Fears of an AI bubble have picked up steam in recent months. Some point to surging AI capital expenditure that lacks corresponding immediate returns. Others have raised red flags about the influx of circular deals between OpenAI and the likes of Nvidia, Microsoft and Oracle, which could be artificially inflating demand. Still others question the sky-high valuations of AI-linked stocks, despite the recent market pullback.

Are we in an AI bubble reminiscent of the dotcom era? If so, what are some signs that it may be about to pop? INSEAD faculty analyse how we got here, whether investors have cause for concern and what to pay attention to in the months ahead.

The market is pricing in continued exceptional growth

Ben Charoenwong, Associate Professor of Finance

The question of whether we are in an AI bubble requires some definition. A bubble exists not only when prices exceed current fundamentals, but when prices exceed what future fundamentals can realistically deliver. It implies that the market is behaving like expected future returns are negative, which can happen when the market is “wrong” or when risk-taking capacity turns from risk-aversion to risk-lovingness – distinguishing one from the other is notoriously difficult.

So, bubbles can form when market participants unrealistically extrapolate recent growth rates to the indefinite future, either by ignoring competitive dynamics or when circular validation (rising prices confirming the narrative that justified the price increase) crowds out rigorous analyses of sustainable cash flows.

By this framework, the current AI market is not obviously a bubble – yet. Earnings growth has largely matched price increases for AI infrastructure leaders, but the market is pricing in continued exceptional growth for years to come. Whether those expectations prove realistic or represent overextrapolation will determine if current valuations are justified or become the foundation of a correction.

There are warning signs. The web of circular financing deals in AI has reached a scale and complexity that warrants serious scrutiny. These arrangements bear uncomfortable similarities to the vendor-financing structures that characterised the late-stage dotcom bubble. Consider the arrangement between Nvidia and OpenAI. Nvidia invests up to US$100 billion in OpenAI. OpenAI uses that capital to build data centres, which are filled with Nvidia chips. OpenAI receives cash to expand, and Nvidia is guaranteed to be the supplier for that expansion. The arrangement between Oracle and OpenAI follows the same pattern.

The distinction between "flywheel" and "house of cards" comes down to whether real end-user demand is being generated, or whether money is simply moving in circles. Bull markets and elevated sentiment are forgiving of circular financing in hopes of future growth. But if demand fails to materialise in earnings, these hopes can be dashed. What appears as a virtuous cycle today becomes the mechanism of collapse tomorrow.

While markets have focused on generative AI and cloud infrastructure, the next phase of the AI cycle may be physical, comprising humanoid robotics and embodied intelligence. Goldman Sachs projects that the global market for humanoid robots could reach US$38 billion by 2035. 

This wave is beginning to materialise commercially but seems less of a factor in current valuations compared to the AI darlings – and could be a significant upcoming catalyst for capital investment beyond AI data centres. Investors focused solely on generative AI may be underweighting the longer-term transformation that physical AI represents.

Investors are already asking firms to “show me”

Lily Fang, Dean of Research and Innovation, Professor of Finance and the UBS Chair in Investment Banking

It’s always dangerous to say, “this time is different”. From a pure valuation perspective, there are some parallels between current market conditions and the dotcom bubble. For example, if you look at the Shiller P/E ratio, which is the ratio of the current price over the trailing 10-year inflation-adjusted earnings of the S&P 500, we are at a level (40) that’s very close to the peak in 1999 (45).

But there are some fundamental differences. In the dotcom era, the most expensive stocks were loss-making, newly listed tech stocks that drove up overall market valuation. Today, valuation is propelled by massive – and massively profitable – firms such as Nvidia, Google and their Big Tech peers. These firms’ stock prices have risen a lot, but so have their earnings. Based on the Shiller P/E ratio, they look expensive. 

However, if you look at the regular P/E ratio of the S&P 500, which is calculated as the current price divided by only the trailing 12-month earnings, then we are at 29 – quite a bit lower than the peak of 45. Though 29 is hardly cheap, it indicates that current earnings are significantly higher than historical earnings, so valuations regarding current and future earnings are more reasonable compared to the dotcom era.

That being said, the market wants assurances that its investments are yielding the desired ROI. Investors are already asking firms to “show me”. Recently, both Meta and Microsoft reported top- and bottom-line earnings that beat expectations. But Microsoft’s stock price dropped by 10 percent, while Meta’s rose by 10 percent the next day. 

Why? Meta was able to convince investors that AI was improving its ad targeting and profitability, whereas Microsoft couldn’t show a clear ROI. Let’s not forget that not too long ago, the market was unforgiving to Meta. In the previous quarter, when it couldn’t show such data, its stock price was duly punished.

I think the bubble is less in the listed market and more in pre-IPO AI start-ups. Many early-stage investors, including venture capitalists, will lose money if they continue pouring capital into loss-making, cash-burning AI start-ups with no clear path to profitability. In this sense, there’s a parallel to the dotcom bubble: countless VC firms disappeared after it burst. 

The valuations of AI firms are eye-watering

Boris Vallée, Associate Professor of Finance

I see many similarities between the current AI boom and the dotcom era. First, there’s no question that we are facing an innovation that will have a profound impact on our economies and societies, much like the internet. Contrast this with the crypto ecosystem, where compelling use cases have been challenging to find. 

Second, the valuations of AI companies are eye-watering and have risen very quickly, despite the absence of profit for most of these firms. This is reminiscent of the dotcom era. On the other hand, Nvidia is a highly profitable company, which reminds me of the gold rush in some ways – shovel-sellers were the ones with the best business case, given the uncertainty of gold prospection.

There’s a key difference between the hundreds of billions that companies are pouring into AI-related capex and the spending that occurred during the dotcom era. Large and extraordinarily deep-pocketed incumbents are participating in and financing the development of AI technologies and capacity, either on their own or by partnering with the most innovative firms. This capex is mostly being shelled out on infrastructure, which should be largely redeployable. This is very different from the dotcom bubble, during which marketing spend, for instance, was particularly large.

General academic research has shown that complex financing structures often serve the purpose of hiding risk. My understanding is that the circular transactions we’re seeing aren’t simply providing financing across the supply chain (e.g. a client financing a supplier) but are structured in opaque ways to limit recourse over the general assets of the large tech firms. There are some concerns here that deep-pocketed incumbents are trying to design options for themselves, thereby fuelling the financing of investments while limiting their own downside exposure. So, if things go awry, the losses would be transferred to the financial system.

Looking ahead, I’d personally keep an eye on adoption trends by both corporate clients and individuals and the associated revenues, to monitor their willingness to pay. We’ll need to be able to start rationalising both valuations and capex with more precisely estimated cash flows. 

An interesting parallel is that Nvidia’s market cap, at roughly US$4 trillion, is as large as the entire amount of PE assets under management in the United States. When we think about the footprint of firms under PE ownership, we have to start seeing AI building a similar one.

“Strategic necessity” isn’t a blank cheque 

Lin Shen, Assistant Professor of Finance

On valuation alone, today’s AI-led market looks hot, but it still doesn’t match dotcom extremes. The current rally is more tightly anchored to profits and cash flow. The Nasdaq-100 is currently trading at a trailing 12-month price-to-earnings multiple of just over 33 vs. around 60 in March 2000.

The dotcom peak was more extreme because a large share of market value was tied to companies that hadn’t yet built real earnings or cash flow, so prices were driven more by future stories than by current results. In the current AI cycle, many of the key beneficiaries are cash-flow machines. 

Nvidia – arguably the emblematic AI winner – reported a record US$57 billion in revenue for its fiscal Q3 2025, up 62 percent year-over-year, and US$23.8 billion in cash flow from operating activities for that same quarter. These numbers make it easier to justify paying up via discounted cash-flow logic, not just “eyeballs”.

That same focus on fundamentals emerges in how quickly the market pushes back when the cash-flow story gets stretched by spending. For instance, when Alphabet declared it would target US$175 billion to US$185 billion of capex in 2026 (roughly double 2025’s level), it triggered an immediate sell-off as investors questioned whether AI spend would convert into earnings.

Alphabet isn’t alone. The scale of the arms race is now explicit: Big Tech players are estimated to shell out US$660 billion on AI this year, a number that crystallises both the ambition and the risk. The market’s reaction to these capex disclosures – sharp repricing rather than applause – underscores how different today’s mindset is from the late-1990s: Investors aren’t blindly buying a vision; they are policing the cash conversion path.

So, why are firms spending so aggressively anyway? Because AI is a “winner-takes-most” game: distribution, proprietary data, model quality and developer ecosystems can compound, and the prize for becoming the default platform is enormous. That belief turns AI into an arms race, with each firm fearing that under-investing today risks irrelevance tomorrow.

But the capex-driven sell-offs are a reminder that “strategic necessity” isn’t a blank cheque. The past couple of years were largely about building capacity. Going forward, investors will demand proof of utilisation, pricing power and profits.