Transcription
Thank you very, very much. Uh, it's a bit like St. Pauli with 5,000 standing places. Thanks for coming. I'm very happy. You know who I am. You know why you are here. Uh, we'll get straight into it. Um, it's about Beyond the AI Hype again. We started with the Hype Cycle last year. There isn't a new one yet. Last year we went into this unpleasant valley. Um, people were asking themselves, so generative AI has entered this valley, so to speak. People were asking if AI is hitting a wall, if AI is still developing at all, if we are putting far too much money into it. I said that's nonsense. And I believe we know today that it's nonsense. There hasn't been a year in which AI has developed as strongly as the last one. Um, this here is the Artificial Intelligence Index, which shows the capabilities of the models on a scale of 0 to 100. I've marked the ones that were the best models when I was last here. Um, the companies are now um there. If you look at this on a graph, these are different benchmarks where the capabilities are measured, um, you also see that there has been a completely steady development, and no, no, uh, wall against AI is in sight. There have been really significant breakthroughs, especially recently with image generation models, for example. Chat GPT or Midjourney has just released the Images 2 model, which is simply in a league of its own. It's, uh, if you know how this ELOScore works, being 300 points ahead is simply a completely new league. And these new image generation models no longer make these old mistakes. The text, the writings are no longer broken. It's no longer blurred, smudged, washed out, there are no artifacts in it. I can create tourist maps that are not perfect, but much, much better than before. I can generate historical posters, build super PowerPoint slides within a minute in one shot. Uh, I can paint picture puzzles for children. Um, it's simply a completely new order of magnitude, uh, that has been created by OpenAI in this case. Another thing that has changed significantly, uh, is the time window in which AI can work autonomously. When I was last here, AI could write a small piece of software for 2-3 minutes, write a small automation for Excel, spit out a longer text. Today, AI can autonomously work for half a day with an orchestra of agents and then complete a task with a 50% probability of success. So you go to bed in the evening, tell Claude or Codex what to do, and the work is only finished the next morning because the AI has worked through the whole night. AI is not only developing linearly, but the development is even accelerating. I believe it's due to three things. Firstly, because we are not only working on better software and hardware, but also because the data quality is constantly improving. There are now many companies actively working to enrich data even better for AI. The other is that AI is now improving itself. We'll come back to that later. And the introduction of Reinforcement Learning in AI makes the AI better and better while you use it. This is what Google has been doing for 10-15 years, learning from your behavior whether the results are better or worse. And that makes AI better and better, so to speak. That's why you should say yes or no after each result. Uh, your activities with AI are tracked. If you have to keep asking the AI, the company notices that it was bad. If you are happy, copy-paste or share the result, then they know it was a good result, and that makes AI even better. Um, last year I explained that AI can already do everything that humans can learn, better than humans, under two conditions: that it can learn from the internet or from data, and that humans have built a test for it, is AI better at it today? That's a fact. That was a fact a year ago. That's why the last exam of humanity was created. 2500 super tricky questions from all areas of mathematics, biology, history, um, logic, which the AI could not solve at the end of 2024. I'm telling this again because AI is now solving 50% of these questions. Yes, so we will probably need a new final exam soon, uh, because AI will approach 100% there. Um, what I also, what I'm teaching the zero line, are the costs, uh, of applying AI. If I'm satisfied with the same performance as last year, I can get it 10 times or even 100 times cheaper now. Not the training, but the application of AI is becoming cheaper and cheaper. For example, who knows or has anyone read Leo Tolstoy's War and Peace? I certainly haven't. Ah, Lis, very good. So it's a real tome. 1000 pages, 300,000 tokens or so, that's about 250,000 words. Tolstoy wrote it over 7 years. It would take a minute today and cost 15 cents to churn it out for AI. Yes. Uh, there are even thicker tomes, 4000 pages. Um, that would cost 70 cents, to generate all German tax laws, um, overnight. We have now reached the point where the generation or the generation of text has become so cheap that last year, for the first time, we had more AI texts on the internet than human content. Uh, that's where we are now. Um, and it's definitely taking over. There was a recent analysis. There's a very typical artifact that, when you work with ChatGPT or Claude, there are these long hyphens, you probably know that, or words like "delve" that occur very often, but also what I call fake dialectic, meaning saying it's not this, but that. You read this again and again, and if you read it, then someone used AI. You can say that quite clearly. And so a typical example, basketball is not just a game, it's a global cultural engine. That's not just a tech upgrade, it's a cultural shift. You read this everywhere, if we, you can't unsee it anymore once you know it, yes? And it has now been measured, and these are only figures up to 2025. And this isn't even about LinkedIn or Instagram, but it's what companies are telling their shareholders. So, it's real CEOs, managers, who are apparently mostly getting their speeches written by AI. Politicians too, by the way. And it's now completely exaggerated. Here, for example, T Online Jobs says: "Recruiting doesn't start with the job advertisement, it starts with visibility." Huh, what is visibility for, if I don't have job advertisements? Or here, our dear libertarian friends from the Initiative Neue Marktwirtschaft, who deal with jobs, as we know, um, these are not statistics, these are résumés, yes? Or this anonymous entrepreneur, I've blacked it out, comes from a pyramid scheme in the broadest sense. Um, he has now really, give me a break, this doesn't show competence, but insecurity. This is not strength, this is fear, this is not leadership, you can tell it's not fun, no give me a break. Hello. So, and it's not just one person, you see it everywhere. In the past, and this is the opposite of leadership, if you let AI write your texts from the default settings. It costs you a markdown file to explain to the AI how you speak naturally. So, if you do that, which is totally okay, maybe your output will be better, but do it in your style and don't use the default. Forbid the AI to do that, because nobody will take you seriously if you do that. Now that you've seen this, you'll never unsee it. LinkedIn is ruined for you. Um, what is also increasingly being ruined is Spotify with AI Slop. That's why Spotify has now introduced a seal for human artists. And that's the right way. In the past, people thought a lot about how to label AI content. That's over. It will always be outsmarted. AI watermarks don't work. We have to label human content with a seal of approval, because in the future, human content will be 5% of the content, and the rest will be completely natural AI. So, if anything, we have to use signaling and label the few humans who still produce content. Um, I said the costs of using AI are becoming cheaper and cheaper. The costs of training AI are becoming more and more expensive, or rather, for every billion dollars, I get less progress. Um, this means it's becoming increasingly expensive to participate in this race and to build the best model. And at the same time, the best models are only the best models for a few weeks or months. This means the time at the top is often only a few months. It's like being a pharmaceutical company. I've just invented Osemp, and after a month, everyone can use it for free, or rather, I can copy it because there's a better drug. That's roughly how AI research works, and that's why these companies need so much money. They need so much money because to build the best model every year, you need five times more hardware. Yes, if I only have three times more hardware, I'll be out of the top league. No more Champions League for us. Um, five times more hardware every year, which has led to 400 billion US dollars flowing into data centers last year. This year, it will be 790 billion, presumably. I've updated this figure once already, and it will probably be increased again over the year. That's at least what happened last year. Um, that's 187% of the cloud revenue of these companies. They are spending multiples of what they earn from cloud at the moment on AI. The entire cash flow that Google, Amazon, Microsoft generate is now being exceeded by investments. This means these companies have to incur net debt to invest even more than they earn themselves. Because 790 billion is a figure that no one can do anything with, I've tried to make it a bit more tangible and compared it with a few historical projects. This is always inflation-adjusted. So the Manhattan Project, the invention of the atomic bomb, actually only cost 2 billion. In today's money, that would be 50 billion. The Marshall Plan, the European Recovery Program, the program that rebuilt infrastructure and industry in Europe after World War II, would be 190 billion in today's dollars. The Apollo program, 17 missions to put humans on the moon, cost 290 billion. The largest project in humanity, the International Space Station, cost 155 billion. These four projects combined are still less than what BigTech will spend on new data centers this year. The EU budget, which is too high, say others, is only a fraction of that. The special fund that is supposed to save our German economy over 5 years, so what will be spent over 5 years, is smaller than what BigTech is spending on data centers. 790 billion is 2 billion per working day. Every working day, new data centers worth 2 billion are being built in Germany. Calculated for approximately 1.3 billion AI users, that's $600 per user. This means that's the expectation of what you will eventually have to generate in AI revenue, um, so that these companies can amortize the investment. It's 2% of the US Gross National Product. For these companies, this is not bad, they are becoming more and more valuable. Um, BigTech is now worth 20 trillion, 20,000 billion. The DAX, in comparison, is about as big as Meta alone. So the 40 largest German companies are worth as much as Meta. Uh, the rest, uh, of the US companies, only the Big Tech companies, is ten times larger. They have added 6 trillion, 6,000 billion in value in the last 12 months alone due to the AI boom. Um, those who don't have so much money yet are the frontier labs, i.e., the research pharmaceutical companies, I would say, the research AI companies, AI companies. Um, in the past, it has been critically questioned many times how much money Amazon burned to eventually become profitable. Um, Amazon invested 2 billion US dollars to eventually become profitable. It was always criticized for that. Spotify also 2 billion. Um, Tesla burned 5 billion before they made money, over 30 billion. It was a scandal back then. How can you invest 30 billion before seeing a return? Compared to our AI companies, that's all a joke. Yes, OpenAI will accumulate at least 250 billion in losses before they earn the first dollar of positive return. One large Hong Kong Shanghai, HSBC, a large bank, even estimates 500 billion before. OpenAI becomes profitable. That's why these companies are raising more money. Um, than ever before. OpenAI has just raised a mega-round with 122 billion, I believe, um, of new capital. This means that for all the other startups, there is hardly any money left. So a large part of venture capital now flows only into AI companies, and there's hardly anything left for others. Why they are now looking for new sources of funding. Um, for other companies, so OpenAI is worth about 850 billion. Anthropic could do its next round at 900 billion. Other companies took 40 years to get there and had to go public. And of course, AI companies will also go public this year. Uh, they are ripe for it. In the past, companies, so the BigTech companies, went public after an average of 6 years. Uh, in that respect, both are ripe for it. Of course, it is also questioned whether they are not valued far too highly. Um, one can compare it with other software companies that have recently gone public or like Databricks, which will do so soon. And if you compare the capital invested so far with how much revenue they have actually generated, you always see a ratio of 3:1. So, I have to invest $3 in a startup to generate $1 in revenue. And these, um, AI companies are totally in this area. This means they are not far, uh, from this standard. Um, the valuations, um, are of course higher, but measured by revenue, they are not that large. They are also growing much faster than historical companies. Uh, this is an estimate now. We don't know how fast they are currently growing, but certainly several hundred, um, percent. And in this respect, the revenue multiplier, how the company is valued, is actually relatively fair, with one exception, and that is SpaceX, because they are only growing at 18%, they don't double or triple. Um, they have been reasonably profitable, um, but they show from this and the valuation of SpaceX. SpaceX is supposed to go public in June for 1750 billion. The valuation of SpaceX, 80% of this valuation was made in the last 12 months, and not by external investors, but by Elon Musk himself, who merged X with XAI, then XAI with SpaceX, and invented some valuations in the process. But I am sure that SpaceX will go public for 2 trillion and that the stock will even rise. I wouldn't invest there, but I'm sure the stock will, um, rise. Um, the question with all these IPOs is where the money will come from to finance these companies, to finance these IPOs. SpaceX wants to raise about 100 billion. That's only 5% of the company being sold on the stock market, so to speak. Um, and at a valuation of 2 trillion, that would be 100 billion. Uh, if you add the other companies that also want to go public this year, you get an open bill or a liquidity requirement of 250 billion. Um, but the money isn't there. So, where should the 250 billion come from? Um, all IPOs of the last four years combined have only raised 100 billion. So SpaceX alone will raise as much money on the stock market as has been raised in total through IPOs in the last four years. And what's left for the others is highly questionable. Who pays for all this? You with your savings plans. Um, this is, so to speak, the share that these three companies, SpaceX, OpenAI, and Anthropic, will have in these major indices if they go public, in a simplified representation. This means they would have 6% in the MSCI World, the most diversified index, immediately. You are venture capitalists from the day they go public. Even 8% in the S&P 500, over 13% in the Nasdaq. People who know about the stock market would say: "Yes, but there are rules, you have to be profitable, otherwise you won't be included in the index, and you have to have a high free float, meaning you have to have many shares on the market and a lot of trading activity, otherwise none of this will work." Um, that's correct, but SpaceX is trying to resolve exactly that right now. So the US stock exchanges are accommodating these companies massively, and the rules that were supposed to ensure hygiene in the financial market in the past are being massively softened. Especially SpaceX, as I said, is actually a very profitable company, has a positive adjusted EBITDA, many investments made will now be adjusted away, but nevertheless, SpaceX is certainly making money. But by saving XAI, meaning XAI was not bought because it was a super merger or because there are synergies, that's all bullshit, I think. The problem was that XAI is the big AI company with the least revenue. This means I believe Elon Musk was fully aware that he would not find new investors for XAI and therefore merged it with SpaceX, because there is a lot of investor interest in SpaceX and SpaceX generates surpluses, and this merger was, so to speak, the rescue program for XAI. Um, we're coming back a bit to the main problem of AI, which is still that you all don't want to use it, and your parents and your grandparents. Um, the actual picture is, perhaps some of you have seen this presentation, this presentation, hopefully you haven't seen it yet. Um, some of you may have seen this slide on the internet. Each of these dots represents 3 million people. The gray dots are people who have never used AI in their lives. The good thing is, this is the potential we still have. The green dots are people who have already used AI for free. The yellow dots are the few people who pay for AI to this day. Yes, this shows how small AI still is today, even though it's already a trillion-dollar market. The red dot represents people who have already built something with AI, who have built an app, a program, automated something on Excel. This means we are still at the very beginning, and if you have already done anything with AI or paid for it, then you are automatically in the top 3% of the world's population. Um, so to speak, it's a huge opportunity. Um, these green dots of AI users are still very deceptive, because many of those who started using AI stop after a month. Two-thirds, so this is a retention graph, which shows how many people are still there after 1, 2, 3, 4, 5 months, and after the first month, two-thirds of the people have already given up using AI or have switched to another tool, which can also mean they have changed their tool, so they are very disloyal. This means people are still experimenting a lot and are actually just playing around with AI. 75% of these green dots who use AI don't use it properly at all. That is, they have fewer than ten conversations a week. Yes, heavy users, someone who really uses AI, I think has at least ten conversations a day. People don't even have ten a week. This means they are actually just playing around and asking for football results or whatever they are doing. But in any case, they are not really using AI intensively yet. They are not using its full potential. The absolute majority has never seen an AI think for longer than a minute. Yes, creating an image or doing a proper research task all takes over a minute now. People today have virtually no access to this. Yes, so these green dots are totally misleading. Nevertheless, AI, these are Gartner figures. You can see the sources in the top right, by the way. Um, these are Gartner figures. Uh, AI is already a 2.5 trillion dollar market today. You can subtract the lower half. The purple is, so to speak, the data center business, i.e., the hardware and infrastructure investment, but even services, software, etc., are, uh, huge multi-multi-multimillion dollar markets. This is because companies, especially, are already integrating AI quite strongly. So these are McKinsey figures. They now see the adoption rate of AI in companies at well over 80%. What is also true, however, is that most companies are still in pilot or experimental phases. So, in the real scaling of AI, in the rollout, it has not really taken hold yet. This has led AI companies to come up with a few new tricks that have already been tried before. For example, in the future, more and more of their own developers and project managers will be sent into companies. This is a concept that Palantir, I wouldn't say invented, but has rolled out very strongly. Um, this means that because the customer doesn't know how to use AI, because the data structures are not ready, OpenAI and Anthropic will send their people there to do the work, like an agency, like an Accenture. Uh, this has been tried and tested very well with Palantir and has worked well, and they will learn from that. Um, then, of course, one tries through consulting firms, right? So at the beginning, it was said that AI would eliminate management consultants. It was realized relatively quickly that AI is creating a special boom for management consultants, because of course everyone is having AI implemented by management consultants. And the third thing, which has been happening recently, is working with private equity firms, i.e., with Hellman & Friedman, TPG, KKR, and all the others, because companies led by private equity generally have a higher drive for innovation, are more productive, and are more open to new technologies. This means private equity is pushing this a bit into its portfolio companies and has therefore built joint ventures with OpenAI and Anthropic to drive this forward. If you look at who is actually winning B2B customers, the picture has changed significantly in the last year. You see a meteoric rise of Anthropic during the last year. You also see that OpenAI is practically stagnating in the B2B market right now. These are, uh, financial data from Ramp. Ramp is, so to speak, a credit card and account provider for companies. They know exactly who buys what. Um, and they can see 100% whether more money is being spent at Anthropic or OpenAI. And OpenAI is not looking good there right now. In fact, Scott Galloway, I think, also showed the chart yesterday. Three out of four dollars that companies spend on AI are currently going to Anthropic and not to OpenAI. The industries that are, so to speak, most advanced in AI, i.e., Information Technology, i.e., software in the broadest sense, Finance and Professional Services, consulting sector, they generally prefer Cloud versus ChatGPT. Cloud users are also better paid, better educated, generally. This is of course also because developers in particular use Cloud a lot, and creatives, who typically earn more money. Anthropic has, I think, done many things right with its go-to-market, which explains this success in retrospect. Um, they have focused on the B2B market. B2B user companies are not so disloyal. This means they don't cancel from one day to the next. Once they have implemented something, it is much harder to get rid of it. Revenue expansion is easier, so getting a little more money out of the customer each quarter. The, uh, private user pays $ every month. That doesn't change so quickly. It's hard to get more money from them. A company will naturally use AI more and more. Um, you can, if you start with software, i.e., Anthropic has established itself very strongly in the B2B market and the software market. Um, with software, you can immediately measure whether something works or not. You have a faster feedback cycle. You also have excellent training data, i.e., in GitHub, Stack Overflow, in historical code that you can buy, um, there is an enormous amount of data hidden, so to speak, that you can use to build new software. The output is relatively structured and one-dimensional. So, if a word like "function" comes up, I immediately know. Only three or four things can happen now, and not, so it's much easier statistically to guess correctly. It's also good that the developers at Anthropic naturally understand the software market best because they are developers themselves. This also helps. So, it's much, much harder to build AI for medicine as a developer than for software, because I know software, um, well. And the most important thing is that at some point AI starts building AI. So there is this software self-optimization loop. Uh, Anthropic claims that last year it was still a big laugh here when I said they wanted to write 90% of the code with AI. And I said back then, watch out, no company understands how code is written better than Anthropic. And today they claim it's 100%. Whether that's true is another question. Um, but the fact is, it seems to massively accelerate the development speed at Anthropic. Um, I've marked all the days in this calendar on which OpenAI and Anthropic made a major update, uh, this year, and you can see that in April there was hardly a day without it, and these are not small things, it's either a new model, uh, a new industry, a new major feature. Um, a lot is really happening, and every time that happens, um, entire industries go down the drain. So Anthropic has become a bit like the new Amazon. Previously, when Amazon entered healthcare, the stocks of healthcare providers went down. When Amazon said something about banks, then bank stocks lost value. Today, it's like Anthropic generates or produces a new feature. And if they say, for example, we're building something for lawyers, then the legal tech providers Thomson Reuters, Walters Kluwer, LegalZoom, and so on, lose double-digit value. Billions are destroyed because Anthropic has built a new feature. And especially Cloud Code, the programming tool from Anthropic, has become so good that the startups that once made money with AI programming in the past, they are still making money, I don't know if they are making money, but they are still offering it, but they are no longer growing. Since the emergence of Cloud Code, these companies, which were previously absolute high-flyers, have stagnated significantly. So Lovable, Cursor, Replit, and the German automation tool, uh, Nordomation, I believe it's called. Um, they are stagnating relatively strongly. Uh, this is web traffic. Uh, this is relatively representative, uh, of demand. And if you look at the cohort tables, they don't look good right now for other tools, except, um, Claude Code. Um, you can't ignore that more and more code is being built with AI. To doubt that would be nonsense. You can see it from the output alone, how many more new websites are being created, how many more new apps are being uploaded to the app stores, there's a very clear trend right now that this is massively increasing, and it's of course because everyone can now build products themselves with AI. GitHub, where developers upload code, is currently experiencing an explosion of new commits, i.e., new contributions where new code is being sent. And the reason for this is, of course, that this code was generated with AI. The good thing about the B2B sector, as I've already said, is that customers are more loyal. They don't give up the tool so quickly, uh, as in the consumer sector, and you can see that very well in this retention curve. In 2022, when ChatGPT was just released, most companies, half of them, gave up after a year and stopped paying the bills. So they said, we tried AI, it's not for us anymore. Now, 80 or almost 85% of companies are sticking with it. This means AI seems to be getting better if more and more companies are not switching tools or abandoning the AI project. And what we hope for, of course, are retention curves like those from software. So, it's called a smiling retention curve because at the beginning, losing between 20 and 40% of users is totally okay. For some, the product just doesn't fit. But the users who stay should spend 20% more each year. Then you get to the point where this smile in the retention curve emerges, and thereby the growth improves, so you grow through new customers and because existing customers spend more money each year. And I believe what we will see at the IPOs of Anthropic and OpenAI at the latest, rather at Anthropic, I believe, is that this smiling curve will look like this. Um, a revenue expansion of 120% is expected. This means the company will spend 20% more in 2027 than in 2026 for software. With AI, it will look like this. This is also real data from Ramp again. This means in 2023, the average AI contract was still worth $40,000. Now, the average company spends $140,000 on Anthropic or OpenAI. Last year it was already half a million per company, and this year it's expected to be a million. And it's not because you sign larger contracts when you start, but because once you start using AI, the costs explode on their own because you realize what you can do with it. The software also needs more and more tokens, the bills get bigger. People are now considering hiring people again because AI is so expensive. Um, but this revenue expansion will drive Anthropic's revenue brutally, and OpenAI's in the B2B business as well. You can also see this quite well with Google. This is now quarterly, not yearly. Um, Google always releases how many bill, how many million, excuse me, how many billion tokens are generated within a minute. Tokens are, so to speak, the output product of AI. Um, and this has increased by 60% in the last quarter alone. So, on an annual basis, probably around 300% output growth. And token growth usually also means revenue growth. Um, so we will see unimagined revenue expansion in this area. What is astonishing, when you know that only one in five employees still has access to a chatbot from their employer. This is actually the saddest figure, but at best it shows how much potential is still there.
is stuck. We are now at slide 70 out of 150. There are still 20 minutes left. My favorite slide, I've been showing it for about 3 years every year, and I won't stop showing it until I believe it's no longer true. I said back then that having the best LLM isn't that important. Much more important are your own data, hardware, and distribution. I want to explain two things about it. Firstly, why Google has become so good again in AI, why the Gemini model has become extremely good, and secondly, where we stand in Germany. I believe Google has perfectly leveraged these advantages: hardware, data, and distribution. Google started building its own AI chips 12 years ago, 13 years ago, and that's why they are the only ones who can offer AI super cheaply, for free. Essentially, anyone currently using Google gets AI Overview for free or can use the AI mode. No other company could subsidize this permanently, I claim. Google undoubtedly has the most data, we don't need to talk about that, there's no company in the world that has more data than Google, and it has nine apps with a distribution of over a billion people. Yes. Every company would be happy about one of these features. Google has all of that: its own chips, the most data, nine apps used by over a billion people. Furthermore, with Deepmind, they have a Nobel Prize-winning research lab, and they have the cash flow to sustain this, to build new data centers, and so on. The problem with Google's rise and its subsidization of AI is that the rest of the internet is slowly dying out. To be clear, Google would strongly disagree with this statement. Therefore, one should look at the numbers. The blue line represents the revenue Google makes from web search. The red graph is the revenue Google makes from the rest of the web. So, when I integrate Google ads, that's the so-called Google Network Revenue, Google's display business. This is the only business line at Google that is continuously shrinking by 3-4% per year. And this is because the rest of the web no longer receives traffic from Google, because these AI Overviews, no matter how you optimize for them, nobody clicks on the links there, or 1% of people click on the links. This means it doesn't help the rest of the web. At the same time, AI bots are completely destroying websites by scraping them. This is data from Cloudflare; we calculated how often a chatbot visits your site before a visitor arrives. In the case of Anthropic, the Cloudbot visits your website 71,000 times before sending a single visitor to you. Yes, if you use Claude, it's not easy to find a citation, a reference, or a link. The content is simply reproduced, regardless of its origin. They might write where it came from, but why should I visit the website anymore? The numbers don't look much better for the other players; Claude is probably the worst in this regard. Let's look at Germany again. This is the model performance by country. Germany is shown in yellow. We don't have a large LLM. I said I wanted to describe our situation in Germany again. So, in the LLM league, we have nothing. We've just sent Shopic to Canada, we've just sent Alpha to Canada. We definitely have no distribution. We have no own hardware. 90% of the world's hardware is manufactured in Taiwan, on this small island. Most people don't know this. Most people don't even realize it's Taiwan and not South Korea. Who knew that? Hands up. Yes, you don't know either. That's actually the Philippines. That's Taiwan. But in any case, nine out of ten AI chips are manufactured there, and they can't just be manufactured in Germany. A few can be manufactured in the USA, but we won't solve the hardware problem in Germany. The only solution we have is to hold onto our data and ensure that we generate value from this data, and that it's not US platforms that contribute minimally to taxes in Europe that do so in the future. The good news is, we might not even need our own model because open-source software could solve this for us. Here you can see quite well the blue models, the open-weight models, or open-source models. With a slight delay, they become just as good as the commercial models, and we can use them for free. Here again, in a different representation, blue is open source again. The problem is, they all come from China. This means we now have to consider whether we want to work with the big US platforms or with free Chinese models. Many countries in the world have already made that decision. The red countries, like Cuba, Venezuela, whatever you want to call them. In any case, there's a kind of Warsaw Pact there, where they only use Chinese models and not Western models. We'll have a laugh again because it was so well received. Last year in China, there's a model called SetGPT because it has so perfectly internalized the Chinese party's politics. And that shows the problem a bit, because open source isn't necessarily truly open. The openness of a model is defined, among other things, by its transparency. Whether I know what it was trained on, where the data actually comes from. You wouldn't take a medication without a leaflet, would you? I want to know if something was adjusted during fine-tuning, where the data comes from at all. And truly, truly open are actually only the models. The one model from Nvidia on the top left is essentially Nvidia's best, and the K2 model, the dark blue dot, is MBZUAI. MBZ stands for Mohamed bin Zayed, so it comes from Abu Dhabi and is probably not better either. Nobody invests more in models than the USA. The USA invests about 20 times as much as China and 60 times as much, yes, 60 times, 70 times as much as Germany. The problem is that China, with 1/20th of the cost, gets relatively close to the performance of the USA, or becomes just as good with a slight performance lag as the USA. And what China does is essentially distill the US models, extract the knowledge, and then offer it cheaply. I've compared models here. The one on the left is Claude, Claude Sonnet from Anthropic. It costs $15 per output token, so $15 per 1 million output tokens. The same from DeepSeek costs less than $1, the same from Alibaba Qwen costs less than $1. Yes, this means many startups are no longer using the expensive US models, but are waiting a bit or are satisfied with slightly less performance and are buying the China discount version. Here again, the costs are shown. On the far right is the latest ChatGPT, on the far left is DeepSeek, so it costs 100 or less than 1% for output if I'm satisfied with slightly less performance. These are many companies; this is usage data from OpenRouter. I can control models via an API through it, and you can see that the top spots, or many of the top 10, are now occupied by China. And China is becoming the world's token factory, replacing AI at the cheapest prices, and that's the biggest problem I see for US companies right now, honestly. Also present here is Grok from Elon Musk. Why is Grok so cheap? It turns out Elon Musk had to admit in court recently that he too distills models. So XAI is also stealing OpenAI's intellectual property by distilling from the models, just like China does. And the question is, when will China essentially put on the right blinker and overtake? Right now, they're always lagging behind, as they might have done in the past in production or with cars. But at some point, China might overtake and then be cheaper, faster, and better. The good news, the green bar, is the performance of the best model currently running in the world's largest data center. The red line is the performance of the best model that can currently run on your gaming PC. With a delay of 6 to 7 months, you can use on your home PC something as smart as what would have been running in a data center half a year ago. This should lead to a lot of dispersion, a lot of freedom, so that everyone can work with AI. It is logically the cheapest option and the one where you are least dependent. We'll talk a bit more about adoption. The different models are getting better and better. Google stands out here, having become significantly better. Last year, looking at the past year, Google Gemini has improved the most from a relatively small base, but you can see how the market share of Google Gemini, this is again SimilarWeb data, is really eating into ChatGPT's market share. So ChatGPT is still the market leader for private users, but Google Gemini has taken a lot away, and we also need to question these numbers. ChatGPT says, or OpenAI says, they have 900 million users. The question is how they can measure that, because Nick Turley said on the BG2 podcast that the best growth hack they did was removing the authentication wall. If you work a bit in marketing, you know how incredibly difficult it is to track people when they no longer have to log in. So, how can OpenAI know, if I use ChatGPT on the iPad, then in an incognito browser, then on my phone, how do they know who is a user and who isn't? So, how do they measure their weekly users? And just as easily, Google could say they have 3 billion users because everyone who uses Google uses AI nowadays. In general, there isn't one market for AI, but many different markets. I would divide the market into consumers, professional users, and companies. And then, I think, one must further differentiate into whether I want to use it for free or have it included in a package, whether I want to pay a subscription for it, or whether I want to pay based on usage. It's relatively clear that Shopic is currently leading in the usage-based B2B market. I think it's also clear that Google Gemini will win because only Google can offer AI for free permanently, because they have a super strong advertising model behind it, because they can offer it cheaper than anyone else with their own TPUs. Other winners might emerge in other segments. In companies, Copilot might win, just like Teams won, simply because it's inevitably provided. For individuals who are private users and want to pay, ChatGPT might be the leader. In principle, Shopic is pushing people forward very strongly. Everyone is currently narrowing their strategy. OpenAI is doing fewer side projects and is also trying to move more into the B2B sector and copy Anthropic's strategy. At Google, a strike team has been formed to improve coding, to catch up with Anthropic's advantage or lead. So Anthropic is really setting the market. And if you look at the latest numbers, you can see that this year, in February, Claude from Anthropic actually gained the most traffic. This is probably due, in part, to the dispute between Anthropic and the White House, where they said they don't want to provide software to the Pentagon that the Pentagon could use to spy on American citizens or make lethal decisions on the battlefield. Nevertheless, Anthropic is integrated into Palantir, and Palantir and the Pentagon have likely used Claude in Venezuela and Iran in the past, based on available information. Why this is important: I want to point out that the restriction is for US citizens, meaning they only want US citizens not to be spied on with the help of AI. Since Edward Snowden, we know that the whole world is actually being monitored. Our conversations can all be overheard. The only problem is, nobody can evaluate all conversations in the world. Nobody could evaluate all conversations in the world because AI is the technology that will make it possible for every conversation, every text, every video chat, every phone call to be fully automatically evaluated by AI. And I believe that's the reason why Anthropic is resisting it, but they are only resisting it for US citizens. Whether our conversations are not already being automatically evaluated by AI is a huge question, and we will hear a lot about these problems in the future. I recommend the book "Surveillance Capitalism." Exactly after Anthropic resisted the Pentagon, they saw a storm of signups for the app. That was the best PR campaign they've ever done, that can definitely be said. Last year, the graph is about a year old. Last year, OpenAI was still far ahead in revenue compared to Shopic. Today, Shopic is about to overtake OpenAI. Below that, you can also see ZI, XAI, Mistral, and so on, they are also playing along. But this is a bit misleading because the scale here is logarithmic. In reality, it looks like this. This means there are these two companies. The others have long given up or are no longer competing in terms of revenue, but what you see is that Anthropic is massively gaining revenue against OpenAI. If you were to shift Anthropic a bit to the left, as it started a bit later, then Anthropic would be there today. And while I was building this presentation weeks ago, new news came in over the weekend, namely that the company Semianalysis reported that they believe Anthropic would now be at $44 billion in revenue. They built two-thirds of this revenue this year alone. Yes, they are really, if there's a hockey stick mode anywhere, then this is it. Whether the numbers are true, we don't know yet. We'll hear about that in the future. But Semianalysis understands very well what chips are used for and should have a good insight. Also, per employee, these companies are earning an incredible amount of money now. OpenAI makes about $5 million in revenue per employee, compared to what other companies made when they were this old. Shopic makes about $8 million per employee. Now comes the obvious question, what about jobs? If one employee can make so much money. Rishi Sunak in the UK, the Prime Minister, warns of fewer jobs for young people. Others see entire careers in danger. In fact, the hiring rate is massively decreasing in many jobs. There is this chart that has instilled fear in many. It has many errors. Two are in the chart, two are behind the chart. The two that are in the chart are these two lines. So, they are supposed to say that juniors will have a harder time than seniors in the future. What you can see is that this development started before. AI cannot explain this. It has nothing to do with AI. The other is that this COVID hump you see there is a bit distorted. It would look much smoother and less threatening if you removed the COVID effect. The other two things hidden behind it are that this chart applies to the 3% of companies that have already adopted AI first, i.e., startups and so on, and therefore doesn't affect the other 97% of companies at all. There is another study called "Canaries in the Coalmine." Previously, canaries were taken into mines because they died earlier from CO2, carbon monoxide poisoning than miners. And the idea was to say, this is the first sign that AI is destroying jobs. This also concerned junior jobs. The problem is, the study couldn't explain why companies saw that they had to lay off people 6 months before ChatGPT. So, it's hard to say ChatGPT is to blame if people are being laid off half a year earlier. What is much more likely is the interest rate hikes in the USA. During that time, interest rates were increased from 0.25% to 4% in the USA. This is a much more likely indicator of why these jobs were lost. If you look at the interest rate curve overlaid, it looks like this, and this seems to me to be a better indicator for this crash than the launch of ChatGPT. What the scientists also realized, why the study is now called "Canaries Interest Rates and Timing." You just didn't read the correction in the media because it's not that exciting. This entire wave of layoffs is also because many tech jobs are actually disappearing that were over-hired during COVID. So, they hired too many people, whom they now need to get rid of. This is much more important than the effect of AI. Here you can also see well, especially tech layoffs, they start long before ChatGPT. This is because tech companies in particular are much more dependent on interest rates and financing than others. That's the good news. The bad news is, none of this helps you at all. Because this is the actually important study, it's from 2010 and it says it's always bad to come out of university when the economy is doing poorly. That is the actual explanation. So, one way or another, it doesn't look good for you. Another forecast, this is a Goldman Sachs study, which says legal, i.e., lawyers, are particularly threatened by AI. 44% can be automated. Never before have more lawyers been hired than right now. And what also doesn't fit the narrative that your jobs are all gone is that precisely since Claude Code was launched in May 2025, hiring of developers and customer service people has increased again. This doesn't fit either. Nothing makes sense. Nevertheless, tech companies have laid off an incredible number of people, and the reason is that these people had to leave because of AI, but not because their jobs are being taken by AI, but because the money that pays for their jobs has to flow into data centers. Companies no longer have money to pay for data centers, and therefore they are laying off people to become leaner and invest more money in data centers. What is true is that it now takes significantly fewer people to build a unicorn. That means startups, so we will see a separation of the economy. Old companies will be incredibly slow in implementing AI and freeing up people, as they did with digitalization, while startups will not hire these people at all. This is also because it is much easier to build a unicorn, as more money is now available. Let's skip that. What you also need to be aware of, if you truly believe there are people who still go around saying we're not hiring juniors anymore. Juniors are naturally the generations that use AI the most. It's not the Silent Generation, the Boomers, who use or trust AI, but it's the juniors who understand this technology. And I believe if you are faced with the decision of whether to hire Günther or someone young, ambitious, who is much more involved with AI right now, I would tell you that you absolutely must hire the young person. Nothing is worse than not hiring people right now. We saw this with digitalization. If I don't hire people who know about Google, about Facebook, my company won't have the important know-how. Amazon has recognized this; Amazon is hiring 11,000 new interns. So, nothing is worse than only hiring old people. It's like going to a newspaper kiosk and saying, I want to pay double for yesterday's Hamburger Abendblatt. You need the new people. So, that nobody wants data centers, you've already heard from Scott Galloway. We won't talk about that. We will build approximately 80 new nuclear power plants; all of these are needed to build all these data centers. You can also calculate this quite well if you simply look at Nvidia's forecasts. Nvidia certainly wants to grow by at least 40%. Currently, they are growing by 60%. And if Nvidia continues to grow like this and you simply extrapolate how much energy these chips need and the data centers around them, you arrive at the same figure, over 100 gigawatts. However, Nvidia no longer builds all AI chips. Half of the AI chips, by unit count, are now built by Google, Amazon, and other companies. So, that's also a bit at risk. AI brings entirely new cyber threats. Large companies will be hacked. Anthropic's new Mythos model is considered too dangerous to unleash on humanity. Only governments and large corporations will have access to it. And so-called zero-day exploits, meaning a vulnerability is discovered and immediately exploited, are massively increasing. So, it takes less than a day on average before an open security vulnerability is exploited. This means every software will be immediately vulnerable in the future. I'm almost finished now. Yes, these are all the topics I didn't manage today. We're still through, 15 minutes. I'd like to thank my team very briefly. Jan built two slides off-screen. Many, many thanks for that. With Claude. Actually, Claude built it; when I say Jan, I mean Claude. But otherwise, I've never used as much AI as for this presentation. Two small tips for you, and then I'm really done. I can say all of this. I'm very relaxed. I was recently at a company that wanted to hire me as a consultant, or was considering it. I told them they shouldn't do that. You can build your own board of directors with AI today; you can all do this if you have a startup, no matter what kind of company you have. You can take people who are already dead, like Steve Jobs, people who are soon to be dead, Warren Buffett, you can take the best advertiser in the world, you can take people who never existed, Harvey Specter, or someone who has been dead for centuries, and build a board of advisors from them. If you don't know how to do that, you can all do it. Just ask Claude how to do it. Go to Claude and say, Philipp Klöckner said I can build a board of advisors from dead people. How can I do that? Claude will explain how it's done. And many important personal decisions are also coming up for some of you this year, and you can build your own election advisor or ask the AI if you don't know how you want to vote. For example, I once asked the AI, you are a senior researcher at a leading economic research institute. What would happen if the AfD wins the federal election and implements its economic program into policy? What would happen to economic growth, unemployment rate, inflation, purchasing power, export volume, inequality, domestic demand, exchange rate, insolvency rate, political disillusionment? And the result is, as one would expect, I would say. The insolvency rate will, thank you, the insolvency rate will partly double, triple. Almost everyone agrees on this, except for Elon Musk's little whiner, almost everyone agrees, but even Grok knows that we will fall into recession and that the unemployment rate will rise. So, we will have double-digit unemployment rates again. We will have double-digit inflation, some models believe. Everyone is sure that it will be the grave of the German economy. So, I don't give a damn if you are the last heartless, heartless assholes, but don't vote for this party if you care about the economy at all. Here are the slides. I will post the slides on LinkedIn in the coming days. I hope a video of the talk will be released. You will definitely get the slides if you follow the newsletter, follow our podcast, follow me on LinkedIn, or follow the Substack. Thank you for the extra 4 minutes. Philipp Klöckner. Wait, Philip, I have a very quick question, just before you all stand up and leave. Number one. I am always impressed by how you manage to share so much information so quickly. You've gone through 148 slides again this time. And who is in favor of us having a Philipp Klöckner Hall at OMR 2027? Philipp Klöckner, thank you very much. Great again. Thank you. Thank you very much.