What if the AI bubble burst?

If there's a trendy topic in tech right now, it's AI. And it's normal, given the amounts invested on one hand, the societal transformations on the other, not to mention all the issues and risks related to employment and ecology.

But there's one theme that seems to unite just about everyone, pro and anti AI, it would be a bubble, and it's going to end up bursting.

Well, now's the time to ask the question. Is it a bubble? Can it burst and when? And what would the consequences be if it did?

Is it a financial bubble?

If the question seems obvious to many, it deserves a closer look. Are we really in a financial bubble? Is the value we're giving to AI companies realistic?

Are we witnessing a historic transformation... or a speculative bubble?

First clue: valuations that seem absurd

A financial bubble isn't simply a sector that rises very quickly. Otherwise Amazon in 2010, Tesla in 2020, or Nvidia today would all have been bubbles. The central question is one of sustainability: is the value attributed to these companies realistic? Are massive investments in AI based on solid financial foundations?

At first glance, it doesn’t look good.

Anthropic generated about 850 million dollars in revenue in 2024. Yet, a few months later, the company was valued at nearly 19 billion dollars. More than twenty times its annual revenue.

You need to understand that in finance, there are two logics:

• For a traditional company, the value of a stock corresponds to known performance in the present. It depends on profits generated and physical assets (factories, stores, inventory) or intellectual assets (patents).
• For a tech startup, however, the value rests on the promise of future gains. Its valuation doesn't reflect its current revenues, often low, but anticipates their explosion to come. A "multiple" is then applied to its revenue or recurring revenue. If this multiple rarely exceeds 10 to 15 times revenue (especially since the interest rate hike post-2021), AI pioneer companies today reach levels well above that, sometimes ranging from 20 to over 50 times their revenue.

However, just because these multiples are enormous doesn't mean they're unusual and abnormal. They're abnormal for the period, but not abnormal in absolute terms. In the past, Shopify, Snowflake, Amazon, all these companies exceeded x50 and the bets turned out to be winning.

So, wrong trail.

Verdict: ❌ inconclusive

Second clue: an avalanche of capital

Over the past three years, more than 1,000 billion dollars have been committed to the AI ecosystem.

According to the BIS (Bank for International Settlements), 5% of US GDP is today linked to investments in AI. This amount exceeds what was observed during the previous internet bubble in 2000.

And in absolute value, the Internet bubble represented approximately 600 billion dollars of the time. Adjusted for inflation, we're again approaching current investments in AI.

On the other hand, by contrast, the subprime crisis involved over 10,500 billion dollars in residential mortgage debt. But be careful, we're comparing apples and oranges in this specific case.

What's certain is that the amounts at stake are colossal. But colossal doesn't mean abnormal. Telecoms in the 80s/90s, mobile in the 2000s also attracted gigantic sums without people talking about a bubble at the time.

Verdict: ❌ concerning but inconclusive

Third clue: what if the customers were... the investors?

Here we touch on one of the specifics of the potential current crisis: the revenue of AI companies increases through their own spending on AI.

AI actors aren't just companies that create LLMs. It's also cloud hyperscalers or chip manufacturers. It's a perfect form of circular economy (round tripping)

NVIDIA invests in OpenAI
        ↓
OpenAI reserves computing
        ↓
OpenAI signs cloud contracts with Oracle, Microsoft, etc.
        ↓
These providers buy NVIDIA GPUs
        ↓
Revenue comes back to NVIDIA

The problem with this circular economy is that it self-feeds and we don't yet know if real demand will follow. If the only customers for AI are the other AI players, let's just say that's at least suspicious.

All projections for new datacenters, investments to build new chips, etc... would potentially be thrown away in case of market reversal since for now the only customers for these companies, at this scale, are themselves. And the sums we're talking about are in the tens of billions.

Verdict: ⚠️ concerning

Fourth clue: non-existent profitability

Investments in AI are not profitable except for NVidia but that's a bit natural when you're selling the hardware (see clue 3). For the major players (Amazon, Google, Microsoft, Meta), for every euro spent, less than 10% of the sum comes back in revenue, with Meta winning the small palm for having only 1% coming back in revenue.

Even major players like OpenAI, Anthropic (or Mistral in France) struggle to exceed 40/50% of the sum in return.

In short, these players aren't profitable.

But this clue is very shaky. It's rather common to over-invest in the startup phase, especially in industrial fields that require investment.

Amazon wasn't profitable until its 9th year, Tesla until its 17th year, Uber 14 years. A phase of significant losses is nothing unusual when a sector is building the infrastructure for a new market. It's even almost expected.

In reality, the problem isn't that AI isn't profitable today, that's normal. The real issue is whether it can be profitable tomorrow, once investments have stabilized. This question remains largely open. Are today's investments sustainable or do we need to reinvest the same amount every year?

For now, given the obsolescence of the hardware used (chips for training and inference), the answer doesn't seem to be going in the right direction.

Verdict: ⚠️ concerning

Fifth clue: when history begins to repeat itself

Remember the bubble of the 2000s?

In 2000, all you had to do was rename your company with a .com, like sausage.com and boom, valuation increased. Irrationality was so great that investors rushed at anything and everything, and especially everything, just to not miss out on the next gem.

And let's remember that investors aren't always the best when it comes to rationality, because each time, during major crises, we get a new version of "this time, it's different" which basically justifies that economic fundamentals can be neglected.

Well, we're seeing something similar today. All companies are integrating AI, at least in their investor pitches. Some have started reserving .ai extensions just to ride the wave and I'm more than suspicious of the valuations of some companies like Lovable (6 billion), Cursor (60 billion), Devin (26 billion). It's cool for their creators but, making a wrapper on existing AIs, I'm not convinced that's worth that price.

Verdict: ⚠️ concerning

So yes, I can't say 100% that we're in a financial bubble but it looks like one. Extraordinary valuations, gigantic investments, an economy that runs largely on itself, companies still far from profitability and, above all, discourse that's starting to sound oddly similar to that of past major bubbles.

At this point, the most interesting thing therefore maybe isn't to know if a bubble exists.

It's to understand how it could burst.

Because all bubbles don't burst the same way.

The triggers

Bubbles always end up bursting but the detonator is often unpredictable. We can easily identify the weaknesses but it's sometimes a trivial event that ends up bringing down the system.

Despite this, I propose we play a game and try to list the possible triggers for a potential AI crisis.

Option 1: The IPOs of Anthropic and OpenAI go badly

Anthropic and OpenAI are planning upcoming IPOs. While the exact dates aren't yet known, this could happen between the 4th quarter of 2026 and the 2nd quarter of 2027.

During these IPOs, we'll see if "retail investors" decide that the stock price is actually worth what we're paying for it today. And precisely, there are several problems as we've seen:

• valuations seem far too high compared to revenue generated
• profitability is too low

To that I'd add another important problem: the amount of liquidity available on the market.

The theoretical valuation of Anthropic is 965 billion dollars and OpenAI isn't far from this figure. Obviously an IPO won't be for the full value but a part, let's say 10% (100 billion).

The records for IPO amounts oscillate between 35 and 50 billion. To go buy 200 billion (for Anthropic and OpenAI), will markets follow? Will the first raising, which would already be exceptional, not handicap the second?

Option A: both raisings happen in these orders of magnitude, it will necessarily be by selling other assets to mobilize cash, so by creating a drop in the prices of other companies.

Option B: it's impossible to raise these amounts, and the values of both companies drop

Despite this, I want to set this scenario aside because I find it obvious, too easy and therefore unlikely.

Option 2: Big investors reduce their sails

Today it's hyperscalers that are investing the most in the big AI companies, OpenAI and Anthropic to name just them for now. And when I speak of investment, I also speak of material investment with facilitated access to infrastructure.

Except that for now the return on investment is low. So imagine that tomorrow, Microsoft announces it's reducing these investments and lowering these datacenter/infrastructure costs.

The market won't just look at Microsoft and LLM publishers. It will look at Nvidia, AMD, TSMC, electricity producers. The whole chain.

Despite this, this option doesn't suit me either. True, for Microsoft or Amazon, the situation isn't rosy, but it's not catastrophic either, because these companies benefit from public contracts, particularly in the military field. And deciding to cut investments now would be precipitating failure, shooting yourself in the foot, which seems unlikely to me.

Option 3: Physical slowdown

This trigger is based on a very simple idea: datacenters have enormous resource needs. Eventually we know that these resources won't be available, whether for energy or mineral resources.

Imagine that tomorrow:

• we have supply difficulties on rare earths, for example related to a conflict with China (60 to 70% of mining extraction and more than 85% of refining comes from China)
• that available electricity can no longer be supplied for datacenters (e.g. in Dublin, the US, and Singapore)

We could imagine a slowdown not financial... but physical. And we fall back on option 2. We won't just look at the datacenter, we'll look at the whole chain, all planned investments that won't be made.

This option is much more credible, in my opinion, but in the medium term. It seems difficult for me to imagine a significant event before 2028. However, I may be naive because relations with China keep degrading and datacenter construction refusals are already increasing. I therefore classify this lead as credible.

Option 4: AI hits a ceiling

This is a real issue. Imagine that Opus 8 is "only" 3% better than its predecessor? Investors pay to see exponential growth. This is exactly what motivates all the circus we usually see at each release from these giants' marketing teams. They have to sell us the incredible, because they'll only be paid for the impossible.

What would happen if this rate of improvement in model performance slowed? Again, we come back to option 2. A major investor could decide to lower these datacenter expenses and you know the rest.

Is this option credible? The improvement over 1 year has been staggering and almost frightening. Despite this, I have the impression that the room for progress remains significant. But I say that being far from being a specialist. I would however tend to reject this option for at least the next 2 years.

Option 5: A macro shock

This option is more traditional. What if tomorrow a huge shock shook the economy with no relation whatsoever to AI. It could be an energy crisis, a political crisis (an open conflict between Europe and Russia for example), the Chinese real estate crisis that's been brewing for years.

This is typically the option where unknown unknowns are more numerous than others. Since this option is by nature completely unpredictable, I won't consider it either.

At this point you might tell me, "ok but you haven't retained any option, so you don't think the bubble will burst?"

That's not quite it. I think most of these causes are too predictable, or too long term. Again, predicting the trigger is particularly difficult. It will maybe be a combination of factors with a domino effect we don't anticipate today. The trigger for the burst of the internet bubble in 2000, for example, was the Federal Reserve's rate hike.

In any case, I have trouble not imagining a trigger in the next 2/3 years.

On the other hand, the consequences won't necessarily be what everyone imagines.

The consequences

I'm not done bothering you with multi-section chapters :)

We often imagine that the burst of a bubble is like the 2000 internet bubble or the subprime one: a huge market crash, lots of companies on the ground and done.

Well, not necessarily. Bubbles always end up bursting but it can be slow and smelly and we'll see again several scenarios.

Scenario 1: The bubble bursts violently

That's the easy scenario, the one we all have in mind. Valuations collapse, some major players go bankrupt, followed by layoffs in the tens of thousands and project shutdowns.

For this scenario to be as impressive as possible, we can imagine a Big Tech company going under, a sort of modern Lehman Brother, or Worldcom to draw the parallel with the internet bubble.

Beware, this scenario is obviously the most painful because it won't just affect "just" tech. We're talking about all hardware sellers, electricity suppliers, insurance companies that covered transactions, retirement savings plans, stocks etc... because yes, a large part of current investment is based on US values, particularly in tech.

I insist, it will be very painful even if you think you're far from AI. The burst of the internet bubble was between 60 and 80% loss of value of stock indices, millions of jobs lost.

Scenario 2: A dilution

We might not necessarily have as cinematic a scenario as the first. We could simply have a gradual decrease in investments but a maintenance of infrastructure and a reabsorption of large LLM publishers by hyperscalers.

It would potentially be the end of the game for Anthropic, OpenAI as independent entities and the losses would be largely absorbed by GAFAM.

(I'm not talking about xAI which is already part of a consortium or Gemini which is part of Google)

It wouldn't be an explosion but a dilution of the AI bubble into the accounts of big techs

I speak of dilution but we could also speak of consolidation. We'd certainly have fewer actors but a large part would be absorbed by the more resistant ones, Google, Amazon, Tencent for example.

It would be less spectacular but not without consequences either. Part (and only part) of the investments would be set aside. Heavy investments planned for in 5 years would be stopped, so with more time to rebuild correct budgets. There would still be layoffs but less massive.

My little pinch in the heart in these two scenarios is that I can hardly see a European player like Mistral survive this scenario without being completely absorbed, unless Europe decides to invest massively through public procurement. The gap is large and I find their strategy poorly lisible today, and even poorly anchored in the tech ecosystem. But I wish them the best, because we need a European champion on the subject. Even if at worst we'll cobble together things with open weight models.

Now you know what?

If we exclude the Mistral case, despite all this, I'm almost hoping it happens and I'll explain why.

Why it wouldn't be so bad

A bubble, let's remember, is misallocated capital. Now, I'm not saying that AI as a technology isn't worth it. It's completely redefining many professions. But is that reason enough to do anything and everything, I'm not sure.

When Google questions its objective of carbon neutrality 2030, it's a failure for everyone. And then, is it healthy to have this race to the bottom for datacenters that we know we can't all power with electricity, except to reopen gas power plants?

Besides, when I talk about misallocated capital, I'm very happy for the people who created Lovable or Cursor, that's cool for them. But all this capital to create a tech that's already being competed with by tons of products vibe coded 2 years later, that's a shame. The same money could have funded something more useful.

Capitalistic irrationality, excesses on mineral resources that plunge us into a RAM and component crisis, increased tensions between countries, and simply the current climate around AI that's becoming unbearable between pros and antis, I find that's a lot of reasons to hope it stops.

If the bubble bursts it will sign a form of return to rationality. I don't believe in the disappearance of the tech, the same way the internet didn't disappear in 2000 and the train also survived the railway bubble of the 1840s.

What disappears in a bubble isn't necessarily technologies: it's mainly absurd valuations, projects that had no viable business model and investments made because "everyone's doing it".

More rationality means better thought-out investments, R&D more focused on optimizing what we've already learned: mixtures of experts, quantization, pruning, HBM, model optimization. We need to now do as much but with less. It's a necessity.

We already have enough to work with what we have in hand and it wouldn't be bad to pause a bit, think about the uses that work, and there are already some, find a real path to profitability and redirect capital, particularly toward the challenges of decarbonizing the economy (electrification of the vehicle fleet for example) or adaptation measures to climate change.

Will the burst of the AI bubble see the end of AI? Not so sure. It will be more of a landing. But whatever the scenario, it will be beneficial, a return to economic AND ecological fundamentals.

Don't be mistaken, given the sums already invested, the burst will hurt, a lot, but the longer it takes, the worse it will be. The band-aid needs to be ripped off, quickly.