Is AI the next dot-com crash? | Business Beyond

Billions are being poured into data centers to make two magical letters work. AI, AI, AI, AI, AI. Artificial intelligence. The one technology to save us all. You shouldn't be asking what it can do, but what it can do if you buy into the hype. This technology has extremely high narratability. We might have AI that is smarter than any human by the end of this year. But at the same time headlines like 95% of companies don't report profit increases connected to AI make you wonder is the hype actually legit or are we really close to an inflated bubble. To get to the limits and problems of AI and its bubbly character that I'm going to explore with these two, we first need to get what companies can actually use AI for. It's not that BMW builds entire cars with AI. Yet. But they're using AI to manage their production equipment which they hand over to production partners. That can be presses, welding machines, lifting cranes, you name it. And thousands of them. AI helps to keep track of these 250,000 tools that BMW needs to manage. Our AI is like, yeah, it's like small little helpers which do specific tasks. So they are creating the inventory order. They are sending the task to the supplier. They are checking the results. And at the end of the day, if everything is green, so green light, we say, okay, this task can go to our asset accountant. What BMW is using here is called an agentic AI. Even multiple of those agents between BMW and its suppliers. First, a task agent asks the supplier to perform an inventory task. The supplier takes a photo of the QR code and this data runs through a validation agent that checks if everything looks correct. If not, a human gets alerted. If everything looks good, a documentation agent takes care of the paperwork. The agents are based on large language models like Chattopadhyay, Gemini and Claude. But they work a bit differently. We say, hi Chattopadhyay, please give me three recipes that I can cook this week for myself. And then it will do that from whatever information it has in its system. All right, let's get cooking. Quick, colorful and full of flavor. An agentic AI, however, would ask for my taste preferences, my caloric needs, habits, scrape supermarkets for cheaper offers, check how potential items would be ranked against WHO health recommendations, verify its own decisions and create a shopping list tailored to the layout of the supermarket that I'll be going to. Simplifying it, an agentic AI does more with less human input and makes decisions on its own within set limits. You can automate more and need less workers. In narrow scenarios, it can even operate without ongoing human input. In BMW's AI automation process, QR codes have an important role. This QR code is in the middle of everything. The supplier gets the inventory order, then he goes to the tool and scan the QR code. Ah, so you literally just take the phone and then you take a photo and then you have all the information. And we get all the geolocation and metadata through the photo and have the inventory number with the QR code. That's it. By using AI, BMW wants employees to be able to focus on cases outside the norm but not spend time with the ones where everything is clear. But such an integration into an existing workflow takes time. For BMW, it took one and a half years to get kickstarted on agentic AI. And this integration challenge is often underestimated in the hype of things. This technology relies, for it to be potent, it relies on redesigning the way those employees do their work. And that ends up being a business problem, not a technology problem. This is David Crawford. He helps enterprises integrate AI into their workflows. Unlike some more recent, you know, technology disruptions, you can't just deploy the technology and expect to get efficiency. You actually have to do the hard work of mapping out your business process, changing how you want to do it differently, etc. If successful, Crawford says, companies can see a rise in revenue of up to 10 to 25% before taxes. But during my research, I came across a lot of companies that promised the use of AI or agentic AI. But when I asked them about showcasing it, many were like, we're still testing, product isn't ready, all sorts of excuses, which makes me wonder again, is all of this AI stuff a bit of crap? And it's not just me. There are also some well-known over-promises of AI. DoNotPay claimed that it could automate anything a lawyer does. The founder even offered $1 million to any lawyer that would argue in front of the Supreme Court with the help of DoNotPay. Eventually, they got defined from the Federal Trade Commission for false promises. In other lawyer-replacing AIs, Stanford University scientists who double-checked these AIs found severe hallucinations in their answers. Sometimes the hallucination was so bad that it reversed the meaning of the laws. And even benchmarks don't seem to match real-world performance. Cognition's agent Devin, a software engineering AI, had great benchmarking scores, but in the real world, they found that Devin was only successful at three of the 20 tasks it was built for. This GPT-5, for example, and even Gemini 3 can do well on, or actually do better than, humans on international math olympiads. But then they can fail on extremely silly questions that have never been in any benchmarks. This is Rao Kambamputty. He works on fundamental problems in planning and decision-making by AI systems. There is such a thing as, you know, studying to the test. And when you study to the test, will you generalize beyond that? If you actually happen to see specific exams, the questions that will come. The general, you know, the entire premise of education has been that if you somehow do well in the test, then the capabilities that you got in doing well on that test will help you generalize and, you know, handle other situations. This notion of generalization is something that people, that is like still remains like an Achilles heel for LLMs. So it has become difficult to make a connection of how good an AI performs in benchmarking to what it actually can do in a reliable way in the real world. One of the biggest concerns around AI is always hallucinations. How do you deal with them in these kind of automated workflows if anything like goes wrong? So the important part to control that is to have really strong guidelines and our agents are fixed to one process step in the workflows. So they have clear rules so that we can ensure no hallucination is happening and we have always these testing possibilities in every single process step. Whoa, AI and no hallucinations and 100% reliable. Sounds like a dream, but I'm going to check with Rao Kambhapati later. But it's certainly a fit for David Kirsch and Brent Goldfarb's analysis of 58 historical technological innovations like electricity, aviation, railroad, dot-com boom. They checked why some of them created bubbles and others didn't. And they ended up with four reoccurring factors. First, a narrative to buy into like... We might have AI that is smarter than any human by the end of this year. And I would say no later than next year. And then probably by 2030 or 2031, call it five years from now, AI will be smarter than all of humanity collectively. I have engineers within Andropa who say, I don't write any code anymore. I just let the model write the code. I edit it. I do the things around it. I think, I don't know, we might be six to 12 months away. Second, uncertainty around how to use the new thing. And if 95% of companies don't see a return on investment on AI, that's a lot of uncertainty. Third, the presence of pure play companies, the ones that do nothing but AI. And fourth, novice investors. So what we did is we took this set of technologies and said, well, and for each one, we did a kind of deep dive into the history of that technology. Where did it come from? Who were the promoters? What were the opportunities associated with it? How did it change economic activity? And these four pillars emerged as associated with increasing the likelihood of a bubble. It's not a deterministic model. We can't say, oh, they're narratives, technological uncertainty, pure plays, novice investors, therefore, there will be a bubble. And the first pointer for a bubble is narrative. You and I are living through this once in human history transition, where humans go from being the smartest thing on planet Earth to not the smartest thing on planet Earth. These narratives or stories ultra simplify complex processes. And part of the reason why everyone is like, yeah, that might be true, is because we've been surrounded by these narratives for quite some time. For example, by Hollywood movies like Star Trek, Her, Space Odyssey, Axe Machina, The Matrix, where technology can do almost anything. This technology has extremely high narratability. It can support so many narratives that, let's say, I don't buy the one about replacing the need for human labor, but there's some other narrative that AI does kind of plug into that connects to my preferences. And then all of a sudden I'm like, OK, I'll write a check for that. So check for the narrative factor on the AI bubble. BMW also uses agentic AI for a different task. Automated offers for companies that want to buy, lease a lot of cars. So there are all sorts of questions about car models, specifications, engine sizes, lease length, rates that usually would all need to be read and categorized and answered by a human employee. But at BMW's subsidiary Alphabet, all these emails, inquiries get fed into an AI that lays out a rough answer and creates quotes that a human double checks. So what was a difficult thing to work with when you need an AI to access all these kind of different data points? Well, the challenge was to make the system work for the most cases, most of the times. So it's easy enough to make the AI make a perfect job just for one configuration. But trying to do this for all of them, adjusting the prompts and the settings and the configuration of the systems and then you have run it against all different documents again to see how it performs against them. In AI lingo, that is called the reliability of a model. The holy grail here are the five nines, meaning a reliability of 99.999%. Even top models today have failure rates of at least three, but up to 10%. So a lot of hallucinations when it comes to summarizing short documents. And many experts even argue that these hallucinations are here to stay. You might think 90% doesn't sound too bad, but let's transfer that to a real world example. Flights. What if 10% of flights in the US would crash? That would have been about 1.7 million crashes in 2025. Absolutely horrific. And so there is nothing that anybody can see that can theoretically remove hallucinations from such a model. You can try to improve it with training, but even actually there was just a recent paper, I believe in Nature Machine Intelligence talking about the fact that LLMs cannot reliably distinguish between fact and belief. And the belief basically is if they hear somebody saying something that becomes a belief. And so there is always this question of you can try to make the data more clean, or you can try to do additional sorts of training. But if you just let the machine, let the model, you know, come up with a long completion of a prompt, there is very little reason to believe that it will have 100% accuracy. But this makes AI harder to utilise for companies, because there's always the looming danger of something going wrong, also known as AI slop. That's when an AI does a bad job and humans need to spend more time fixing the mistakes than it would have taken them to do the task originally. David Kirsch and Brent Goldfarb described this factor as uncertainty. How will a new technology translate into real business? Something that also happened when radio came along. Why would we want to broadcast from one to many? And the thought was, well, we do that in church, so therefore it'll be used for religious services. But then the question is how to pay for it? Because how do you know if somebody is tuning in? And this became a huge problem for early radio. So big that there even was a public contest for ideas on how to make money with radio broadcasts. And AI could be running into similar issues money making wise, because to scale AI, you need loads of money. Loads. To build the necessary infrastructure, data centres. Until 2030, McKinsey says, data centres alone will need 5.2 trillion dollars in investments. That's 15 times the amount of the Apollo space project costs. 17 space missions. Just for comparison, in 2025, AI tech is expected to generate around 60 billion dollars in revenue by one estimate. I'm just getting dizzy from these numbers. We need to find new uses for this technology that transcend employee productivity and actually do some of those really novel things. Accelerate drug discovery. Accelerate, you know, health and wellness. Accelerate healthcare. And the AI industry needs to find new revenue streams ASAP, because they might be living on borrowed time as... Even companies like Google, Microsoft, Amazon and Meta, who've got a lot of cash sitting around that they can spend on data centres, even they are taking on debt to finance their data centres. In one of the largest private debt deals on record, Meta received 72 billion US dollars to finance its gigantic data centre Hyperion, which is going to be built in Louisiana. Also, Oracle took on 18 billion in fresh debt for its AI infrastructure. Until now, it was one of the main things that separated the AI boom from what it's often compared to, the dotcom boom and bust. Back then, many companies racked up debt as they raced to lay fibre optic cables into the ground, that years later would become today's high speed internet. But we all know what happened before, the dotcom crash. And not everybody is turning a blind eye to this. In October 2025, the UK's central bank, Bank of England, put out a warning about the increasingly debt-financed AI infrastructure and AI stocks valued at close levels seen at the peak of the dotcom bubble. And there's a really interesting graph in this report. This one. It shows the exact moment what happened to AI stocks when the news broke that the Chinese model DeepSeek can achieve similar results like with a fraction of the computing power. And as you can see here, everything took a bit of a nose dive. The stocks recovered after some time. But a breakthrough like that could render these gigantic data centres and the expected revenue from them somewhat useless. So that's definitely a check for uncertainty. The last bubble sign from Brent and David has something to do with all of us. Because people, while not really having a clue what's going on, still throw money into AI stocks. In 2024, 30 billion US dollars was poured into NVIDIA by retail investors and made it the most popular retail stock. But Brent and David even go beyond this definition of novice investors. We're all novices. That's the take-home lesson. There was a saying on Wall Street that you're a novice until you go through a bear market. Just to recognize that that can really happen. Nobody's done AI before. I don't know exactly how useful it's going to be in each application. I don't know how expensive it's going to be over time. I don't know how... We don't know actually what business model is going to be the most effective way for AI. So where does this all leave us on the bubble thing? Undeniably, AI has an impact. David Crawford said some results from Bain and company clients range from 10 to 25 percent uplift in their earnings before taxes. And BMW says... We see this percentage also within the inventory workflow. And when we think about how we can scale now over all of our processes, I think Gen AI will be able to bring that potential. But you don't get these results by throwing ChatGPT or Copilot into a company. Also undeniable, there is some trash talk going around. Some lofty company valuations. And we checked all the bubble boxes from David Kirsch and Brent Goldfarb. But is it March 2000, just before the burst of the dot-com bubble? We don't know. And according to many experts, no one really does. I'm using AI models as much as I can to see what they are capable of. And the advancements in the last years have been... But I'd say that we're missing a breakthrough in reliability to match the crazy data center spending and the valuations. Microsoft's stock just took a 12 percent nosedive because it became clear how much they spend on AI and data centers. The whole AI house of cards seems to be pretty fragile. And maybe we're just one deep-seated moment away from everything going...