The numbers that make your head spin
Here are a few figures gathered for this article. In 2025, AI startups raised close to 150 billion dollars, more than 40% of total global venture capital (venture capital refers to investments made in companies with high potential but also high risk, often in tech). The Big Tech companies (the American tech giants: Microsoft, Google, Amazon, Meta, Apple) announced 725 billion in AI investment for 2026, up 77% on 2025. OpenAI is valued at 852 billion dollars. Anthropic is targeting 900 billion. xAI (Elon Musk's company) more than 200 billion. Cursor 9.9 billion. Perplexity 20 billion.
And yet, as we documented in our article on the big failures, the overwhelming majority of consumer AI products fail. Sora was burning 15 million a day against 2.1 million in revenue. 85% of AI hardware devices launched in 2024-2025 failed. Apple paid 250 million for having sold non-existent AI features. The gap between investment figures and profitability figures is historic.
The question now haunting financial analysts is simple: are we in a bubble that is going to burst, or in an infrastructure boom that will eventually pay off? The answer is not obvious, and it matters to pretty much everyone, including those who are not investing a penny in AI.
Why some are talking about a bubble
Several converging signals feed the bubble hypothesis. First, valuations far outstrip revenues. OpenAI is valued at 852 billion against roughly 25 billion in annualised revenue (ARR: Annual Recurring Revenue) in 2026, a multiple of 34. For comparison, Apple trades at a multiple of around 8. Anthropic is valued at 380 billion against 30 billion in annualised revenue in April 2026, a multiple of 12 (already high but more reasonable).
Second, unit economics are often negative. Unit economics refers to the cost and revenue generated by a single user or customer. Sora cost more than it brought in. Even ChatGPT is under discussion: according to several independent estimates, OpenAI continues to lose money on every free user, and the margin on paying subscribers remains thin. Anthropic is more profitable, but is investing massively in the next generation of models, which keeps margins negative.
Third, the market is concentrated among a few players. 80% of the 242 billion invested in the first quarter of 2026 went to just 4 companies: OpenAI, Anthropic, xAI, Waymo (Alphabet's autonomous vehicle subsidiary, Alphabet being Google's parent company). This concentration resembles the end of the dot-com cycle (the Internet bubble of 1999-2000, when the valuations of web companies exploded before collapsing), where capital clustered around a few players before the explosion.
In 1999-2000, internet companies' valuations reached levels similar to what we see today. Pets.com, Webvan and Drkoop raised hundreds of millions before collapsing. The NASDAQ (the American stock exchange where most large tech companies are listed) lost 78% of its value between March 2000 and October 2002. Several analysts see a similar pattern today. The key difference, however, is that today's AI players already have real revenue (sometimes enormous), whereas many dot-com startups had none.
Why others talk about an infrastructure boom
The opposite argument is just as well documented. First, AI is already generating real, measurable economic value. Cursor saves hundreds of thousands of developers several hours a day. ElevenLabs (a synthetic voice generation platform) powers thousands of voice applications in production. Anthropic powers tens of thousands of enterprise applications. That value does not disappear if a bubble bursts.
Second, the infrastructure that has been built remains. Data centres, chips, open-source models and the accumulated human skills are durable assets, even if some companies go bankrupt. That is what happened after the dot-com bubble: the fibre optic cable installed in 1998-2000 fuelled growth from 2003 to 2010. Today's AI infrastructure will probably play the same role over the next decade.
Third, companies themselves are adopting AI en masse. As we showed in our article on global adoption, 16.3% of the world's population already uses generative AI regularly. In the Emirates it is 64%. In Singapore, 60.9%. This adoption does not look like a passing fad, but like a structural transformation of digital usage.
The most likely middle scenario
Between these two theses, the most likely scenario is neither a sudden explosion nor uninterrupted growth. It is a painful but sector-specific consolidation. Several categories will probably collapse over the next 18 months, while others continue to grow.
| Category | Risk 2026-2027 | Reason |
|---|---|---|
| Consumer generative apps | High | Broken economics (the Sora case) |
| Standalone AI hardware | Very high | 85% failure rate already observed |
| Foundation models (frontier) | Moderate | 3 to 4 players out of 10 will survive |
| Vertical pro tools (Cursor, etc.) | Low | Sound unit economics |
| Infrastructure (chips, cloud) | Very low | Structural demand |
| Voice/image platforms | Low | Stable platform position |
The signals to watch over the next 12 months
If you want to get ahead of the curve, here are the indicators analysts are actually watching.
Signal 1: the leaders' margins. If OpenAI, Anthropic and Google report clearly positive margins on their API (Application Programming Interface, the interface that lets other software connect to the model) in 2026-2027, the infrastructure boom hypothesis holds. If they keep losing money despite growth, the bubble scenario gains credibility.
Signal 2: valuations on new funding rounds. A "funding round" is an event where a startup raises money from investors. The slowdown or acceleration of valuations sets the tempo of the market. If Cursor raises at 30 billion in its next round, the bubble continues. If the same company has to accept a down round (a funding round at a lower valuation than the previous one), that is the signal of a reversal.
Signal 3: Anthropic's IPO planned for October 2026. An IPO (Initial Public Offering) is the moment when a private company goes public by putting its shares on the market. If Anthropic's IPO succeeds with an enthusiastic reception, the public market validates private valuations. If it disappoints or is postponed, that is probably the signal that public investors are no longer following private multiples.
Signal 4: layoffs in Big Tech. In 2025-2026, Meta announced 14,000 job cuts to fund its AI. Microsoft and Amazon are following. These layoffs can be interpreted as a sign that AI is genuinely replacing human labour (proof of real value), or as a sign that Big Tech is squeezing its costs because AI investments are not yet paying off (proof of financial strain). Both theses can coexist.
Signal 5: the behaviour of central banks. If the Fed (Federal Reserve, the central bank of the United States) or the ECB (European Central Bank) raise interest rates, the cost of capital increases, and speculative valuations adjust downwards. Monetary conditions in 2026-2027 will be decisive for the fate of AI valuations.
What this changes for you
For the general public, the main question is not so much the possible bursting of a financial bubble as the day-to-day consequences. If a bubble bursts, here is what would concretely change.
First, several services you use would disappear or change owner. That is already happening: Microsoft Designer, Sora, Humane Pin. Any worsening would affect more mainstream services.
Second, prices would rise sharply. The players that survive consolidation will have reinforced pricing power (pricing power refers to a company's ability to impose price increases without losing its customers). ChatGPT at 40 or 50 dollars a month in 2027 is not at all impossible if OpenAI comes out on top. As we analysed in our article on the freemium trap, these increases rarely happen in one brutal jump, but through an accumulation of small adjustments.
Third, open source would become more attractive. Mistral (a French lab that develops AI models), DeepSeek (a Chinese lab), Llama (Meta's model): these open-source models would be the natural alternative for developers and companies wanting to free themselves from the giants. It is probably the best hedge against bubble risk: mastering at least one open-source tool that you can run independently.
The real lesson: decouple your decisions from the bubble
Beyond the bubble-or-no-bubble debate, the practical lesson is probably the same as for the dot-com in 1999. The right behaviour is not to bet on the burst or the continuation. It is to decouple your own decisions from the general mood of the market.
If you are a user, stay in control of your workflows by avoiding excessive dependence on a single supplier. If you are an entrepreneur, as we analysed in our article on the startups that outperform, build a vertical product with a viable business model from the outset, rather than chasing funding rounds. If you are an investor or shareholder, look at real margins rather than headline valuations.
The bubble, if it exists, will not sweep everything away. It will clear out the players with no business model (defensible economic model), as previous bubbles did. The tools, services and skills that are genuinely useful will continue to exist. The whole trick is knowing how to tell one from the other in 2026, because the sorting will happen either way in the next 24 months, whether the bubble bursts or slowly deflates.
To dig deeper into this subject, see also our analysis of faked AI demos, which shows how bubbles are built through communication, and our article on the 1, 5, 10 and 30-year scenarios, which places these financial issues within a long-term trajectory.