Did we go from a broad bubble to a gen AI bubble? What is the current state of AI and generative AI? What has been commoditized and what is still distinctive? What does the future hold?

Navigation:

Our co-hosts:

Our show: Tech DECIPHERED brings you the Entrepreneur and Investor views on Big Tech, VC and Start-up news, opinion pieces and research. We decipher their meaning, and add inside knowledge and context. Being nerds, we also discuss the latest gadgets and pop culture news

Bertrand Schmitt

Welcome to Tech Deciphered, episode 57. This episode will be about generative AI, and we will be asking the question, “Are we in a generative AI bubble?”

Bertrand Schmitt

In our last episode, we talked about AI, what was the latest happening in terms of endpoints, PCs, Macs, iPhones, and is it a risk? Is it a benefit? Today, we’ll talk about what’s happening with GenAI. Is it overhyped? Is it too much investment? On the plus side, we’ll be wondering, okay, maybe it’s not a bubble after all, or even if it’s one, is it such an issue because basically is it laying the foundation for a new stage, a new scale of AI for an act two of generative AI. Good to talk to you today, Nuno.

Nuno Goncalves Pedro

Nice to talk to you as well, Bertrand. I’ll start with my answer. It is a bubble and it isn’t. I’ll come back to that. I’ll leave you guys on that cliffhanger. Let’s start maybe a little bit with where we are in the state of AI and GenAI. Are we at commodity level or not? Adoption levels that we’re seeing in the market. Interesting report from McKinsey. Some around the state of AI, they did over 1,300 interviews. It was really a survey format around AI and GenAI adoption.

Nuno Goncalves Pedro

Some, like the conclusions, increased adoption of AI in at least one business function. Over the last year, this dramatic increase. I’m not really sure. Everyone that said, yes, we’re using it is actually using it. I take that with a grain of salt. This is self-reported, again, and so everyone has to be using GenAI, but everyone’s aware of it. I’m sure there’s an increase in adoption for sure.

Nuno Goncalves Pedro

The second piece that I feel is a little bit more exciting is what are the functional areas of companies that are using AI and generative AI more actively? Maybe not super mega surprising marketing and sales, which is core to some companies, but relatively support function. A little bit surprising to me that people are seeing product in our service development as number two. Surprising to me that software engineering is so low. Again, maybe no software engineer has actually filled in anything around that. That’s why it’s so low or middle of the board, or they don’t know what their engineers are doing, really, which is also interesting.

Nuno Goncalves Pedro

Then very low on strategy and corporate finance were business analysis, triangulation of data, or using ChatGPT, et cetera, I would have taken a little bit for granted that people would be using it. A bit surprising on that. Just feel it’s an interesting… Again, self-reported, it’s a survey. Some interesting conclusions on both sides in terms of the functions, et cetera. Some the conclusions as well on the rapid ascendancy of generative AI.

Bertrand Schmitt

Yeah, for me, what I’m quite impressed, I must say, is how fast generative AI has picked up. I don’t think I remember any new technologies that move so fast in terms of adoption, because here we are already talking about adoption metrics. I mean, it moved from nowhere in 2022, 33% in 2023, 65% in 2024. Basically, as high as the adoption of AI, the general AI adoption took seven years to get there.

Bertrand Schmitt

Here it’s two years. I’m not sure I remember any technologies that move so fast in terms of adoption. It took years and years for smartphone to get there. It took 5, 10 years for cloud computing to get there. It’s really fast, I must say for me, is my first reaction.

Bertrand Schmitt

In terms of function, marketing, I’m not surprised. That’s where you can use generative AI to generate content relatively easily. It’s not mission-critical content, it’s content you can review. At the same time, it can give you some clear benefits in terms of moving pretty fast and needing less resources. I guess for some other function like software engineering, I think right now, I guess the biggest issue is that you cannot trust it. It’s not something that you are ready to deploy in production so quickly, very different from content for marketing.

Nuno Goncalves Pedro

Maybe perhaps a little bit more interesting is the article from Tomasz Tunguz from Redpoint. I hope he’s still at Redpoint, and I didn’t say anything wrong.

Bertrand Schmitt

I’m not sure, actually.

Nuno Goncalves Pedro

Around the evolution. He was at Redpoint.

Bertrand Schmitt

He was.

Nuno Goncalves Pedro

Around the evolution of AI and this notion that it might be following the sigmoid logic, so log scale sometimes. Then sometimes exponential in terms of curves, so the exponential improvement and then log curve, which is flattening out for those who are not familiar with log curves. I would even say maybe it then looks a little bit linear, because if it’s exponential and then log, then it’s basically it looks a little bit linear. A little bit of argumentation around, are we going for the next S-curve? Are we going for the next log curve? Where are we on development? I think our intuition is probably telling us that there’s more stuff to come still on this technology stack.

Nuno Goncalves Pedro

Although this technology stack, as we know, the whole use of transformers to get us to where we got to buy OpenAI and all the other players is definitely, according to everyone that I talk to, not what’s going to take us to AGI. It’s definitely not what is going to take us to the big innovations, the big next themes and the next levels.

Bertrand Schmitt

I really like the way OpenAI has categorized the different level of AI similar to self-driving cars. As a reminder, level one was a chatbot layer, robots that can chat with you. You could argue, interestingly enough, this level one is what used to be actually considered something very hard and the proof we have made AGI, I guess everyone now realize that Turing test is absolutely not enough.

Bertrand Schmitt

Level two: reasoners, robots that use logic to solve problems. Level three: agents, robots that act independently of humans. Level four: innovators, robots that create new ideas independently. Level five: organizations, robots that replicate the work of AI.

Bertrand Schmitt

At this stage, where are we? OpenAI say it’s close to move from level one to level two. We will see. As you say, I tend to agree that at this stage, I feel that we are getting some plateau. Just to be clear, things keep improving. Don’t get me wrong, there are very clear metrics that we improve stuff, but it feels more an improvement through brute force. Let’s move from that many terabytes of data to 10X that terabytes of data. As a source, let’s move from 1,000 GPUs to 10,000 GPUs.

Bertrand Schmitt

Yes, of course, you are improving stuff and the magic things with the current foundational models is that they scale very well with more data as long as quality stays good, and they scale very well with more GPUs. I guess the only issue at this stage is that it doesn’t feel that there will be quantum leaps with this technology in terms of moving to level two, level three, without any separate breakthrough.

Nuno Goncalves Pedro

Yes. We know OpenAI is using other methods and other technology pieces into their technology stack beyond transformers, et cetera. It’s already very apparent that they are. But still, the analogy of this with the self-driving car levels, which if you remember, is also level one through level five. Magically enough, it’s very similar.

Nuno Goncalves Pedro

For those who are unaware, probably today, the best case scenarios that we see out there is like a level three. Some cars don’t even have beyond level two. It’s like they’re telling you, there’s a car on your lane or whatever, that’s nothing. A little bit like the Tesla experience today is on the level three thing.

Nuno Goncalves Pedro

We’re very far from level five. I’m sure, Bertrand, you’ve had similar stories. We’ve been promised self-driving cars for a long. I remember from my perspective, the penny dropped when I looked very heavily into the space, spent over a year looking at companies that were deeply embedded in the architecture of self-driving. Then at some point in time, I ended up spending time in this specific space with just simulation. The area where you simulate behaviour of cars to get those driving miles, but there are actually no cars on roads. It’s actually a simulation.

Nuno Goncalves Pedro

The first thing is, 80% of the players I said was just BS. It looked like gaming simulations using Unreal, and there was no factoids to it. There was no physics to it. There was nothing that they were really testing. Then you got to the other 20%, and you realized these guys are a little bit more legit. Then you realize, well, there’s things that they’re not simulating, for example, like rain and snow, and what happens if there’s ice on the road? At some point you’re like, “But that’s pretty critical, right?’

Nuno Goncalves Pedro

That’s when you realize I spent probably a lot of time around this, 2016, 2017, when you were like, “Oh, my God, this is going to happen.” You’re like, “No, not anytime soon.” Again, this promise that we’re going to achieve AGI with a stack right now, I feel is totally overblown.

Bertrand Schmitt

Yeah, I must admit that one of the rare times where I felt duped by a Silicon Valley bubble was about self-driving around 2015. At some point, I genuinely believe we are one or two years away from self-driving cars. Pretty quickly, I realized I was being lied to. I don’t know if it was really a lie or wishful thinking. For sure, it was a very strong reality distortion field in place. We are in 2024. Today, what we have is level three.

Nuno Goncalves Pedro

At best, it’s not pervasive, right?

Bertrand Schmitt

The latest version of self-driving for Tesla is not too bad, I must say. As long as one is there, I’m fine. I would say that we have one that managed to find out the way to do it. We are still not at level five for driving automation. As a reminder, the Tesla robotaxi has been delayed. We’ll see how many times it gets delayed. But to the point, 10 years. In 10 years, we are still not there.

Bertrand Schmitt

Here with generative AI, let’s not forget that these are the people who told us maybe a year ago that we were very close to AGI and you have to be very careful because these models can become very dangerous pretty quickly. I think it’s totally the same BS. Sorry, but this is not going to happen in one or two years. If it’s 10 years, I guess we’re lucky.

Bertrand Schmitt

I’m certainly not saying it will not happen. What I’m just saying is that there are ways to scale current models, bring in more data, bring in more GPUs, but we are still not at the stage where we can really provide some leaps that bring us to significantly higher level of autonomy and going to a full level five AGI, if you want to call it that way. I think that it’s a similar parallel. It’s a good parallel in a way, self-driving versus the current craze of GenAI.

Nuno Goncalves Pedro

We now have a name for this thing. I remember In previous episodes, I referred to this as the strategist dilemma. There’s a name called Amara’s Law, so a guy called Roy Amara, a futurist. Basically what he said is we tend to overestimate the effect of a technology in the short run and underestimate the effect in the long run. I’ve mentioned to this a bunch of times as a strategist dilemma. Apparently, there is a name for it. It’s Amara’s Law, and it’s absolutely spot on.

Nuno Goncalves Pedro

We tend to overestimate the effect of technology in the short run. We underestimate the effects in the long run. We don’t really know what’s going to happen in the long run. But right now, what we’re seeing is we’re grossly overestimating the impact of generative AI in this short run.

Bertrand Schmitt

I want to highlight that we just talk about AGI, how far we are, self-driving, perfect example of how bullshit it can be early on. Another recent example for me was Klarna. They put a press release not too long ago claiming that now they can handle two-thirds of customer service chats in the first month releasing their AI. I was like, “Seriously?” I have not heard any company being able to achieve that. They are the only one. I realized now that people keep talking about them, “Oh, look at how much they did.”

Bertrand Schmitt

Some people claim if the only reason they got great results were how bad was their automation previously. That any bank, any insurance, any other company that is a bit more advanced on any of this stuff, just by using the regular tools of the trade from the 2010s even, you will, of course, have improved a lot your autonomy in supporting customers, in providing some level of autonomy, self-service.

Bertrand Schmitt

Basically, these guys only got great numbers because their base were so bad. That for me is very interesting because most people just copy and paste to press release, it’s great, and don’t try to look deeply into, “Okay, how does it work? Where were they, really?” Of course, if you come close to zero in terms of how you manage stuff, you may make great improvement. But most industries, most companies are not in that spot.

Nuno Goncalves Pedro

Maybe just close the cycle on the BS around it, right? I think it’s a natural segue to what we’re talking about right now. It just takes a really long time. At the end of the day, product market takes a long time. Things to get nailed down takes a long time. There’s been a lot of risks taken in certain things, but honestly, this is the beginning. There’s going to be a lot of these BS stories. “I’m using AI, I’m using whatever,” and you’re like, “Are you really using any AI algorithms?”

Nuno Goncalves Pedro

Then in retrospect, also the other way around, there have been things are using what we now call AI for several years that we never called AI because they were just there. Someone was running algorithms on their data sets, and they weren’t calling it AI. But random force regression, for example, which is part of statistics, actually qualifies as deep learning. There’s all this nomenclature that will make all this bullshit emerge of BS stories around the table.

Nuno Goncalves Pedro

Moving to the structural issue of, are we on the commodity side right now or not of foundational models? Do we think foundational models are a commodity or not? Bertrand?

Bertrand Schmitt

I think that’s a big question because a lot of the craze started with foundation models. The first one to really be visible was ChatGPT. Probably ChatGPT V3 was really a breakthrough for sure. At this stage, my belief is that, yes, they have become a commodity. We have a few of them from OpenAI, to Claude, to Mistral to Llama, that are able to perform at similar level. If we take models of similar size, some of them being actually open weight and free, like Llama, like Mistral, some being closed source.

Bertrand Schmitt

It feels at this stage that if there is some regular improvements, like multimodality, for instance, using tools, controlling tools, in a matter of six months, the competition has reach similar levels of features and performance. It feels like the recipe now is well known. You build used data centres, you operate used data centres, you put in place thousands of Nvidia GPUs, you have a few hundreds, if not thousands of AI engineers, and you bring at scale a lot of data that you scraped from the internet. Sometimes it was authorized, sometimes it was less authorized. Basically, you run that for a few months. If you apply what is relatively state of the art, you end up at similar places.

Bertrand Schmitt

It’s not an indictment. I think it’s great what they are all doing. It’s not saying it’s easy. It’s not saying it’s cheap. It’s not cheap. But it’s saying that it looks like no one seems to have been able to really separate itself from the pack by more than a few months. As a result for me, it’s raising the question, is it a commodity? Probably an answer being yes.

Nuno Goncalves Pedro

I tend to agree with you. There’s anecdotal evidence. I was proposing an episode to Bertrand. I’m eventually going to convince him to do this episode, which is we, for a couple of weeks, run everything in parallel. With Claude, with ChatGPT, with Gemini, we see the answers across a variety of topics, and we start figuring out what are the ones we like the most, who’s performing the best in different topics, et cetera.

Nuno Goncalves Pedro

I started the inklings of that, running a couple of questions across the three, and honestly, it’s shocking how many of the answers are very, very, very, very similar. Obviously, same prompt, but very, very, very similar. These are very open questions.

Nuno Goncalves Pedro

I asked one question, which is, “What should we do in Tech Deciphered in our next episode as ideas?” They all came up with really great ideas. Magically enough, they were all very similar. It’s grand that they knew what Tech Deciphered was, so that was cool. It’s surprising to me in that sense. That’s the anecdotal piece of the puzzle. I don’t know. I just still think from a data set perspective, there’s still private data sets out there, et cetera. I’m not sure that extrapolation of it’s all commoditised, it’s truly there. There will be vertical AI applications with private data sets that people won’t have access to.

Nuno Goncalves Pedro

Today, we have copyright infringements on public data sets or relatively public data sets once people get access to them, which they shouldn’t. But certainly on private data sets, I don’t know. I’m mostly there with you, Bertrand, that maybe we are closer to commodity than we think we should be. But I can’t believe that all the data sets in the world have been encapsulated by these guys, so we’ll see.

Bertrand Schmitt

Yeah, we will see. Also, another piece of the puzzle is that we have seen a new wave of computing device that are natively AI-enabled, so new APUs, new CPUs with NPUs, neural processing units for the latest Windows PCs, for the latest Macs, the latest iPhone, iPads. They are all AI-enabled in hardware, at least the most recent devices. The software is also becoming there with the latest version of Windows, of macOS, iPadOS. But it’s not really obvious in terms of end user features that there are breakthroughs, to be frank.

Bertrand Schmitt

Of course, we have some improvements in how the camera is working, how noise-canceling is working, that stuff. But in a way, it’s more of a natural progression. Of course, there are some new tools to write documents, to generate some graphics. But it’s not clear-cut that we have seen some killer apps at this stage that justify this level of investment and new hardware deployment.

Nuno Goncalves Pedro

Yeah, and it might be that it’s just on a different curve, right? In some ways that this generative AI thing, which for us seems like it happened overnight, is now on the mass adoption curve, so it went very quickly through all these different hoops of getting with early adopters, then early majority, and now the mass majority.

Nuno Goncalves Pedro

In some ways, a lot of these tools, both on the software and on the hardware side, are going through similar cycles, but maybe they’re going at a slower pace. People are not then willing to take risks on those products that are out there. I, as a developer, am then not willing to use them because there is an intrinsic risk. It’s not an early majority play yet, and so I’m like, “I’m just going to wait and see what’s going to take off.” Because then you start having noise, right? Maybe that was the advantage that ChatGPT had. It didn’t have noise early on. It took a while for others to catch up. Maybe they just not jumped, but they went very quickly through the initial stages of adoption, whereas now it will be increasingly difficult.

Nuno Goncalves Pedro

Maybe for the huge elephant in the room, is there overinvestment in AI startups in public markets? Obviously, Coatue, for those who don’t know Coatue, they’re a significant hedge fund. They also have a growth equity practice that’s well known, and they, I believe, still have a Coatue Ventures, which is more focused on the venture capital side of the things. Philippe Laffont, the founder, and I don’t remember his title, but he has a funky title. He’s Portfolio Manager. I think that’s what he says in the interview. I’m a Portfolio Manager. I want to be a Portfolio Manager when I grow up then as well, like Philippe.

Nuno Goncalves Pedro

For you guys who haven’t watched it, there was an interview with him and David Rubinstein, who’s obviously super well known, mammoth of the investment industry as well. It’s worthwhile watching. It’s a really interesting interview. There’s a lot of little titbits that come through on how someone behaves, how someone takes it and scales it to the next level, et cetera. But his point is that we’re close to this notion of we’ll now have several companies at 2-3 trillion market cap when a trillion was the big number back in the day.

Nuno Goncalves Pedro

Then there’s all these market caps out there, like S&P 500 is at 50 trillion, and global market cap at the 110 trillion. Is this a, “Something’s got to give”? He talks a lot about this notion of the winner tends to win, the big tend to get bigger.

Nuno Goncalves Pedro

We used to have this rule of three, which is in a specific industry, depending on the industry, it might be on a specific geo as well, or geography as well. The number one wins clearly, the number two does super, super well, the number three barely lives, and then all the others over time get killed. Obviously, there has to be some market definition into this. What is a market? Is it just the geography? Is it very specific sub-sector, et cetera? But I feel that still applies.

Nuno Goncalves Pedro

The question is, “Is now everyone rising to the top because it’s unclear who the winner is?” But once the winner wins, then we have this notion, again, of number one, number two, number three, and therefore a lot of these valuations will need to get rebalanced? That’s one thing. If there’s that, then there may be some significant overinvestment. A lot of these players are getting kicked in the ass, pardon me for my French, kicked in the ass on their valuations immediately with their latest earning reporting.

Bertrand Schmitt

Yes. One point he was making is that AI was clearly the dominant driver for stock market return in H1 2024. That’s a very important useful metric to keep in mind. At the same time, it’s stock market return, so this can change. At the same time, what was true is that some of these big companies were increasing their earnings significantly as well. Of course, Nvidia is a clear leader, but in that quote-unquote AI space, a lot of other companies were included there, like Microsoft, like Google. It’s clear it has had a huge impact in H1, and not just H1, but also 2023. The question is, can it keep running that way? I think on the street, it feels less like this, to be frank. There are big questions. Can it really sustain itself?

Bertrand Schmitt

Another question he was raising was, is it too hot right now in terms of VC in AI funding? It’s clear that there are some significant questions you can ask yourself. Some metrics are pretty impressive, because if you take AI startups right now, it’s 3% of total deals, but it’s 15% of total capital. It’s 5x the valuation of other deals, and it’s 6x the round size. I mean, this is difficult to sustain, especially if you don’t generate the metrics that go with this level of expectations.

Nuno Goncalves Pedro

I think there’s definitely a bubble in AI VC funding. I mean, it’s silly. I mean, the numbers you quote are silly. Just to be clear on the valuation, the 5x, it’s for Series B and Series C equivalents. There is a capital intensity to it. But then there’s areas that are actually underfunded. I mean, semiconductor has received very little funding, because everyone believes semiconductor is very difficult to scale.

Nuno Goncalves Pedro

But there’s pieces of the architecture of delivery of AI that are just totally broken. Switching, et cetera. It’s like, shouldn’t semis also be funded and stuff? We just funded a company in the space, not on the switching side, but on another side. There’s a little bit of a parochial speaking to myself here in some ways, but definitely there’s underfunding. The rest of the VC market funding is stabilising, and then AI is nuts.

Nuno Goncalves Pedro

Part of it, I believe, is what I call intrinsic noise. People just want to get into the latest big thing and put a ton of money in the area. Shockingly enough, that’s been the best funded is the models area, according to the COA2 analysis, at 14 billion. That’s silly. If we believe that models are actually getting more and more commoditised, then it’s like, “Really?” Is this an arms race and someone’s going to buy these guys out, and are they going to buy them at the premium?

Nuno Goncalves Pedro

Again, it just feels like a tremendous bubble where in a very classic lemming mentality, because public equities have gone through the roof. We VCs are now jumping into this boat and just saying, “I’m going to put money in this, then money in that, money in that, and whatever.” Maybe. I think it is, honestly.

Bertrand Schmitt

I think the models space is not really a VC game, it’s a big tech game, because you can afford the scale, the size, the CapEx, the continuous financing that you need, and you just generate cash. That’s a way to use that cash. Actually, like Mark Zuckerberg was saying, if you are a big tech, you cannot afford not to do it, because you don’t want to depend on your competitions, competitions’ model and data centre and the like. You have to do it anyway.

Bertrand Schmitt

The space left for startups, and I would say that OpenAI might be a special case. They started so early. They have been there for so long. They have made so many breakthroughs. So they might be the exception to the rule. But it will be a tough one to keep investing. I mean, we might see a Mistral, because they are EU-based, they get EU financing, and that is a special category. Maybe xAI can have a different approach, more focus on freedom of speech, and that generates some interest from that perspective. But I think it will be a tough one, because you need to be able to convince that you are going to spend billions while you are facing the biggest competitors it can be, and performance just seem to converge.

Nuno Goncalves Pedro

Yeah. Again, to our question, why the hell are all these models companies getting funded? No answer yet. We’ll soon find out.

Bertrand Schmitt

I would not bet on this. Going to that Sequoia article a few weeks ago, that $600 billion question. I like big questions. It was a nice blog post from David Kahn, and basically was simply running some metrics. Nvidia, there is a run rate for the data centre revenues, that is around $90 billion for Q1 2024, estimated to be $150 billion by end of year for Q4. When you consider that it’s only one half of the cost of running a data centre, that means that the implied data centre spend is $300 billion for AI data centre, for AI spend.

Bertrand Schmitt

Then you have a software margin of 50%. That means that if you build software on top of this, then it’s $600 billion that are required for payback. From $150 billion of run rate revenues from Nvidia, that’s real, we need to generate $600 billion of AI revenue for that payback, $600 billion revenues. The big issue is that we see billions today from OpenAI, from Microsoft, from a few others, but we don’t see hundreds. I’m not even sure we have reached tens of billions.

Bertrand Schmitt

That’s a big discrepancy that everyone is talking about. It’s not new. It’s been a year that some have raised the question. To be frank, it’s normal. You start to invest first, and you have to do this CapEx investment, and step by step, you generate revenues. I think here, maybe the question is that we don’t really know where it will come, when it will come, and if the guy who has spent all that money are going to reap the revenues.

Nuno Goncalves Pedro

Yeah, I think the points that are made, this is a revisiting. If you guys didn’t read the early articles from Sequoia, it’s a good read. I think it’s based on the Act 2 doc written by Sunyu Wang back in the day. I think he published something on AI’s $200 billion question, and now it’s $600 billion question. I agree with most of what he’s saying. I think there is a little bit of underpinning. When he talks about the infrastructural piece, maybe there’s a little bit conservativeness. There is this, for example, the notion that all data centres are going to be commoditised, that GPU computing, obviously, is turning into a commodity.

Nuno Goncalves Pedro

A lot of this is correct, but as things develop, there’s frames and things that come out of this that are not totally true. There’s pieces of this puzzle architecturally that don’t tend to commoditise immediately. I don’t know. I have some doubts that the notion of it is, “Oh, my God, this is all going to implode,” whatever. It’s, again, back to the Moore’s law, as I was saying, I think there’s pieces of what we’re building today that will be valuable in the future.

Nuno Goncalves Pedro

Now, obviously, there’s pieces of tech that will become obsolete, because as we know, tech becomes obsolete, and there’ll be new pieces of tech that will replace those pieces of tech. Chipsets, et cetera. But still, there’s a little bit of, I believe, the take it on the low side and, “Oh, this is not going to scale.” I’m not sure. There’s definitely still a lot of runway, and saying this is a tens of billions of dollars play. I can’t believe this is a tens of billions of dollars play. It has to be a hundreds of billions play. Now, the question is, are these the players that are going to dominate those spaces or not?

Bertrand Schmitt

I think that’s a big question. I agree with you that I also believe it would be hundreds of billions. The question is, is it in two years from now? Is it in 10 years from now? Do you need to spend even more to get there?

Nuno Goncalves Pedro

Maybe, yeah.

Bertrand Schmitt

Who will get the money ultimately? Because if you spend a lot of this money, but you are not yourself generating revenue, that will be definitely some issue.

Nuno Goncalves Pedro

There’s clearly a subsidisation right now of certain players in the market. Nvidia is definitely getting subsidised. It’s getting subsidised by the big players. It’s getting subsidised by VCs giving money to startups to give money to Nvidia. Definitely, there’s a huge amount of subsidisation going on. Now, well done by Nvidia, and it’s amazing what they’ve built, and we all count on them to be helpful to us and all that stuff.

Nuno Goncalves Pedro

But at some point, the money needs to ripple back to the other layers. If it doesn’t ripple back, this is not sustainable, therefore it implodes. The valuations don’t make sense, and all of this is write-off investment. I think that’s the question that in some ways David is asking. It’s like, “Are we sure, or are we going to have to write off 90% of all investments that we’re doing around some of this stuff?”

Bertrand Schmitt

It’s good to see a VC asking this question, because in some ways, you could argue they are part of the problem. Sequoia is one of the lead investors in a lot of AI runs. They put a lot of money to work there. I think it’s a good point in some way. I don’t know if it’s really raising the alarm, but definitely raising questions. Goldman Sachs also had a report recently raising similar questions. Are we spending too much money for too little benefit? I would say here that some that are writing that don’t seem to believe at all where AI is going. I think that’s a different issue and that does not make me afraid.

Nuno Goncalves Pedro

I read it thoroughly. It was more in an interview format, right? Talking to different specialists, some internal analysts and some external experts. There’s a little bit of naysaying there that I feel is more religious than truthful. It’s more like, “Oh, I’m not part of that religion.” I’m like, “Cool.” But it’s like, “This is happening like this.” If we’re asking someone who’s from a different religion, “Do you agree with this?” Well, of course, I don’t think it’s going to happen.

Nuno Goncalves Pedro

Some of the points I think that were made really around quantitative analysis and what’s happening are interesting to me. For example, around the energy lack that we have, the build-up of energy we need to have to develop these things and to get to that level. I think that analysis was quite interesting and strong. There’s a couple of things that are saying, “Look, even if you want to go really fast, guys, there’s a couple of other issues you need to solve first.” Like, “Oh, cool. Data centres, how do they get energy?” Right? “Oh, energy. Electricity.” I think there’s just this ripple domino of you touch one, and all of a sudden, you realise the problem’s down at the level that you thought was solved, but it’s not solved at all, right?

Bertrand Schmitt

Yes.

Nuno Goncalves Pedro

I think that part was interesting. That part of the report was quite interesting from Goldman Sachs.

Bertrand Schmitt

Yeah, I agree with you. I felt that some interviewees, especially one MIT professor, were disappointing in the quality of their analysis, their understanding of AI. But I agree the energy piece was probably the most interesting in this report. Energy is interesting, because it’s not as if you can launch new power station that easily.

Bertrand Schmitt

Interestingly enough, one point that resonated with me was they were explaining quite clearly why we didn’t have an energy issue with cloud computing. Basically, it is because cloud computing was a replacement for private data centres. Cloud computing actually is inherently more efficient than these disparate corporate private data centres. In a way, the transition to cloud computing was a net benefit in terms of energy consumption. Even if cloud computing itself increase in volume, the gain in efficiency was a huge compensation for this. Therefore, we didn’t get a real energy issue.

Bertrand Schmitt

However, we’re at this stage where we are back to where we used be with private data centres, because now cloud computing has scaled so much that it has not simply replaced, but it’s beyond in energy consumption. We’re adding a new layer with AI that is clearly new and was not there before, and that is consuming enormous resources in some cases.

Bertrand Schmitt

If we start to plant 10 years ahead in terms of what could be the need, if we see that ongoing growth in AI build-up, then we actually have issues. We have issues, because unfortunately, in the West, our capacity to build new energy power source is not good. We don’t build that much. We don’t build that reliable. The most reliable of all, most efficient of all, nuclear, is not being built at scale these days, unfortunately. We really have an issue in terms of how we scale our computing for AI if basically we have a wall in front of us in terms of energy consumption.

Bertrand Schmitt

Reading a recent paper from Meta, actually, it was interesting to see that it’s not just the energy consumption itself, it’s a grid. Basically, when they are starting and stopping some training exercise they do, when you are managing thousands of Nvidia GPUs at scale, and they need to be synchronised. When they all stop and when they all start, we’re talking about megawatts of powers going immediately up and down and apparently straining the grid significantly. It’s not just how much capacity you have, it’s also how you use the energy and how you learn how to better use the energy grid in new ways.

Nuno Goncalves Pedro

Indeed. Moving to the positive side, assuming this is not a bubble. This is really just the beginning. The bubble is really not there. I mean, one point that can be made is the classic Gartner Hype Cycle, because we’re going so fast through these transformations. We’re already at the peak of inflated expectations, and we’re about to go into the trough of disillusionment or value of disillusionment, as I used to call it back in the day, in the next few months, even a few years, because we’re just going faster.

Nuno Goncalves Pedro

Now, the good news is if we’re going very fast through the cycle, the next part is going to be the slope of enlightenment, and then we’ll have the plateau of productivity. Anyway, the cool stuff is maybe it’s just because we’re going so fast, we’re going to go thrust through the cycle as well. Therefore, all this money then doesn’t go to waste because of speed, right? The money doesn’t go to waste because of speed in some ways, and we go on the other end.

Nuno Goncalves Pedro

There is also a little bit of an argumentation that there’s other S-curves coming. That there’s a lot of tech that is and can be deployed and new methodologies that can be deployed today that will give us that next exponential acceleration of innovation in the space. Do you buy it, Bertrand?

Bertrand Schmitt

I certainly buy the fact that we are in a very, very, very typical hype cycle. It has been maybe the strongest hype cycle I have ever seen coming out of Silicon Valley. There have been many, and they try hard to make them very big, very fast all the time. But this one was insane, to be frank, totally insane in terms of scale, in terms of speed. I can see that we are past that peak of inflated expectations. I think we are getting soon to the trough of disillusionment.

Bertrand Schmitt

The question is, how fast will we go to that slope of enlightenment? That’s really the question. Is it 3 months, 6 months, 12 months, 24 months, 48 months? I don’t know. What is clear, however, is that if you look at the story of AI, there has always been some AI winters. Will it truly be a winter? I’m not sure. I think there is that belief that we have delivered something new. A new technology, a new layer of infrastructure has been put in place, that we will all benefit. In a way, you could argue it’s like laying down fibre in some ways. Now we have GPUs everywhere. We have put new tools to make them at scale.

Bertrand Schmitt

As you said before, definitely we have an issue that it’s not like fibre, that once it’s laid down, you can keep using it and improve it on both ends. Here, you have to replace the GPUs, and we know they improve very fast. An investment of $1 billion today might be worth $200 million in two years from now. There is that issue. But I feel we have laid down some very important foundations, and that is what is getting me probably quite excited for what’s coming soon.

Nuno Goncalves Pedro

Instead of a winter, which is three months, it might be like a Beijing autumn, which is like 1-2 weeks. It’s going through Beijing autumn. Best time of the year in Beijing. Very short, though.

Bertrand Schmitt

Very short, yes.

Nuno Goncalves Pedro

Very short. Other points that have been made is that there’s no bubble, because there isn’t a huge market in technical risk, that a lot of this is tagging along solutions set that are well known. They’re going to have immediate applications, et cetera. I have some view that maybe that is true for some of the stuff we’re seeing around consumer. Maybe that’s the reason of success for something like ChatGPT, because it really tapped into something that’s pretty core, that consumers are figure out, like question and answer chatbot interactions. I’m not sure that would be a great justification for why there isn’t an actual bubble. There still might be.

Bertrand Schmitt

Yeah, on this one that there is no tech risk and no market risk, I think it’s missing the point, because there is no tech risk, quote-unquote, yes, if you have billions to spend and if you just look at the current scaling law for foundation models, yes, you can keep improving by adding more data, adding more GPU. You will get better metrics. But as we discussed before, it’s still missing that piece around significant new change in algorithm in order to get to a totally different level. I believe that we don’t know, we don’t have a clue how to do the next significant jump.

Bertrand Schmitt

I think for me, it’s quite different if you compare to the history of computing. There is always some level of path to the next gen of CPU, the next gen of GPU, the next gen of modem. We know how to optimise the build-up, and we have passed to get there. Even if some people are pessimistic once in a while, there is always some new stuff coming up and some stuff that you see where it’s going, and it could help you in 2 or 3 years. I don’t think it’s true in AI. I think in AI, what we know is how to scale the current models approach, but we don’t know at this stage what could be next. That’s a layer of uncertainty in tech risk.

Nuno Goncalves Pedro

The product market fit part of the question depends on the product aligning well with the market need. I think there are areas where, again, we’ve already seen it, the channel is very obvious. I will use this because that’s just a different channel for me. It just gives me a better response and a different response. We’re talking about the consumer side and the ChatGPT use cases, for example, as just one of the examples on that. But then there’s other areas where it’s still very unclear, that business use level and where does it integrate and how does it use the data sets. Things are still moving very, very rapidly there.

Bertrand Schmitt

Yeah, I think the no market risk point makes no sense. I mean, it’s like saying that, “Yeah, because now everyone has a computer, everyone has a phone, there is no market risk. As long as it’s cloud AI, you can scale it to the world in a matter of months.” Yes, that’s true. ChatGPT demonstrated you can scale very quickly. At the same time, it’s clear they have retention issues. It’s clear they have usage issues.

Bertrand Schmitt

For me, it’s clear that there is still, as you say, a product market fit question for the foundation models as well as for apps on top of them, because it’s not as if we have seen thousands of them scaling. There is still a market risk, but yes, the underlying market is there for the taking.

Bertrand Schmitt

But it’s true of many things in tech today. It’s true of SaaS, it’s true of so many things. Yes, today you can assume that you have billions of smartphones, hundreds of millions of computers, and that they can be used to access your products, whether it be generative AI or something else.

Nuno Goncalves Pedro

I think also we’re potentially not in a bubble. We’ve mentioned this before in different ways. Obviously, we’re sharing some of our readings like the foundation Capital article and the Sequoia article, but it aligns well with some of the discussions we’ve had before at Tech Deciphered.

Nuno Goncalves Pedro

The build-up of this app economy, that’s how I would basically frame it. The fact that we have these foundational models that are more cheaply accessed today, but were sci-fi several years ago. The fact that as a startup, I can now figure out really more around what’s the application of the models. It can do a little bit of algos or algorithms on top of it, but I can really focus on the application of them and go after very specific use cases that generate value to customers in that space. For example, in B2B, to consumers in that space in the case of consumer, I think is a really good case.

Nuno Goncalves Pedro

I’ve mentioned several times. I do think there’s an app economy emerging. It’s an AI app economy. There’s good things about it and there’s bad things about it. Maybe we will get back to the bad things in a second.

Nuno Goncalves Pedro

But the good thing about it is there’s value. There’s value add there. There’s going to be SaaS companies that are going to emerge that are going to do well. There are going to be consumer apps that will emerge that will do well. There won’t be a lot, but there will be some. That’s one of the key pieces that I think is quite exciting about where we’re at. Now we have platforms that everyone can lever or leverage to take it to the next level.

Bertrand Schmitt

I totally agree with it. Myself, I’m very excited. In a way, I feel some of us are getting for free hundreds of billions of dollars of investment.

Nuno Goncalves Pedro

Well, some of us are investing as well, so it’s not totally free.

Bertrand Schmitt

To be clear, the models we’re having I mean, take that GPT-4, take Llama 3.1, I mean, it will have been sci-fi two years ago. Just two years ago, this will have been complete sci-fi. You will have asked people, When do we get this functionalities? People I’ve said maybe 10 years, maybe never. It’s clearly amazing what we managed to build. We have said a few times now that we don’t know where the next Leap in LLM is going to be in order to really improve the intelligence.

Bertrand Schmitt

At the same time, they Definitely, new features have been released that significantly change the game, like Multimodal models. Now you can combine not just text, but video, audio, presentations as input, and you can also generate that. It’s very exciting, and it certainly expands the type of products you can build as a result.

Bertrand Schmitt

Two, there are some new technologies like multi-agent systems, like new model architecture that are happening and that are being in place. I must say that there are stuff that feels next level. The question for me, if I’m putting my entrepreneurs at, my tech would be, today you have some new tools.

Bertrand Schmitt

If you had started a startup, let’s say, 15 years ago, your new tools in front of you would have been either cloud computing or mobile platforms. If you had started 20 years ago, your new tools would have been computing at scale, you would have new database capabilities.

Bertrand Schmitt

My point is that as an entrepreneur, you should consider some of these foundation models as new tools that you can leverage. I don’t think, personally, you should reinvent the wheel at this stage. You should let big tech spend hundreds of billions of hard-won dollars in building new foundation models at a relatively cheap cost for you and think about, “Okay, I have these foundation layers, there are customers in front of me. How do I bridge a gap between the two? How do I bridge a model that in itself has some value, some use?”

Bertrand Schmitt

But practically is very brutal. It’s very method in what it can do in terms of what it can truly solve today in the sense of stuff that is repeatable, that is useful, that can significantly increase human work or even replace human work and run with it. I think that’s the right approach, and I think that can generate actually a lot of returns.

Bertrand Schmitt

I mean, if you compare again with SaaS, at the end of the day, yes, the hyperscaler built a lot of value, generate value for themselves, but also the new class of SaaS companies to develop, to expand at a pace never seen before. I think that’s something similar that we’re going to see in the years to come. Companies that smartly bridge a gap between foundation models and real customer needs.

Nuno Goncalves Pedro

Totally in agreement. What’s our take, Bertrand? Do you want to go first? Do you want me to go first? Do we think we’re in a bubble or not?

Bertrand Schmitt

I guess I will say like you started, yes and no. I think we are in a bubble from a valuation perspective, amount-invested perspective, and expected return for the same actor perspective. At the same time, it might be the right thing to do. If you are the meta of the world, can you afford to stay on the side? Can you afford to depend on somebody else’s platform to run your business in the years to come?

Bertrand Schmitt

Probably not. You might have no other choice. You know what? Good news. You are printing a lot of money, so you have to use it. But it means also that your margins might decrease. On the bad news side, for sure, if you’re an investor who put a lot of money in some public companies or private companies, and the expectation that it might radically change our business? I don’t know.

Bertrand Schmitt

What I know is that there is a new business that is going to grow up. Is it OpenAI that will transform it business model? Is it some new actors that will, again, bridge that gap between foundation models and real customer needs? I don’t know, but personally, I would bet on some new startups, definitely.

Nuno Goncalves Pedro

My take is, again, I gave my take at the beginning, yes and no, we’re in a bubble. I think there are areas where we are in a bubble. Large language models, the guys who are going to disrupt the space around LLMs, et cetera. Not sure how are you going to make money out of that in particular, as you said, because the incumbents have a huge incentive of just putting the CapEx in and whatever. This is very interesting because it reminds me of telecom. It’s like when You launch 3G, and you launch 4G, you just have to do it because that’s the only way. That’s the only way.

Nuno Goncalves Pedro

These guys are doing the same. It’s like, “Okay, it’s Gen AI, we’re going to all do it. We’re all going to do Gen AI. It’s like the next platform for everyone to be served by us.”

Nuno Goncalves Pedro

In some ways, that’s the analogy. Pockets like, for example, large language models and the building of foundational models, I have a lot of skepticism around that. There maybe some areas that are less broad, that are more verticalized, that are interesting. I think in infrastructure, it’s badly distributed. The money is all going to NVIDIA, but we need more innovation in that space.

Nuno Goncalves Pedro

It’s clear that there will be stuff around FPGAs, ASICs, et cetera, that will emerge, that will in some ways address the methods of the future and will help us with things like inference and stuff like that. There are definitely a lot of things happening in the InfraSight that I believe it’s not in a that I believe it’s not in a bubble, but it’s just inadequately distributed.

Nuno Goncalves Pedro

It probably will still grow even more. There’s probably the need for more innovation there. On the app side, on the application side, I think there is a bubble. Again, it’s a bubble that is maybe unevenly distributed. There are companies that are raising a ton of money, and they are literally just apps. I’m not really sure. I don’t want to diss any specific investors, but I’m not really sure if their investors have figured it out yet.

Nuno Goncalves Pedro

If they’ve been confusing capital intensity due to actual development of platforms with the capital intensity of becoming a successful app, which are totally different things. One is a channel discussion, the latter, and the former one, it’s a channel discussion because it’s about getting users, customers, et cetera. Whereas the former, it’s a discussion around technology differentiation and moat.

Nuno Goncalves Pedro

If you’re in the ladder, and you’re putting a ton of money into a company, you should know. Because you might still not do well, and you don’t have a tech moat. Cool. 

Bertrand Schmitt

Yeah, it’s actually dangerous because you might create really bad habits and stuff to come back. It might not be clear at all. It wasn’t a necessary investment.

Nuno Goncalves Pedro

That’s where the subsidization to NVIDIA happens It’s probably at the most significant level. It’s not just subsidization to NVIDIA, it’s going to be subsidization to NVIDIA. It’s going to be subsidization to whatever cloud provider they have, to Microsoft, to Amazon or to Google. It’s going to be subsidization for the models that you’re using. You’re basically passing all your money as an investor to the startup so that they can pass it along to the entire value chain of big tech. It’s a waste of money.

Bertrand Schmitt

To be clear, I believe NVIDIA is a really fantastic company. I mean, it’s a company I follow for decades. It’s an amazing company. I mean, they made the whole AI revolution possible with invention of the GPU and leveraging the GPU as an AI computing unit. But yeah, a lot of money is going their way because everyone has to spend. But in some ways we are lucky they are there, and they enable all of this. I don’t know, to be frank, if it would be easy for anyone to compete against them because they have been at the game for a long time.

Nuno Goncalves Pedro

No, but people can… Again, the big guys, the apples of the world, not unknown for doing their own Silicon. At some point, might start doing their own Silicon. Why not?

Bertrand Schmitt

They are already doing their own Silicon.

Nuno Goncalves Pedro

That’s begs to say, I mean, I’d be shocked if Google is not looking into it. They’ve been doing a lot of stuff around ASICs for many years now.

Bertrand Schmitt

They have their own TPUs, Google.

Nuno Goncalves Pedro

At some point, these players, let’s see what comes out of it. In conclusion, this is episode 57 of Tech Deciphered. Are we in a Gen AI bubble? The conclusion, as you guys just heard, is we are and we aren’t. Great conclusion, but a nuanced conclusion nonetheless.

Nuno Goncalves Pedro

We went through the state of AI and generative AI. We discussed the case for the bubble, so the negative case. We discussed the positive case, the case that we are not in a bubble. This is just business as usual. It makes eminent sense. Maybe there should be even more investment. Finally, we shared our take on the bubble, whether we believe that we are in a bubble or not. Thank you for listening to us today. Thank you, Bertrand.

Bertrand Schmitt

Thank you, Nuno.