If the recent wave of TV commercials promoting artificial intelligence gave you a sense of déjà vu, you are not alone. Dot-com start-ups in 2000 and cryptocurrency companies in 2022 dominated TV ads shortly before both suffered epic crashes. This year’s ads aired amid a sharp increase in investments, increasingly fueled by debt. Alphabet has raised over $30 billion in USD-equivalent debt, including a 100-year bond, a rare structure not seen since 1997 when Motorola issued one.
Given the large amount of money flowing into AI, it’s natural to wonder if today’s investments signal a speculative bubble. It’s difficult to assess these moments in real time, and even when bubbles burst, the underlying technologies can ultimately change the world. Here are six areas we are watching:
1. Bubbling IPOs
This year, high-profile IPOs are expected from leading companies in the generative AI field. Among those reportedly considering their options: Anthropic, OpenAI’s ChatGPT producer, and SpaceX, which recently merged with xAI. This is a logical step that has so far been missing in the AI boom. The anticipation is accompanied by concerns about valuations, circular financing, and these companies’ ability to meet investor expectations. One of the elements that inflated and sustained the tech bubble of the ’90s was accelerating revenue growth, with promises of future profitability. These pre-IPO companies are the modern equivalent. When they go public and investors can get a more detailed view of their financial data, high growth rates will likely be rewarded. Investors are particularly eager to track OpenAI’s earnings growth. The company has committed $1.3 trillion in purchases through 2031 from suppliers such as Oracle, CoreWeave, Microsoft, Amazon, NVIDIA, and others. This concentration has made OpenAI a benchmark for AI growth. As long as growth is sustained, OpenAI and other AI leaders will likely continue to have funding. Sooner or later things will slow down and the environment will become more challenging, but we are not there yet, as we are still at the beginning of adoption and monetization curves.
2. Debt-fueled growth
Alphabet’s 100-year bonds highlight the growing trend of hyperscalers relying more on debt, even though commercial returns remain uncertain. AI-related debt issuance increased 112% in 2025 compared to the previous year, and 2026 looks set to be even stronger. Although the size of today’s issuances is making headlines, these are high-quality companies representing a small part of the investment-grade debt market, in stark contrast to their dominant position in the S&P 500 index. Most of these companies currently have low debt levels and are issuing debt to finance AI-related capital expenditures (capex), which helps optimize their capital structure. Given their substantial cash reserves and healthy cash flows, they can probably finance these projects on their own, even considering increased capital spending. In our view, this greatly reduces systemic risks.
We believe this is in stark contrast to dot-com era companies. Many tech companies in the late ’90s operated with limited or even negative cash flows, relying heavily on issuing shares and more speculative venture capital. Companies like WorldCom accumulated debt and significantly increased leverage to build their fiber optic networks, while Pets.com raised large sums despite unproven demand.
It is important to note that many current investment-grade debt issuances have been made at the parent company level, which offers many advantages. The main one is that their value is tied to the collective cash flows and value of the enterprise. For example, Alphabet is the parent company of Google, YouTube, Waymo, DeepMind, and other subsidiaries. This is an important distinction because a loan is not granted to a structure that exists solely to finance AI investments. However, investors have demanded an additional yield for holding AI-linked bonds compared to those with similar ratings, a premium reflecting the high volume of bonds issued, the issuer’s slightly higher leverage, and uncertainty about whether AI demand will continue at the current pace.
3. Creative financing
Another concern is so-called vendor financing. Money circulates among the same companies, with start-ups and hyperscalers buying from each other and helping each other increase revenues. A fitting example: Amazon and Google have invested billions in Anthropic, an AI systems start-up. In return, Anthropic agreed to use Amazon Web Services and Google’s services and products. In the ’90s, similar circular arrangements occurred with Lucent Technologies, which extended excessive loans to cash-strapped start-ups so they could buy Lucent’s equipment. Those customers ultimately could not pay, forcing Lucent to restate revenues and take huge write-downs.
We consider a short-term scenario similar to a house of cards unlikely for hyperscalers. Unlike Lucent, they lend only a small portion of their cash flows. Their financial strength generally gives them the flexibility to pursue alternative financing methods for their expansion plans, which may include off-balance-sheet arrangements or project finance operations. Meta, for example, has a joint venture with Blue Owl Capital called Beignet Investor to build a large data center in Louisiana called Hyperion. Microsoft, meanwhile, has signed short-term agreements with data center providers called neoclouds, which are considered operating expenses rather than long-term capital investments.
Since AI development is still considered in its early stages, these non-traditional arrangements will likely increase over the next year, particularly with private credit. They can be interesting in some cases but require further analysis by potential financiers because they are structured to limit financial risks for the parent company. While we believe in the transformative power of this technology, we are in no rush to invest in these arrangements. Everything will depend on the individual structure and contract terms, including an assessment of the financial support provided by the hyperscalers.
4. Overbuilding
If you build it, according to this logic, growth will follow. In the early 2000s, telecom companies invested billions in fiber optic cable networks, convinced that internet data transmission demand was unlimited. What happened instead was an oversupply that led to massive asset write-downs and losses for investors.
It is important to remember that overinvestment is a feature, not a bug, of every major technological advance. At some point, companies will shift their focus to more efficient investments. Today, hyperscalers believe building more data centers is essential to expanding AI inference, i.e., the ability to run generative AI models for everyday use. Infrastructure is needed to handle increased AI workloads related to training and inference, with the latter requiring reliable, always-on computing to serve users in real time. From the AI workload perspective, physical infrastructure must be built before it can be populated with graphics processing units (GPUs), networks, and storage to run the next generation of large language models. Therefore, investors pay close attention to the performance of new AI models and their updates, they add. If gains begin to stabilize, that could signal AI demand may not keep pace with spending. However, we believe the current hyperscaler capacity to be built over the next two years could be repurposed in other sectors if scalability laws do not hold. For some companies, there will be demand for that computing power even if AI demand slows.
5. Resource constraints
Electricity availability has become an urgent issue for AI’s growth potential. This is because data centers require memory, power, chips, copper, and water. Bottlenecks in these areas could impact infrastructure development, slowing hyperscaler capital spending and putting pressure on development timelines. The main bottleneck is the shortage of skilled workers able to build new power plants and transmission lines needed to connect new data centers to the grid. For now, the United States appears to have a sufficient electricity surplus to meet growing demand through 2028 or 2029. After that, without a significant increase in electricity production, data center growth rates could slow. Utility companies have started investing more in their electric grid due to aging infrastructure and rising demand. Overall utility capital spending could rise from about $130–135 billion in 2020 to around $200 billion, with as much as 20% of that amount related to AI. Since utilities want to maintain their investment-grade rating, we expect many of them to resort to equity or equity-linked financing to fund investments.
6. Economic growth slowdown
It is impossible to talk about a potential AI bubble without understanding where we are in the economic cycle. The hallmarks of late-cycle phases typically include sustained increases in inflation, interest rates, wages, and other conditions that push the Federal Reserve to tighten monetary policy. We believe the U.S. economy is currently in the mid-cycle phase, a generally resilient phase. We forecast long-term U.S. productivity growth to rise to 3–4%, which could keep wage increases elevated and support broader economic growth.
Of course, there is always the possibility that other external shocks cause a drastic drop in AI-related stocks, such as innovations that substantially reduce training costs or investors withdrawing from data center financing. In our view, we are in the early phase of AI technology development, with demand continuing to outpace supply.
Over time, competition for AI dominance will produce winners and losers, but current tech hyperscalers generate substantial earnings and free cash flow; therefore, a large speculative bubble comparable to the dot-com era is less likely. Every cycle will have some level of excess and likely moments of volatility, but this one seems more likely to grow.
