Rise of AI Conference 2025 · Berlin · Applied AI Stage

The AI Tipping Point

Gilles Backhus · Head of AI, Recogni

At Rise of AI 2025 in Berlin, Gilles Backhus, Head of AI at Recogni, argues that AI adoption is approaching a tipping point defined by cost and speed rather than raw capability. Recogni builds AI chips and data-center systems, which gives the company a bottom-up view of what AI actually costs. Backhus explains that global AI compute is shifting from training to inference, projected to reach 95-99% inference by 2030, and that inference costs for advanced multi-model agent stacks have risen roughly 100x per million tokens because accuracy compounds negatively across chained models. He walks through enterprise dilemmas around where to integrate AI, what it costs, and how to manage internal and external change, citing Shopify and Klarna's hiring reversals as examples of unstable dynamics. He introduces real-time GenAI, especially real-time video generation, as the next unlock, using a hypothetical $18 trillion AI-education market to illustrate current costs of roughly $100 per hour of AI video. He closes by describing Recogni's approach to cutting cost 10-100x through logarithmic math, faster chip-to-chip networking, and purpose-built systems.

Key takeaways

  • Global AI compute is projected to shift to 95-99% inference over training by 2030 as models grow too large for smaller companies to train.
  • Chaining multiple 98%-accurate models into an agent stack can drop combined accuracy to about 90%, driving up model prices roughly 100x.
  • Shopify and Klarna both told staff to stop hiring and use AI, but Klarna partly reversed course and rehired human service agents because top AI remains too expensive.
  • An hour of real-time AI video currently costs about $100 because a $5 million Nvidia NVL72 rack supports only a few simultaneous video streams.
  • Recogni's logarithmic math approach replaces costly multiplications with additions, cutting AI compute cost by roughly 4x at the transistor level.

Detailed notes

Recogni's bottom-up view of AI cost

Backhus introduces himself as co-founder at Recogni, an AI chip and systems company that he describes as, in a broader sense, an AI infrastructure company. Because its data-center rack systems cost millions and take years to build, Recogni has to understand both the true cost of AI and where the technology is heading. He notes the company has innovations such as logarithmic math and a hyperfast chip-to-chip network that make its systems more efficient than Nvidia's, without going into technical detail.

From training to inference, and the accuracy problem of agent stacks

Backhus recalls the early excitement with simple LLM tasks giving way to the realization that much still does not work. He says that in 2023 global AI compute was dominated by training, but this has flipped and will keep flipping: by 2030 he expects 95%, possibly 99%, of global AI compute to be inference, since models are growing too large for smaller companies to train. Everyone wants the best models with mixture of experts, reasoning and multimodality, and compute cost per task is exploding. He gives an example: a text-to-text model that is 98% accurate in isolation drops to about 90% accuracy when five such models act together as an agent stack, which is why some models now cost $50 per million tokens instead of 50 cents, a roughly 100x price increase.

Enterprise adoption dilemmas and hiring dynamics

Backhus describes the questions enterprises face: where to integrate AI, what it will cost, and how to manage change internally and externally. He says internal developers often reject a hardware vendor, including AMD, because they do not want to learn a new programming language, which is part of why Nvidia has such a large market share. He cites Shopify's CEO telling staff to stop hiring and use AI instead, reducing hiring of junior computer science graduates in favor of senior developers using AI, and the Klarna CEO's similar call that was later partly reversed by rehiring human customer service agents because top AI is still too expensive to use everywhere. He warns that companies who delay AI adoption risk falling behind competitors who have already raised their margins through it, and suggests looking for new products rather than only AI-ifying an existing business.

Real-time GenAI and the cost of video

Backhus introduces the term "real-time GenAI," describing how tasks such as five-second video clips or complex reasoning by ChatGPT o3 currently take minutes rather than seconds, and how Recogni is building systems to bring this down to seconds. He argues that once video becomes real time, interactions such as an AI twin joining a video call become possible, and describes video-to-video as the last modality needed for AI to understand a highly visual world. As an illustration, he estimates a market of $18 trillion if every one of 1.8 billion students received two hours per week of AI video-based teaching at $100 per hour of AI video, calling the number completely unrealistic but useful to show how much cost still needs to fall. He also cites AI companionship and therapy as already the largest GenAI use case.

Why an hour of AI video costs $100, and how to cut the cost

Backhus breaks down the economics: an Nvidia NVL72 rack system costs roughly $5 million, runs at about 100 kilowatts, and can support only a handful of state-of-the-art video streams at once. Over a five-year depreciation period this comes to about $1.1 million per year, or roughly $125 per hour at full utilization, close to the $100 per hour price companies charge with margin. He identifies three levers to reduce cost 10-100x: keeping high numerical accuracy (16-bit rather than 8- or 4-bit) rather than shrinking models, increasing speed since usability often follows a step function once an application becomes fast enough, and cutting raw hardware cost. Recogni's approach includes logarithmic math, which replaces expensive multiplications with additions and cuts cost roughly 4x, a partnership with Juniper Networks on ultra-fast chip-to-chip networking to parallelize across up to 144 chips, and purpose-built systems stripped of unneeded generality.

Q&A highlights

Asked which industry will transform most in the next two to four years, Backhus says therapy and companionship, partly because errors matter less there and it is already, per some studies, the largest GenAI use case. Asked about the state of chip development in Germany, he says Germany's great silicon days are largely over but notes some startups working on analog AI chips, and that Recogni itself is split 50/50 between Silicon Valley and Munich; developing its first chip cost $25 million, which he calls cheap by industry standards. Asked for advice for young people, he says being a specialist in one thing will matter less than becoming an excellent user of AI tools. On education, prompted by an audience member from Exponential University, he agrees it is a major opportunity, particularly for bringing translated, localized instruction to rural areas, though he expects parental concern to fade over roughly a decade. Asked where Recogni will improve most in the next three to five years, he says software, despite being a hardware company, because usability is why Nvidia dominates over technically comparable but less usable AMD chips.

Quotes

It's a giant mathematical formula with hundreds of billions of variables. That's really what an AI model is.

Gilles Backhus

You'll be looking at this paradox that if things become cheaper actually it'll be used more and not less.

Gilles Backhus

The great silicon days of Germany are a little bit over.

Gilles Backhus

Questions answered in this talk

Why is AI compute shifting from training to inference?

According to Gilles Backhus at Rise of AI 2025, models are growing too large for most companies to train themselves, so by 2030 he expects 95%, possibly 99%, of global AI compute to be inference rather than training.

Why do multi-model AI agent stacks become less accurate?

Backhus explains that if five text-to-text models, each 98% accurate alone, act together in an agent stack, the combined chance all of them behave correctly drops to about 90%, which is why capable models now cost far more per token.

What is real-time GenAI and why does it matter?

Backhus defines real-time GenAI as generating outputs like video or complex reasoning in seconds instead of minutes, which he says is a step-function unlock: once an application becomes fast enough, such as real-time video-to-video interaction, it suddenly becomes usable in ways it was not before.

Mentioned

Inference vs. training compute, Mixture of experts, Real-time GenAI, Logarithmic math, Chip-to-chip networking, NVL72, AI agent stacks, Tokenomics

Keep exploring

More from Rise of AI Conference 2025

More on Enterprise AI & Adoption

Transcript

Auto-generated transcript, not edited. · 5,244 words

Show the full transcript

co-founder at Recognize and we're building a AI chips and systems. So, we're a hardware company. Uh, in a broader sense, you could say we're an AI infrastructure company. So, we look at the cost of AI from the bottom up. Uh, we're building systems like this one, which go into data center racks. They cost a lot of money, millions. And so, uh, we believe we have a fairly good understanding of the cost of AI. Uh, and also some other things that make AI successful. And of course uh because the the time from conceptualizing a system like this all the way to launching it is actually years. So we also always have to have a very good understanding of where AI is heading.

So these are the topics that I want to talk about today. Um and uh yeah we have a couple of innovations that make our systems better than Nvidia um such as logarithmic math connecting a lot of chips in a hyperfast network bunch of other things. I'll be mentioning them briefly but I won't go into too much technical detail today. Um now if we look at uh the AI explosion of the last years um yeah I think we all remember how it started right you were super excited and happy that you could tell a little LLM tell me a joke and it actually was able to come up with something that if a human would have done it we probably would have said that's quite creative uh and then we kind of uh you know came down from the party and realized okay there's still a lot of things that actually totally don't work.

uh and yeah um the the important part is that uh the compute is exploding um I think uh 23 um the global AI compute was absolutely dominated by training not inference inference being uh you've already trained a model and now you just execute it and you're making money with it or that's the hope at least um but now it's flipped right and it will flip even harder um as we as the next years will pass until 2030 probably 95% maybe even 99% % uh of the global compute for AI will be inference not training anymore uh because the models are getting larger and larger and uh it'll it's becoming harder and harder for smaller companies to even train models that's very different from the previous age of AI where very often it was about vision and uh everybody with a little gaming graphics card could could train their own model these times are over so inference is the big market and you can see that the compute and these are fairly realistic numbers it's really

exploding right so for a single task. A task could also be tell me a joke, but a task could also be uh please analyze this huge report and then make a buying decision based on this or so for high frequency trading company. Um the the cost of compute is exploding. Um and ideally we want always we always want to use the best models with mixture of experts and reasoning and multimodality and multiple models acting together. Right? That's what you ideally always want to do. We have not yet reached a ceiling of AI capability. Everybody wants to use the best model ideally as long as the price uh fits. Um interesting things are going on right now in the industry and some of you might have read these headlines in the last couple of months that uh the Shopify CEO for example told its people stop hiring uh rather use AI.

Um for the first time um these graduates from uh from universities who learn to be a computer scientist um are not being hired as much anymore. Uh but uh the idea now is if you have a senior developer who understands what good code looks like um they can use uh AI which kind of then replaces the junior programmer and you don't need these young people anymore. Uh so very weird new dynamics are going on. And then um what's quite funny is that the Clar CEO also said stop hiring, use AI and now they actually went back to uh hiring uh human and customer um sales agents uh or service agents. And uh but that's also because the best AI is way too expensive currently to always use it.

And so we're a company that's trying to uh address that. Um the big questions that we always see um enterprises asking themselves is um where do I integrate AI? Do I do it for all the internal stuff? Do I actually put it into my product? Um right, this is a big question. Uh what will it actually cost me? Definitely more about that today. And um how do I also change manage internally and externally? Uh we constantly because we're an infrastructure hardware company, we we very often hear even from huge companies that are optimizing the hell out of their production AI stacks. Uh we hear from them that their internal developers do not want to learn any new programming language to you know deal with AI models.

And that's a simple reason to uh reject um a certain hardware vendor. um and not necessarily us also other huge fir firms like AMD who are not as good in software as uh Nvidia for example. So you're also dealing with these kind of things as you know you're turning into an AI company. Um and even companies that are absolutely already considered an AI native company are still struggling with these things. And of course, external change management also needs to happen, right? If you suddenly switch your customer service from human to AI, um chances are that if you don't understand your demographics really well of your customers, um you will have a couple of maybe more senior folks who will be very weirded out by that and might leave you and you you're losing customers.

Um and then also, am I too late to the AI race? Um the question for that is maybe but you should nevertheless try otherwise try to find out what maybe to rather focus on on like we see a lot of companies that are just trying to take their current business into the AI age and sometimes it might actually be better to not try to do that because their competitor already started doing that one and a half years ago and they kind of didn't know about it because it was a little bit stealthy and so on and uh now they're just trying to catch up and and the competitor who has already AIified a lot of their stuff like customer support and so on has increased their margins right and it's really going to be really hard for them to catch up so also sometimes think about uh what else could I do with AI right that my competitor is not doing maybe create a new little business branch some new product idea or so on

um and then also a problem that we as a company deal with uh every day can I trust the model's output quality um with agents right I mean there's this whole idea of AI high co-pilots um the a co-pilot normally uh does not have to be perfect right because the pilot the human in most cases today um can take over uh but if AI agents become actual pilots um mistakes become very u relevant and so you need to pay a little bit more attention to accuracy in models and to to model quality um now text to text to image these are things that most of us understand and and use maybe even every day uh in AI.

But there's a couple of new things on the horizon. Um AI agents, of course, everybody knows that word. Everybody is maybe, you know, without even realizing it, using it every day. Uh lots of companies are actually building them. Uh what's important there is um if you have a model that for example in isolation, let's say some textto text model is 98% accurate. That sounds really good, right? Um but if you have five of these models acting together um the chance that all of them are doing the right thing together as a multimodel agent stack uh then drops to 90% and then suddenly it's rendered useless. So this is why you're you're seeing models that instead of costing 50 cents for a million tokens as it might have been maybe a year ago quite normal um now you're paying $50 per million tokens.

So the price has increased by 100x even though we would expect in a in an exploding a technological field that price also comes down a bit. Now for the existing models yes it does come down over time but um the models are actually getting bigger and bigger. So the hardware that people develop needs to also be able to catch up with that or keep up with that and really truly enable it. The next thing uh and this is a term that we as a company are trying to put out there uh you probably don't see this otherwise very often is instead of geni real-time geni and what we mean by that is um right now you have a lot of things that take minutes um so for example if you want to create a 5-second video clip with an AI model it takes minutes instead of these 5 seconds it's not real time uh if video became real time then me in a business call uh when I don't want to join that video call

at during work because I'm actually busy cooking my lunch right now. Uh I could have my AI twin join that call and instead of just being a fairly fixed avatar that is very minimally animated, it could be a completely every pixel is generated by an I model. But today that's impossible. We don't have the machines to do that and the cost is actually extremely high and I'll show you an example of that later. Um and and also reasoning, right? If you use chatp03 as part of your $20 a month subscription and you let it handle some really complex task, it might come back after minutes. Uh we're actually building systems that can reduce that to seconds.

If that becomes a reality, you as a human worker can interact with these systems in a completely new way, right? Instead of, you know, hitting, okay, now submit the task and now I'm going to wait for a couple minutes. Hm. What do I do? Maybe I take a sip of coffee. I go for a smoke or something. No, now it's going to come back fairly quickly. it's going to be a much more immersive um experience uh in the B2C, B2B, in the business world, in the private world. Um and this is also uh something that um we also as a company want to enable the increase in immersion. Today, talking to an AI or interacting with an AI model is mostly through text and audio.

Um in the future, you'll have real-time video to video and that future is not that far away. um the NVL72 system by Nvidia and our system will be the two only systems that can accelerate this that have enough chips in one network so that you can run it fast enough to even generate a video stream in real time. Um and uh that of course will have huge implications. Genai I think this or next year um the number one use case is now therapy and companionship. So people are using ChatGPT to have a friend uh during uh difficult times in their lives. when this becomes an actual video instead of just text and maybe audio, the immersion and the relationship you'll have with that AI will be completely different.

That sounds like a very nichy private thing, but also in the business world when you when you do sales, right, or customer service or um anything like that or even when you talk to a colleague and it's actually an AI twin of you doing that right now, um you also have a sort of relationship. Sometimes it's a micro relationship that just lives during that one day or that one conference interaction. Um but AIS once they get come to video they will be able to replace us humans way more than it is currently the case. Um replace is a very bad word. Uh but let's say um enhance or um yeah ex add to to our existence basically in the work world and in the private world.

Um now let's look at one really crazy example and of course this is a market estimation that is completely BS and it will never happen. it'll never become 18 trillion because that would simply be too much. But let's say we were to tackle education with AI teachers and the AI teacher is actually video rendered. So it would be a video portal uh and every student on the planet for just two hours per week which is absolutely a realistic amount of AI usage that most students are already doing today. You would get uh an AI to talk to you and it's a video stream, right? The advantage of that is uh they can interact with you with much more emotions.

Uh they can also see you because it's video to video. Uh you can maybe um suddenly they can generate a block diagram, right? All kinds of visual input um that maybe a human could not even quickly whip up. And um how many students do we have on the planet? 1.8 billion. Uh so 100 hours per year, $100 to create an hour of AI video. That's actually a very optimistic price. Uh today very often would be actually more than a thousand. Uh we're not even creating an hour of AI videos these days. These models have just come out that could actually do that. Uh most of the times just 5 seconds or so of AI video that we create.

But that would be the cost. And then you're actually looking at an $18 trillion market. Obviously, as I said, completely unrealistic. This will never happen. But it's just to show that first of all, the cost of AI still needs to come down massively. Um because video is that last modality that we want to unlock in terms of interacting with a with an AI model and also for AI models to understand the world. The world has been built for us creatures and by us creatures and we're very visual creatures. So um unless vision is unlocked, AI robots etc will not be become a thing. Um it has many many applications and uh this is really to to show we're at the very beginning of AI adoption.

Um now real-time video based AI is a fairly small market today but it's growing very quickly. These are the estimations or basically the the entire total market here includes things like gaming industry um companionship therapy coaching education uh medical consultation consulting in general um things like that. And um and as this market will grow the cost for it will come down. the hardware will be released that can actually compute it and uh but the models will still grow in size and you'll be looking at this paradox that if things become cheaper actually it'll be used more and not less. Um now let's try to understand why does it cost $100 to generate an hour of video.

Uh and I try to be as realistic as possible. So first of all um the systems that for example DL they talked today about having set up now in Sweden uh some NVL72 racks by Nvidia the all-in cost of such a system is roughly 5 million and you need one of these systems to maybe run one two three four state-of-the-art video streams that huge system consuming roughly 100 kilowatts cannot do more than that um and so the electricity bill optimistically speaking will be 100k in the US and Europe with higher energy costs might even be more than that. So over a 5year period, right? Maybe let's say okay over 5 years this hardware will be depreciated.

We're looking at one and a 1.1 million per year and a year has 8,760 hours. And so if you divide the 1.1 million over 8,6 hours and you assume basically 100% utilization, the system is running all the time. Then you're looking at $125 per hour. So as I mentioned, let's say it's computing actually two, three, four streams in parallel. You assume very high utilization, but of course people setting that up like deep and other companies, they also want a margin to their product. And so you're looking at maybe $100 per hour. So that's how this this cost is actually uh uh generated. And so you have very high system cost, right? the 5 million and fairly inefficient processing um because the the electricity bill actually also makes up a fairly large portion of it.

Um and so the three things that are really important is addressing accuracy with new hardware. Um so you need to enable people to run the biggest models at the highest numerical precision. Meaning can I run it actually in 16 bit instead of just eight or 4bit. Um, this is something that Nvidia has been talking a lot about lately, but the truth is that uh if you have a model that especially generates images or video or audio or even just really large models working on very complex tasks, most of the time you don't want to reduce the uh budget that you give that model in terms of its size, in terms of its numerical accuracy. Um, and the next is speed.

Of course, speed is kind of this um enabler. Once something becomes fast enough uh it becomes suddenly usable. So very often the value of an AI model in terms of its speed of execution is not just a curve like this that it's more and more applicable the faster it gets. No, it's more like a hockey curve where suddenly it's fast enough so that in a certain application you can now really start using it. Uh that's of course especially true for real-time applications because before is real time you can't use it suddenly it's real time now you can totally use it. Um so there are these kind of step functions um and that's what we're also trying to address and these really are trillion dollar level um markets.

Uh and then finally the cost it really needs to come down by 10x 100x sounds like a very big reduction that um that yeah we have to somehow make happen otherwise a lot of these applications will not become a reality and the the way that um yeah AI chip companies are trying to solve this is u for example new ways to do AI math. So, um, what is AI? It's a giant mathematical formula with hundreds of billions of, uh, variables. That's really what an AI model is. It's just a giant formula. Uh, so it's all math. And, uh, if, um, if you're doing 50% multiplications and 50% additions, which is what these models normally do, um, the multiplications are very expensive.

Um if you remember how you learned to do a multiplication school, a multiplication has a quadratic amount of steps to do it on paper for the number of digits. Uh an addition has a linear amount of steps. So in hardware that's exactly the same. Multiplications are way more expensive than additions. So um if you can find a way to to make that cheaper without losing accuracy, that is a huge lever in cost. Easily 4x um sometimes more, sometimes less. Um, the next really important thing is I already alluded to it. Uh, one chip can maybe compute a small model. Um, but normally these days you need lots and lots of chips in a very fast network.

And we're not talking about your standard Ethernet network with maybe 100 Gbit or so. We're talking about something that is many orders of magnitude faster. Uh, hundreds of gigabytes, many chips connected to any other. And uh, so these are the most state-of-the-art network technologies that you can imagine. and uh they're not easy to to build and so on. And uh we're actually collaborating with a company um to to tackle the network side of things, Juniper Networks. Um one actually the inventor of fabric networking, chip to chip networking. Um yeah, and finally um the Nvidia systems like the NVL72 for example, Nvidia's best uh most expensive thing right now. It's built for training, it's built for inference, it's built for many different applications.

So we need purpose-built systems because then you can throw out a lot of junk that you simply don't need for just a task that you care about. Let's say it's inference of genai models, large models and suddenly things become again so much it's like having a very heavy car and throwing out the massage seats and you know because you simply don't need them because you're maybe on a racetrack. Um and so the way we try to do that is first logarithmic math. Um so logarithmic math means uh when you you might remember this uh this little formula from school uh if you have a * b which is the expensive nasty thing in AI models that we want to make cheaper.

Um the logarithm of a * b equals to the logarithm of a plus the logarithm of b. Very simple rule we didn't invent that of course um napier uh thought about that many probably hundreds of years ago. And uh but uh you can work this into an AI chip and then the AI chip actually does multiplications that way. Um the additions then become another thing and we spent seven years doing research on AI math to solve that part as well. Uh it's working really well now and uh yeah that that makes right off the bat on the most atomic level that makes AI 4x cheaper. Um so right pretty big uh number 4x it's not like 10%.

Um but it took a long time to to get there uh because it it really goes down to the transistor level and how you want to implement things very efficiently. Um yeah and then as I mentioned we're cooperating uh with a company to solve the chip-to-chip networking. Um again you can really imagine if you have a 100 chips then computing this giant reasoning model becomes almost 100 times shorter than with a single chip. Uh a typical number these days is a parallelization over 48 maybe 16 chips and we're trying to get this to 72 and 144. So definitely go beyond the 100. Uh so this is again another yeah easily 10x if not 20x that uh we want to bring to the market.

Uh and then yeah system level design. So uh in our company that's always the mantra you know think about everything you do in terms of what this means for the system. So for the entire pod that we put in a data center rack, maybe even beyond that for the ecosystem that it becomes a part of. Um, and this comes also down then to making certain different design decisions. Um, yeah, that's it from my side. I just want to advertise something for a second. Uh, actually tomorrow, uh, this is a Pacific time, so you'll have to calculate in your head at what time that is in Germany, but we'll have a tokconomics webinar. Um, and tokconomics is token economics.

So we'll be talking in much more detail about the cost of AI and uh yeah talk compare different hardware solutions and why they mean different costs and uh yeah if you want to join that too then go ahead. Yeah, thank you. That must be 11 p.m. I think. No, hang on. No, it's it's in the afternoon. It's in the afternoon. It's the other way around. The other nine hours. Thank you. There's one. Thank you. There's a question in the audience in the back. Um Fabian in the very far back the other side of it. Uh but while we are waiting for that one question um you talk about affordability and accuracy and me uh from coming from image processing I I really see the word immersion there and computation that gives things pictures or images and I think immediately of the of this movie called her.

Oh, I is it with the robot like the No, that's with one Hin Felix who is falling in love with his AI. That's like 10 years ago. Um and we're like um I look into the ethics of the things. Yeah. Oh, I mean we're looking at a massive cultural shift I would say in all regards. It's pretty crazy how little it's being discussed in politics, I think. But yeah. Oh, which leads me to the question, what do you what do you think which industry would transform the most in the next couple of years? in the next two three four years um yeah I think therapy companionship I think is going to happen in a way that uh like it's going to be bit stealthy bit but hidden right a lot of people are just going to do it in their private lives it's already happening uh it is according to some study already the biggest application of genai yeah I think that is likely um because it's also like in in therapy or coaching or

so or in companionship it's also fairly robust against errors Right? Like if someone tells you, oh, you know, things are going to get better with XYZ, there's many different ways to express that, right? Accuracy is not super super relevant there. So, yeah, I think that is going to benefit a lot from that. You might need a lot of empathy and uh training from um experts who are in the realm of psychology and coaching. Absolutely. Hopefully, knows what they are doing. Um this question now goes to um or the stage is now yours. I thank you for the talk. So I think it's quite exceptional the the biography that you have. So my question would be how do you see the state of silicon development in Germany because um yeah it would very interesting to see that.

Yeah I mean I know a couple of startups in Germany that are doing for example analog AI chips um so very state-of-the-art technology that are still going to need maybe a few years to mature um to a product level. Um, I mean, yeah, let's be honest. Obviously, uh, the the great silicon days of Germany are a little bit over. Um, uh, but not too many months ago, I was, for example, at the AI, uh, no, sorry, at the chip day, uh, in Bavaria. Um, and there's definitely stuff going on. Uh, it's of course an extremely capital intensive uh, thing, right? I mean, our first chip cost 25 million to develop and that was dirt cheap.

Normally, it would easily be 10x of that, right? So, it's very capital intensive. Um, and of course, yeah, it's still all of that traction gets pulled into Silicon Valley. So, we're actually 50% in Silicon Valley and 50% in Munich. Do you have maybe a call to action for the young minds listening? um generally or if they want to get into silicon stuff or um so I think both probably I mean sometimes uh I have friends who have young kids and they're thinking about you know how I'm going to raise these kids in the AI age. I think um being an expert in one thing is not going to be a good thing anymore. Try to be extremely good at using tools AI tools and to try to be a tool user.

I think that's the one advice I could give maybe. Maybe can I add and be a critical thinker because earlier I had these numbers from the earlier speaker and my numbers came from perplexity and churchp and that was definitely not accurate. Yeah. Yeah. Just fall in the pit there. Hi uh thanks a lot. I also enjoyed very much your presentation and discussion. Um I'm Philip Castro Suero. I'm president of Exponential University. We focus a great deal of our attention into AI as well. Um I do follow the um the the application as a high priority in seniors as you mentioned in companionship. I was a bit surprised that you didn't mention education. Yeah. Um I would like to collect some thoughts.

I mean so the the one crazy market example was education, right? And I personally also think that's going to be really great because you can bring education with to everyone who has a little netbook or a little iPad or something in some rural area with a very specific dialect there, right? You're going to have that translated perfectly to their dialect and so on. No, I think it's it's great. I mean there I think a lot of parents will be concerned like, you know, am I really going to let my child be uh educated by an AI? But yeah, I think give it 10 years and then I think it'll be a pretty normal thing. Um yeah.

Um and and one one other question is uh where do you think is the area of the biggest biggest improvement your company will have in the next three to five years that the an improvement that we are going to cause or that you are going to now that you are that you need to have. Yeah. Yeah. Eventually to generate a bigger impact. Yeah. So yeah. So the uh surprising and maybe also not so surprising answer to that is even though we're a hardware company, software is actually kind of the most important part because things need to be usable uh and this is why Nvidia has this big market share because their stuff is usable and others isn't.

AMD's chips for example are often better than Nvidia's chips uh from a spec point of view and also cheaper but less usable and that's why it's not happening and so we focus a lot of our time on software. We have way more software developers actually than hardware developers. Yeah. Okay, thank you so much and that's your applause. Fantastic. So I mentioned a bit I mean there was immersion and um basically that was also a bit about storytelling because that's how a lot of people understood. Oh, hang on there.

Rise of AI 2027 · Berlin · May 2027

Be in the room for the next one.

Ten years of talks started as conversations in one curated room in Berlin. The next edition brings 300 selected leaders together in May 2027.