2025, along with 34 direct listings.

The characteristics of companies listed on the stock exchange have also changed. Median revenue before the IPO grew from about $ 23 million in 1999-2000 – in dollars of 2024 – to $ 90 million in 2025, and the median age of the issuer – from 4-5 to 12 years. At the same time, the median price/revenue ratio decreased from 26.5–31.7 to 11.8. Speculative excitement is also noticeably weaker: in 2025, only five American IPOs doubled on the first day of trading, compared to 115 in 1999 and 78 in 2000. In 2026, individual AI companies are already entering the market with more aggressive valuations, but the mass IPO euphoria of the late 1990s is not yet visible.

The main risk is the economy of the infrastructure cycle

If the public market is not repeating 1999, where does it feel like a bubble? The main risk is seen not so much in the spread of AI itself as in the scale of the investment required to maintain it. AI requires accelerators, memory, servers, networks, data centers, cooling systems and electricity – and the cost of this infrastructure is already measured in the hundreds of billions of dollars.

Microsoft for the fiscal year 2026 ended June 30, 2026, disclosed about $145 billion in capital expenditures, a significant portion of which falls on infrastructure for AI. After the second quarter, Alphabet raised its capital expenditure forecast for 2026 to $195-205 billion, and only for the second quarter invested $44.9 billion in fixed assets. The scale of the cycle is already noticeable at the level of the US economy: according to the Federal Reserve Board of Governors, the ratio of investment in equipment and intellectual property to GDP by the first quarter of 2026 was only slightly below the high of 2000, and the pace of acceleration of investment in recent quarters approached the levels of the late 1990s.

Risk arises from the difference between the moment of investment and the moment of return. NVIDIA is getting paid for an accelerator, Broadcom is getting paid for networking equipment or a dedicated chip today. The owner of a data center or cloud platform will have to recoup these investments for many years through cloud services, enterprise software, advertising, automation and new products. At the same time, the infrastructure economy depends not only on the growth of demand for AI, but also on capacity utilization, the cost of electricity, equipment depreciation, upgrade cycles and the price of computing power itself.

So far, the demand for computing exceeds supply. Microsoft, for example, expects its cloud infrastructure to be limited by available capacity until at least the end of 2026. However, at the same time, huge amounts of new power are being introduced, accelerators become more productive, and algorithms allow more useful work to be done on the same hardware. If the supply of computing resources at some point begins to grow faster than profitable monetized demand, the cost of the unit of computing will decline, and the profitability of the infrastructure built today will be below expectations.

This is where the telecom boom of the late 1990s is useful. The companies correctly assumed that the Internet would radically increase the volume of data transmission, and invested huge funds in fiber optic networks. The mistake was not in the forecast of the development of the Internet, but in the investment economy: the supply of capacity began to grow faster than demand, capable of providing the necessary returns, data prices fell, asset utilization was below expectations, and the return on capital deteriorated. AI may well become one of the most important technologies of the next decades, and some of the infrastructure built for it will still be a bad investment.

An important difference in the current cycle is that the largest technology companies are largely financing investments through their own profitable businesses. However, even capital expenditures are already consuming cash flow. In the second quarter of 2026, Alphabet generated $39.1 billion in operating cash flow and spent $44.9 billion on capital investments, causing free cash flow to be negative at about $5.9 billion.

For companies that are more dependent on external financing, the risk is much higher. At the end of June 2026, CoreWeave had about $35.6 billion in debt, and net interest expenses for the first half of the year alone exceeded $1 billion, some of the company's bonds carry a coupon of about 9.6-9.75%. For such a business, the loading of data centers and the cost of computing power directly affect the ability to service debt.

The same risk arises for model developers. According to Reuters, which reviewed the confidential Anthropic prospectus ahead of the IPO, the company plans to spend at least $518 billion on computing infrastructure over the next decade. About 80% of these obligations are non-revocable or involve payments regardless of the actual use of capacity. A corporate customer can cut AI costs if they don’t see enough returns. The developer of a multi-year fixed commitment model has much less flexibility.

Therefore, the risk is distributed unevenly across the AI chain. Equipment manufacturers monetize the investment cycle before others. Cloud platforms, data center operators, and model developers are taking on an increasing share of long-term commitments and must meet future demand.

So where's the bubble?

At the end of September 2026, there is no visible single bubble of American “AI stocks” on the scale of the mass euphoria of 1999-2000. The largest companies have serious profits and operating businesses, technology companies enter the public market more mature and on average at significantly lower multiples, and the risk of revaluation is distributed unevenly along the AI chain. In itself, the high value of the company does not mean a bubble: the reverse estimate shows that in some cases the current capitalization is explained by quite achievable business parameters, while in others the market requires very high multi-year growth.

In this sense, the “AI bubble” really resembles Schrödinger’s cat: he is and at the same time he is not – until you open the box and see where to look. In some stocks, expectations already look excessive, while the market as a whole is far from the euphoria of the late 1990s. At the same time, the main risk is increasingly shifting to those participants in the chain, whose assessment depends on whether the hundreds of billions of dollars invested in infrastructure will pay off in the future.