The ‘AI Bubble Warning Bubble’ Has Grown Bigger Than the Bubble Itself

The wave of astronomical capital expenditure on AI data centers triggered by the ChatGPT moment in 2023 turned Nvidia (NVDA)—which just a few years earlier had been seen as a company that made graphics cards for gaming—into the most valuable company in the world.

The process, however, has not always been smooth. As explosive demand repeatedly took the ChatGPT service offline, Nvidia and related stocks would surge over and over, only to plunge whenever skepticism such as “ChatGPT doesn’t have human-like intelligence; it’s just a dumb chatbot that’s good with words.” once again dominated the headlines. Three years on, usage of AI—here meaning LLMs based on transformer architectures—has grown far beyond earlier expectations, and companies are preparing to deploy it much more aggressively.

We have pulled together a selection of articles on AI that we published over the past year. At the end of 2024, when ASML, the EUV lithography equipment maker, cut its revenue guidance, AI-related stocks plunged in tandem. We argued that this was not due to a fall in GPU demand, but rather a normalization of the reckless overinvestment in capacity by Intel (INTC) and Samsung Electronics. In February this year, when China’s DeepSeek V3 model managed to deliver performance on par with state-of-the-art (SotA) models far more efficiently, concerns mounted that “we no longer need to invest in GPUs.” In our article on Google Cloud’s earnings miss, we invoked Jevons’ Paradox and predicted that lower inference costs would lead to much greater demand—and this was later borne out as inference-focused models gained popularity and GPU demand in fact surged. Finally, in August this year, AWARE warned that, due to physical constraints, the steep share-price ascent Nvidia had enjoyed over the previous three years was likely to slow, because a company’s revenue growth acceleration tends to be inversely related to its size.

Since November, the “circular financing structures” of AI companies and their use of debt to fund AI/GPU data center investment have become a flashpoint of debate. We had already covered this topic in a research article on Oracle published in September.

In that article, we highlighted that most of the increase in Oracle’s remaining performance obligations (RPO) was coming from a single customer, OpenAI, and pointed out that OpenAI’s revenue would need to grow more than fivefold within five years for it to comfortably shoulder the costs. We also noted that, unlike incumbent hyperscalers such as Microsoft and Alphabet (Google), Oracle could not fund its capex within operating cash flow, and therefore was likely to tap the corporate bond market in size in the near term—a prediction that came true when it issued USD 18 billion of bonds at the end of September. Meta subsequently raised USD 27 billion through a partnership with private credit firm Blue Owl Capital, reinforcing our narrative that we had entered a “GPU investment bubble fueled by debt financing.” Major AI stocks then suffered steep declines from their peaks, and are now in a stabilization phase.

What I want to emphasize here is not that “our calls were all spot on,” but rather that the potential vulnerabilities that had been clearly visible and transparent for months are now at the core of the narrative around an AI bubble.

From a theoretical standpoint, it is impossible to prepare for what investors call a “black swan” event. A black swan refers to an unexpected negative event, and by definition, anything that has already surfaced in news articles or in the market’s consciousness no longer qualifies. In English, such events are referred to as “known unknowns,” a term that aptly captures today’s environment, where we cannot know exactly how the AI bubble will play out, but awareness of the issue is more than sufficient.

As the article above shows, the bond market is already very actively pricing in these known unknowns. Meta’s Blue Owl Capital deal to fund AI infrastructure, along with a combined USD 90 billion of debt issuance by Alphabet (Google) and AWS, sent Oracle’s CDS premium soaring more than threefold. Even the corporate bonds of Meta, Alphabet, and Amazon—companies with near-flawless credit ratings—could only be fully placed by offering higher-than-expected yields. At the same time, in tandem with the sell-off in AI-related stocks, data-center-backed bonds also traded at discounts to par.

All of this suggests that the market is engaging very actively in price discovery to gauge the extent of the risk of AI overinvestment. It is now time to ask whether the “bubble of warnings about an AI bubble collapse” has grown larger than the actual fallout from any AI bubble itself.