Between colossal investments and a ROI that is difficult to prove, we are hearing more and more about a possible AI bubble waiting to burst. It’s true that in the tech ecosystem this wouldn’t be the first or the last time and that, what’s more, the sector loves to one day burn what it adored the day before (remember the metaverse?) and so behind every techno-enthusiast lurks a potential Cassandra.
But it’s still worth looking into the subject.
In a nutshell :
- Artificial intelligence requires massive investment without immediate profitability, a classic pattern of emerging technologies where companies invest to create a market before seeking profitability.
- The sector is characterized by structurally high costs, particularly in infrastructure and model training, while customers remain hesitant to pay the true price of the technology due to a lack of proven sufficient ROI.
- Two dynamics coexist: the large Deeptech companies that absorb billions without being profitable, and more agile startups that take advantage of the former’s products with specialized and profitable products that capture the value of the market.
- Several scenarios could prevent a collapse of the sector, such as the emergence of high value-added applications, market consolidation, technological innovations reducing costs, or more suitable pricing models.
- Unlike the dot-com bubble of the 2000s, the AI sector is based on proven technologies, concrete uses and government support, making a sudden collapse unlikely, although a market adjustment is inevitable.
The normal life cycle of an emerging technology
AI requires a lot of investment, brings in little, and its adoption is not as significant as people would like to think. So what?
This is the nature of the cycle of an emerging technology and a startup.
A lot of investment at the start and a market to be created and convinced. It is therefore out of the question to pass the cost on to the customer, let alone make a margin. You have to create the market, almost “buy” customers by doing nothing more and nothing less than dumping.
The bet is that, over time, the market will grow in size (number of customers), in value (each customer is willing to pay more because they perceive the value) and that, at the same time, the necessary investments will decrease.
But in the meantime, cash is being burned.
But we must keep one thing in mind, and I will keep repeating it: at a certain point, if the market does not grow in size or value and/or the necessary investments do not decrease sufficiently, we are faced with three possibilities.
1°) The customer does not make the technology profitable, stops paying and using it.
2°) The vendor does not make a profit on his investment and goes out of business.
3°) Neither of them makes any money with the technology, and it spells the end for it and the sector.
The only question is when that moment will come, and it’s quite simple: it’s the day when investors decide to call it a day because they see that it’s not taking off or not taking off enough and that they will never get their money back.
AI is a big cash consumer
Unless you’ve been living under a rock for the last few months, you’ll know that huge investments are needed to move the sector forward.
For example, Meta plans to invest up to $65 billion in AI projects by 2025 (Meta Plans to Spend $65 Billion on AI in 2025: What It Means for the Construction Industry) and OpenAI plans to raise an additional $40 billion (OpenAI said to be in talks to raise $40B at a $340B valuation).
You have certainly also heard of the Stargate project in the United States, which aims to mobilize 500 billion dollars over 4 years, involving major players such as Oracle, OpenAI, SoftBank and the Abu Dhabi investment fund MGX (Announcing The Stargate Project) with this detail that it is not clear how the protagonists can invest an amount that they do not have. European response, with a plan of 109 billion euros in France, which represents the same investment per capita as Stargate (Details of 110 billion euros in investment pledges at France’s AI summit).
Impossible profitability in the short term
The current economic reality of the sector shows a flagrant imbalance between investments and revenues. According to CB Insights, AI businesses have attracted more than $183 billion in investment between 2013 and 2023 (The State of AI 2023 report), and then $100 billion in 2024 alone (The State of AI 2024 report), but most are operating at a loss for the reasons mentioned above.
But we need to look at this market with a little nuance. There are two types of AI businesses: to oversimplify, there are those that create models (Deeptech) and those that create products that use them and/or use AI to complement their employees.
While an OpenAI swallows up billions to create and acquire a certain global market, others use OpenAI models in vertical and specialized products that will have a simple go-to-market and an easier-to-demonstrate ROI.
Today, we are talking about a generation of start-ups (Cursor, Midjourney and ElevenLabs) that, with little funding and few employees, manage to be profitable in just over a year and exceed 100 million ARR (Annual Recurring Revenue). Cursor, for example, uses Chat GPT or Claude from Anthropic. By positioning themselves between Deeptech and the customer, these businesses are in tune with the times with vertical and specialized products ([FR]“Generative AIs will soon disappear”) but in a way capture the value of the market without having to make heavy investments.
But if Deeptech collapses, these businesses, which are highly dependent on it, will collapse with it.
When we talk about an impossible probability in the short or even medium term, we are talking about Deeptech, but if they fall, they will take almost everything with them.
In any case, for the behemoths of the sector, there is a widespread practice of using a loss-leader price aimed at gaining free or paying users before seeking profitability. In other circumstances, this would be called dumping, but it is inherent to the sector. Today, AI still means a lot of investment for little income (The Generative AI Con).
This creates an environment where the real costs are not passed on to customers, which is not viable in the long term.
Structurally high costs
There is no magic formula. The sector faces a dilemma: either drastically reduce costs or significantly increase revenues. This is the case with start-ups, as we have seen, but there is something special about AI.
In theory, it is said that AI could generative AI could add the equivalent of $2.6 to $4.4 trillion per year to the global economy (The economic potential of generative AI: The next productivity frontier), but this value remains largely theoretical and difficult to capture or even simply measure. Four years ago, the valuation of the Metaverse was estimated at 13 trillion dollars, and we can see what happened to it (Lessons From the Catastrophic Failure of the Metaverse).
But while the market potential is struggling to materialize, the costs are very real and so is their trajectory.
The learning costs of models were expected to decrease by about 50% per year, and this has been the case. On the other hand, this decrease is offset offset by the growing demand for larger and more complex models.
At the same time, the infrastructure, electricity and cooling costs of data centers continue to grow, as evidenced by the aforementioned colossal investments, which mainly concern the infrastructures in question.
Customers not willing to pay the price
If nothing is to be expected in the short term in terms of costs, perhaps it is time to look at revenues.
Despite the interest of businesses in AI, its uses and benefits are much less advanced than expected (Generative AI in the workplace: revolution or illusion?). Most initiatives remain stuck in a pilot phase, fail to scale up, and despite the promises, ROI expectations are clearly out of step with the predictions of publishers and other consulting firms: “Across all sectors, executives surveyed report limited returns on AI investments at the business level. Only 19% report that revenues have increased by more than 5%, 39% see a moderate increase of 1-5%, and 36% report no change. And only 23% see AI as driving favorable cost changes.” (AI in the workplace: A report for 2025 | McKinsey).
Again according to the same source: “Almost 90% of executives predict that the deployment of AI will stimulate revenue growth over the next three years. But ensuring this growth implies a transformation of the business, and companies have a poor record in this area. Almost 70% of transformations fail.” This is not a reliable forecast; it hardly resembles a prediction, at most a prophecy that one hopes will be self-fulfilling.
Not much more reassuring, only 31% of managers expect to be able to assess the return on investment of their AI initiatives within six months, and none have yet succeeded in doing so (The ROI puzzle of AI investments in 2025). This is a far cry from the expectations of 67% of them to see their organization totally transformed within two years.
But on the other hand, we are told that “97% of business leaders whose organizations invest in AI report a positive return on investment” (Artificial intelligence investments set to remain strong in 2025, but senior leaders recognize emerging risks) without being precise about the ROI figure.
Why such different figures? In my opinion, the answer is obvious: nobody knows what to measure or how to measure it (The True Value of Generative AI: Measuring ROI, And Why It’s Tricky), so it allows people to say anything and everything based on perceptions, because they want to convey a positive message to reassure themselves and others, or even because it’s a case of self-fulfilling prophecy.
Finally, we also realize that the sectors that have the most to gain from AI are the ones that invest the least (The disconnect between AI spend and potential) for one simple reason: the lower the margins in a sector, the higher the expected ROI, and today there are too many uncertainties in the field.
Finally, in the absence of tangible ROI, businesses are only willing to pay for perceived value, and today this is not forthcoming ([FR] No, generative AI is not free, so who is going to pay and for what?).
Towards an economic dead end?
Today, developers of fundamental models (OpenAI, Anthropic, etc.) are investing billions to remain competitive, without a stable revenue model, while business customers are reluctant to pay the true price.
Let’s repeat that this is normal in the context of the economics of an emerging technology. The only question here is the size of the investments, which requires either that businesses be willing to pay a lot for the technology in question or that the threshold for expected profitability be pushed back to a horizon never seen before.
Is the AI industry going to crash and burn?
Not necessarily, because there are ways to avoid it and the tech sector has a certain maturity in the field.
Scenarios to avoid a crash
Here there is only the classic but it is still worth talking about because it is a safe bet that this is how things will happen.
First of all, there is the emergence of applications with high added value.
We are talking here in particular about vertical applications that address either a specific sector or a specific use case or even both at the same time, and as we have seen previously, this works well.
However, there is a limit to this scenario. Unless the publisher develops its own small language model (which is in keeping with the times), being dependent on a Deeptech publisher can weaken it if the latter goes out of fashion or disappears.
There is also the consolidation of the market.
It’s an old habit in the tech world: startups lead the way, some succeed, some fail, and the giants of the sector help to clean up the sector by making acquisitions that allow them to catch up in a particular area, often because they let them create the market without investing too much, telling themselves that they would buy them back if necessary.
This consolidation was announced in 2024, then 2025 (AI adoption to force wave of software consolidation and The AI market in 2025: The rise, the fall, and the reality check) but for the moment we don’t see much coming.
But 2025 has only just begun.
We can also bet on potential innovations that will reduce costs.
We remember the shock caused by Deepseek with its frugal model that caused panic in the Valley. With hindsight, it is not clear that this was an unreasonable move, as Deepseek certainly did not tell the whole truth about the resources actually used, but in any case more and more specialists believe in smaller models (SLM) (What is a small language model and how can businesses leverage this AI tool? and Small vs. Large Language Models: Which One Reigns Supreme?), more specialized, cheaper to develop, and less consuming in financial and natural resources.
Nvidia predicts that the efficiency of GPUs for AI will improve by a factor of 10x in the next three years, but will we still need GPUs as much?
One solution also lies in innovative pricing models.
Speaking of a technology that is very infrastructure-intensive, there is increasing talk of the relevance of pay-per-use pricing (2025: AI, Usage-Based Models, and the Future of Revenue Generation).
But for me it is a model perfectly suited to an industry that has reached a certain level of maturity but is in an emerging phase. It can kill the “discovery” and experimentation phase in which most businesses find themselves today.
Finally, we also hear about public-private partnerships.
Some are in operation in the United Arab Emirates, others in the USA in the public health sector, but here we are addressing the question of investment, not that of future profitability, which is only postponed.
Bottom line
The AI sector is faced with a complex equation: how to balance massive investments, hesitant adoption and delayed profitability. Once again, this is not a new issue for the tech sector, but the scale of the investments adds particular pressure compared to what we have experienced in the past.
The similarities with the dot.com bubble are numerous (The Dot-Com Bubble vs. The AI Boom: Lessons for Today’s Market) but there are notable differences.
We are dealing with mature technologies, clear and often very B2B use cases, the economic models are known and proven with a clear way of making income (even if insufficient) and, above all, governments support the sector.
Today, it is the scale of the sums involved that makes us fear a bubble, but on closer inspection, we are still a long way off and, moreover, there is no stock market speculation on AI at the moment ([FR]An explosive speculative bubble in generative artificial intelligence?).
That doesn’t mean that some AI players won’t disappear, that there won’t be any crushing failures, but compared to the famous bubble of the 2000s, we have businesses with real economic models, and while there will be consolidation or even a return to earth, the risk of a crash is unlikely.
We can therefore continue to believe in AI and explore its potential. In any case, we have no choice (ROI Vs. RONI: why businesses should invest in AI despite uncertain ROI).
Image: speculative bubble by Cherdchai101 via Shutterstock.
