Superintelligence, Bubbles And Big Bets: AI Investing in 2024 | Matt Turck & Aman Kabeer, FirstMark

We are back.

So for anyone that listens to The MAD Podcast on a regular basis, you will know that the usual format is that we interview great founders, researchers, or analysts in the data and AI space. But occasionally we serve as talking heads. We did this last time back in March in connection with the release of the 2024 MAD Landscape.

And we had fun doing it. So as we are, crazily enough, towards the end of 2024, a few weeks before the end of the year, we thought it'd be fun to check in again. And so what we're going to do today is a little bit of an overview of what's happening in AI today, seen from a VC perspective, the news stories that really caught our attention, some of the things that we see in the market from a valuation or dealmaking perspective, some of the big questions that one should probably think about towards the end of 2024 and where we're at, and what we're excited about as venture investors.

Sounds like a plan. All right, so maybe just to start with, it continues to be crazy in the world of AI. It's really interesting, isn't it? We are almost two years into the ChatGPT era. ChatGPT came out on November 30th, 2022, so we're almost at the two-year anniversary. And often you see those hype cycles sort of slow down after a bit, run out of steam. And here it's totally not the case. In fact, it seems to be accelerating.

So everything is big this year. That's the year of big valuations, big numbers, big announcements, big expenditure. $6 billion at a $157 billion post-money valuation.

Boomerang acquihire.

And all those numbers are being thrown around. We're clearly in this phase of huge investment in AI infrastructure. The three top companies, Meta, Google, and Amazon, are on track to invest $200 billion in AI infrastructure this year. One of my favorite things over the last few weeks is Masayoshi Son, the CEO of SoftBank, mentioning just a few days ago that, in his estimate, to reach superintelligence, there was going to be a cumulative CapEx, capital expenditure budget of $9 trillion, which, in typical Masa fashion, he said was actually probably a pretty small number.

Elon, not one to be shy about big bets, he's building the biggest supercomputer of all time. So xAI's Colossus, which is based in Memphis, which, by the way, apparently was built in 122 days. So the fastest time ever to build a data center. Typically those things take years. In true Elon fashion, what-have-you-achieved-last-week kind of approach. They apparently did that in 122 days with a very small team.

100,000 GPUs at first. And then he announced just a few days ago that he had doubled that, or he was in the process of doubling that to 200,000 NVIDIA GPUs. And apparently the plan is to get to 300,000 Blackwell GPUs when they come online. So this is just the beginning there. Interestingly, that has some crazy requirements in terms of overall energy consumption and water. Apparently, it's like many million gallons a day. And in terms of overall energy consumption, it's basically the equivalent every year of enough energy for 100,000 homes.

Which is like a big part of the AI conversation in 2024: the environmental impact of this and the enormous requirements in terms of energy, which is something that I personally didn't have on my bingo card for 2024, has led to a reignition of interest in nuclear power in the U.S. So, growing up in France, this was not completely unusual. There was a lot of nuclear power, which, by the way, was not completely uncontroversial, but that was certainly a big part of energy generation in France.

But yeah, in 2024, it turns out that big tech companies and AI companies are the ones driving this revival of nuclear power, with Microsoft entering into a deal with Constellation to revive Three Mile Island, which was the site of the last nuclear accident in the U.S., and then deals from Google and Amazon. So, really interesting to watch.

Everything big. So obviously, the company everybody loves is NVIDIA. And clearly NVIDIA has had an absolutely wonderful ride in the last couple of years in particular. And it's staggering, right? It's a $96 billion revenue company that generates $53 billion in net income. But you could argue that it is very richly valued. Typically, you measure those companies on a price-earnings ratio, and typically for a mature company that's growing, you have about a 20 ratio. And for NVIDIA, last time I checked, they had a 65 ratio.

So there is baked-in kind of a general idea that they need to grow 3x to get to sort of where the market standard is, with the stock not growing. So it's a tremendous amount of expectation placed on just one company.

And speaking of which, it's going to be interesting to see what happens in the IPO markets in the next year or so. Everybody knows that it's been one of the most terrible periods for IPOs in recent memory, with barely any IPOs. But it seems that there is a generation of especially AI hardware companies, or AI hardware/compute companies, that are starting to think about going public or actually going public. So there's certainly Cerebras. And you and I wrote a little S-1 breakdown the other day that talks about our thoughts on Cerebras. That sounds like a very interesting company with very clear open questions around customer concentration, in particular the relationship with G42.

I think you and I were discussing the other day that apparently this is getting delayed a little bit.

Yeah. So we'll see what happens there. But a positive development in terms of an actual venture-backed startup going public, which is something that hasn't happened much, as we were just saying. It seems that others, possibly like CoreWeave, are very natural candidates for IPOs. It feels like there is this moment for AI compute and hardware companies to go public, given the crazy amount of interest and the crazy performance of NVIDIA. So we're going to see how that plays out. So that's public markets.

So, interesting, kind of frothy, arguably, depending on how you look at it, but really interesting. And then there is private markets, which is the world where you and I, as venture investors, evolve. And it's been pretty crazy as well.

And current chairman of OpenAI.

Yeah. So that's 225x.

Yes, we are back, baby.

It sort of depends on which side of the fence you are, right? So if your companies are getting bid up, it's nice from—at least it's a short-term victory from a VC perspective. But if you're the one paying the price for net new investments, it's certainly rich.

Yeah. Which in some ways is super exciting and shows a lot of positive energy and optimism towards the future. But equally, that's a lot of valuation to grow into. That's a lot of expectation of future revenue that gets baked into the price today. And what's really interesting is that this is just AI, and the rest of the world just doesn't work that way. And when I say the world, I mean actually just the software market, right? So the SaaS market itself, whether public or private, actually is still in that phase of being super reasonable and super rational, with valuation multiples being at the lowest in a very long time, actually still below the historical median.

So there was median, there was like this big spike in 2019, 2020, 2021, and then that crashed. But where we are now is still below where we were before that big spike.

So you had Palantir at 30x as one of the rare plays in AI, and the rest of the market, just to hammer it home, at 5x next-12-month revenue.

You and I were thinking about, hey, what does $3 billion today mean? Like, a $3 billion valuation. So on the one hand, it could mean that you are a pre-IPO company and you've been around for a while and you're SaaS, but you may even have some AI component to it. To justify a $3 billion valuation, let's give you generously an 8x NTM multiple, not the 5x we were describing. Let's call it 8x. That basically means you need to have $375 million in revenue expected for next year to be worth $3 billion in the eyes of public markets.

Meanwhile, some of the companies that we were describing on that chart—and again, we're not picking on those companies, and there's a reason why they were able to command those valuations—but there are several of those companies that have quite literally nothing, that have a vision and a founder. Some of them have a little bit of a product. Some of those products have not shipped yet, and that's $3 billion as well. So which one is it? And look, it's all fine for now.

It's kind of weird, but it's all fine for now. But at some point, all of this is going to need to converge. So either those AI companies that were very richly valued do have those astounding revenue growth rates and sort of grow into the valuation, or public markets need to start behaving in a rational way again and just no longer be at that 5x, 8x, 10x multiple, but actually start valuing companies based on—I was going to say based on vibes, but a little bit, right?

But on future growth and all the things. But something has to give. I was going to say the funny thing, but it's probably not that funny, result of what we're just describing is that, from our perspective as venture investors who sit on boards and spend a lot of time working hand in hand with companies, it sort of feels like you have a split reality in your day. Half of your time, you basically untangle the consequences of the excess period of 2021, where you think of how to help companies get fit and cut burn, extend runway, be efficient, all the things.

That's like one part of your day. And then the other part of your day, for people like us who invest in AI, you sort of live in that completely different reality where you basically are rebuilding another kind of bubble, I would call it, just like the one you spend the other half of your day untangling. So it's a little bit of a strange feeling throughout the day.

The B question.

Yeah, well, so as per the above, given the price and the very rich valuation way ahead of revenue in many cases, it certainly feels like it. But let's argue it either way. So let's start with the arguments for the bubble and then the arguments against the bubble. So look, in terms of arguments for the bubble, so that's the price. There's probably the fact that some of it is driven a little bit by circumstances in capital markets. So, yeah, certainly AI is very important.

It's a paradigm shift. But then separately from that, there is the fact that a lot of people have raised very large funds. Those large funds need to be invested somewhere. AI is pretty much the only game in town where there's growth and enthusiasm and velocity. And it so happens that this generation of AI companies is particularly capital-consumptive. $6 billion round. The minimum ticket size was $250 million. So if you're a very large fund and you need to deploy, and that's the only place you can deploy, it's sort of a match made in heaven.

That you have those companies that want so much money. So the whole thing sort of feeds upon itself, which you could argue is somewhat artificial and a product of the times.

Yes, that's become the $600 billion question, named after a great essay by David Cahn. But not just David. Goldman Sachs had a report a few months ago sort of questioning the potential imbalance between demand for AI and the infrastructure that is being built. So as per the above, people are building and building and building. The question is indeed whether people will have a need for AI that will match what is being built. And what's particularly hard to nail is the timing of those things, right?

Because to build a supercomputer, to build a data center, to build chips, all of this are two-, three-, four-year commitments. This is a lot of hardware that you need to build and put together and order the parts and all the things. And that's a particularly sort of unforgiving kind of cycle. So if you mistime demand, and even if you get the right amount of demand eventually, there's a gap between when that demand actually materializes and all the money that you spend upfront, then you could find yourself in potential trouble, at least short term.

Yeah, absolutely. And then part of the whole demand conversation revolves at least in part around, hey, AI has had this crazy exponential acceleration over the last two to three, four years. The big question is whether that's going to continue. Everybody has an opinion for sure, but nobody actually knows that for a fact. So it's the whole conversation around scaling laws. And scaling laws are really that idea that with all this deep learning and transformer-based models, they have been able to scale dramatically in terms of performance as you fed them more compute and more data and more compute and more data.

And largely, without denying any of the algorithmic improvements, but largely what has made the more recent models better than the prior generation of models is just more data, more data, and more compute. So the whole scaling laws discussion is really like, okay, well, if you feed even more data and even more compute, so we're building those super impressive supercomputers, if you feed more compute, is that going to continue, or do you need a dramatic change in algorithm to get to the next generation?

Right. Is GPT-5 going to be much better than GPT-4? Is GPT-6 going to be much better than GPT-5? And yeah, like I said, I think nobody knows. I mean, just for the anecdote, I got a chance to ask that question to Sam Altman a few weeks ago. I was at an event at OpenAI. Not surprisingly, when I asked him, will scaling laws continue? He looked me straight in the eye and said, absolutely. Which, of course, you would expect.

And look, he's one of the most iconic entrepreneurs of our time and all the things. I'm not going to question anything, but I would probably venture that while more informed than all of us, even he doesn't know, because you're very much at the frontier of science. And we all hope that it's going to continue. But that's going to be a big question. And there's a whole different part of the AI world that basically says, like Yann LeCun, but others as well, that say that all these big models are still largely sort of rote repetition, brute-force kind of things.

And until we really combine them with different methodologies, in particular reasoning, the true progress is not going to continue at the pace we would expect.

We're not in a bubble. All right, so let's see. I guess first of all, you could say that many of the AI startups are doing incredibly well, which seems to evidence some level of demand, even more so, like rabid interest. And if you look at OpenAI, for example, a lot of people have argued, including obviously all the investors, that actually the valuation is not that rich if you think about it. So first of all, the company is growing extraordinarily fast, and the company has extraordinary ambitions to be at $100 billion in ARR in 2029, which is not that far away.

5x. So, to the discussions that we were having a minute ago in terms of market multiples, it's actually in the grand scheme of things not that high. So that's the argument.

And then there was some Stripe data that was interesting as well.

That's for the AI startups. From a Mag 7 perspective, it seems there is tremendous demand as well for AI. And so there's all sorts of things in the press, in particular, that's interesting. Both NVIDIA and TSMC were saying publicly recently that the demand for Blackwell chips was absolutely insane, and that, very far from seeing any kind of slowdown, in fact demand was more insane than ever. And then it's earnings season. So, just in the last few days, all the Mag 7 talked about AI, and there's plenty of interesting statements.

We're five years in, but our AI business is on track to surpass $10 billion in annual revenue run rate in Q2. Azure OpenAI has more than doubled over the last six months. GitHub Copilot increased 55% quarter over quarter. So that's Microsoft. Google had some interesting thoughts as well.

Yeah. Although they got a little bit of heat on Twitter, right?

And then Meta was very bullish as well. AWS as well. Another quote, this time from AWS: AWS's AI business is a multibillion-dollar revenue run-rate business that continues to grow at a triple-digit year-over-year percentage rate and is growing more than 3x faster at this stage of its evolution than AWS itself grew, which is fantastic. Equally, you could argue that by AWS standards, even a multibillion-dollar kind of business is sort of peanuts, considering AWS is at $110 billion in revenue.

As investors, we want to track the trend rather than the absolute number, but it's still very small. Yeah. And then, to echo what we were saying about the progress in AI, so on the flip side goes the argument for why we're not in a bubble. You could argue that actually the pace of progress in AI just doesn't seem to be slowing down at all. And some of the stuff that we were talking about, the interest in reasoning, there's already some element of that happening at OpenAI with o1, which introduced a certain level of reasoning capabilities at inference time that enables the model to do some really interesting and much-improved, in terms of performance, things at inference time around math and geometry and those kinds of things.

And then there seems to be a new model every five minutes.

Yes. A lot of focus on audio, video, images beyond text. So that's really interesting. The multimodal stuff is really interesting. Pace of progress in open-source AI is very interesting. Everybody's waiting for Llama 4, which is supposed to come out in December.

Next month. Yeah. So plenty of things happening. And then the prices of tokens dropping dramatically.

So it's sort of what everybody was expecting and everybody was talking about a year ago. But for once, that was actually true. I think the price per token of GPT-4 dropped something like 90% over the last year, which is fantastic for anybody building on AI, makes generative AI viable. That was certainly one of the open questions. But on the other hand, may not be that great for OpenAI in terms of revenue or margin or performance of the business.

But you would expect that lower price would lead to much more widespread usage and everything working out in the end.

Yeah.

Com and so on and so forth. But on the other hand, if you hadn't invested at all during that period, you would have missed Amazon, which was founded in '96. You would have missed Google, that was founded in '98, and you would have missed out on Salesforce, which was also created in '98. So there's a little bit of, like, play the game on the field, which certainly is very scary because, for sure, billions of dollars were incinerated in the dot-com bubble or the dot-com bust.

And it's possible, perhaps likely, that this is going to be the same thing here, that the same thing is going to happen here, where another few billion dollars are going to be incinerated in this moment. But equally, the next generation of companies is probably almost certainly being built right now. I don't want to say next generation of incredible companies is being built right now.

Yes. And then, look, leaving the whole financial potential discussion, there is the argument that we are perhaps building something of extraordinary importance in the form of AGI. And that's sort of the Sam Altman argument when he said a few months ago that he actually doesn't care if he burns $50 billion a year because the pursuit of AGI is so phenomenally important from a humankind perspective that it's really not about the money.

Hopefully by the time you become a senior VC.

And look, by the way, on the AGI front, the whole discussion was kind of fun, which is that everybody's talking about AGI. So certainly Sam Altman, but Elon said we were like two years away from AGI. But when you dig, nobody actually agrees on the definition of AGI. So if you're DeepMind, you say, well, there's six levels. But then Yann LeCun again said he disagreed with that definition of the six levels. There was someone in the press recently, Fei-Fei Li, who is the godmother of AI, that also said that she doesn't really know what AGI is.

So directionally, everybody agrees that it's sort of kind of human-level intelligence and the ability to generalize and learn, but nobody kind of agrees with it. And then you could argue that AGI is maybe not even the thing anymore.

New and improved. Yes. And yes, sort of the term du jour. And we were talking about Masa a few minutes ago, and that's his thing now. Like he said, he was sort of born, like his mission in life was to usher in the era of superintelligence. So, yeah, that's a Trump lookalike. So again, we're poking gentle fun at all of this, but we love the space. So ASI is this thing. So if you're a new and ambitious company, you need to have superintelligence in your name.

So there's Safe Superintelligence, Ilya's company. There is another company called Mathematical Superintelligence. We saw a handful of others. And it's the same thing as AGI. Nobody really has a definition of what artificial superintelligence is. There's directionally that general kind of agreement that superintelligence means intelligence that's better than any human, and in particular, the ability for AI to create more AI and sort of self-perpetuate. So, sort of AGI creating more self-AGI, which would be characteristic of superintelligence.

That's at least 10x more. But look, a little bit beyond the definition of AGI versus ASI, there's an interesting question from our perspective as investors, which is, okay, even if all of this stopped, or if progress sort of plateaued, decelerated, does it actually matter? You could argue that actually, just with what we have, there is plenty we can do, plenty of great companies we can build. Just taking the current state of models and sort of deploying them in a way that's integrated in workflows and verticalized per industry sort of feels like there's a long way to go.

Yeah, it's sort of unclear what happened to doomerism in the last year or so. That was, like, the dominant discussion not that long ago. I wonder if it's just a function of people just playing around with all those tools, including ChatGPT, and sort of realizing the limitations of it, in that it appears superintelligent, but it's not. It's extraordinary in many ways, but it's not a form of intelligence. So maybe it's that, or maybe it's just people actually trying to deploy AI in their enterprise for their specific kind of use case and realizing that it's actually hard.

And actually, even deploying AI, there's a lot of things that have nothing to do with AI itself. You need to get procurement authorization and compliance and all the things. So far from killing us all, just, like, deploying it in your department is a struggle. So look, where does that leave us in terms of our own investing? So just sharing the thoughts that it might be interesting for people to hear how people like us think about the space. And that's us, but that's a lot of conversations that VCs have, that founders have, that VCs have with founders.

So just, like, a general sentiment of what we think, what people think. So maybe taking the layers of the cake. So it's become sort of, like, a standard that people think of AI in terms of, okay, there's the very base layer, which is, like, sort of compute and chips, and then there's a layer around models, and then there's a layer around tooling, and then there's a layer around applications. And you can divide the application layer between consumer and enterprise. So taking those different layers of the kind of cake in terms of what we're excited about.

So this is the bottom layer, which is, like, compute and chips and all the things. And we sort of talked about that. We feel that from a venture perspective, it's probably outside of our mandate because it requires a specific type of expertise and then very large amounts of investment. But we talked about Cerebras, we talked about CoreWeave, we talked about NVIDIA. Plenty of interesting things happening at that layer. Then above that, there is the model layer. And you could subdivide that layer from frontier general-purpose models on the one hand, and then on the other hand, more specialized models.

And starting with the frontier models, it feels like the market has started crystallizing around a handful of early winners, certainly OpenAI, but Anthropic, arguably Mistral. And that space continues to be fascinating to watch. But from our perspective as venture investors, that's another space where we don't think that is really consistent with our mandate as investors. It certainly doesn't mean that we think those are not good investments. Quite the opposite. Those could be some of the most important companies of this century.

But equally, they require super-large amounts of investing. We're talking about billions, as per the above. There's a little bit of a question around differentiation as well, which is that every other week it seems that one model is better than the other model. Just a couple of weeks ago, NVIDIA announced that in addition to having the best chips, they just released a model that was actually better than GPT-4. So even outside of the usual suspects, there's stuff that comes left and right with better models.

Yes. So, yeah, unclear. Lots of competition, unclear how you build a competitive advantage. And then the question of open source, which I think could be a friend, could be foe, depending on how you describe it and where you are in the stack. But certainly from a frontier model perspective, the fact that Meta is putting all its weight behind something like Llama 4 has to be frightening. So that's how we think about general-purpose models. Now maybe let's talk about specialized models. What do we think?

Yeah, absolutely. H is building effectively AI for RPA, so taking the world of enterprise automation through agents and all those things. So, foundation model, but focused on one specific type of use case, at least for now. Okay, so those are specialized models. So, still on our journey, completing our journey through the upper layers of the cake, the layer above the models would be tooling. And tooling, that's a world of MLOps, AIOps, orchestration, basically how you build those products, orchestrate those products, manage those products, and all the things from a technical perspective.

So those tend to be developer tools. And it's a good old picks-and-shovels approach to the space. When you have a space that's exploding, investing in the companies that provide tooling to the pioneers in the space tends to be a good business. And certainly that was the strategy back in the day for our first investment in Dataiku back in 2016. So, Dataiku being this enterprise AI platform that now serves a bunch of Global 2000 companies, but provides this broad platform to do MLOps, but also governance.

So everything you need to do to create a model, test a model, deploy a model, and manage a model in production. So that type of approach. But yeah, that's an example. But there's been tons of new entrants into the space.

Yeah, which we have found to be both exciting and somewhat problematic, right? Some of the exciting frameworks of a year or two ago that everybody was raving about, now fast-forward to this year, everybody seems to dislike them. And so, look, all of this is fluid and evolves, and maybe those companies will come up with the next version of their product and everybody will love them again. But it's hard. And you and I were looking at the LLM evaluation space as an example.

And look, is there a need for evaluation in LLMs? Certainly. As we all know, those things are stochastic, not deterministic, meaning that they don't get the right answer 100% of the time, and you don't get the same answer each time you ask. Therefore, evaluating performance is more important, arguably, than any other kind of software. So all this makes a lot of sense. And as a result, there's dozens of AI evaluation companies, and they all started by very smart founders and sometimes raised money from very good investors.

Great cultures, and they build very good early product. But which one is it that's going to be the top company is incredibly hard to say.

Absolutely.

Especially with open source, which we love. But we've seen again and again tools in AI just have these dramatic rises in terms of GitHub stars. So, as we all know, GitHub stars is a flawed metric, but still interesting directionally, only to just lose their fans pretty quickly. So it's fast-changing, kind of fickle. And another area where we spent a good amount of time was vector databases. And same thing, right? There was this moment where you had a bunch of vector databases appearing on the market, specialized vector databases, and everybody's super smart and everybody's super thoughtful.

The only problem is, fast-forward a few months, then the general-purpose databases, like the MongoDBs of the world, started launching vector search features. So fast-forward to today, and now you have this tension in the market where the general-purpose databases are saying, well, our vector search is good enough for most use cases, on the one hand. And then on the other hand, you have all the specialized vector database vendors that say, well, actually, if you want to do real AI or anything that has any level of ambition, actually you need our specialized tools.

So the market is evolving super quick, and we're all trying to figure out how that unfolds, how that works as it unfolds. So that's the tooling layer. Another final layer of the cake is the application layer. So let's take consumer first, and then we could talk about enterprise. So for consumer, lots of really interesting ideas, intriguing stuff being done. So the general mental model for this is that each time there is a paradigm shift, there is a whole generation of consumer companies that appear that become the dominant companies of tomorrow.

Web 2.0, all the companies we know and love, right? Amazon and Google and all of those appear on the scene. If you think about the next paradigm shift, which was mobile and cloud, then you saw the rise of the Ubers and DoorDashes. And those are particularly interesting because there are native mobile companies that could not have existed without the very specific capabilities of a phone, right? To order an Uber, you need to have a phone with a GPS and be able to call the car and pay for it and all the things.

So, native mobile companies. So then the question is, okay, for this paradigm shift, which is generative AI, what are going to be the native AI companies that are quintessentially generative AI companies that could not exist without the capabilities of AI?

Yeah, exactly. And look, those products are absolutely amazing. And you and I use them all the time, every day. You could argue whether those are truly generative AI, sort of like net-new ideas, versus doing search better with a different kind of interface. Ultimately, you could argue either way. One company that we discussed a couple of times on The MAD Podcast that could be an early candidate for what a truly AI-native consumer app might be would be Suno AI, which is an AI music generation company.

So if you think of Spotify being the leader in the last generation, meaning a platform that's sort of a marketplace, ultimately, where you have this never-ending catalog of music available at your fingertips, then fast-forward to Suno. Suno would be, well, don't use somebody else's music, but generate your own music, right? I want to listen to something that sounds like Bach, but with an EDM beat and lyrics that talk about the autumn leaves. I'm just making this up, but that feels like a really interesting entrant.

Yes. And I think Richard was saying, and I've heard variations of this as well, the billionaire test is an interesting metaphor for the power of democratization of technology. If you think of what a billionaire was able to do that a kid in the jungle was not able to do, to take two extremes, the phone became the great equalizer in that whether you are that billionaire or that kid in the jungle, you basically have the same phone and with the same exact capabilities. So, using that mental model or metaphor to apply to consumer AI.

What is it that a billionaire can do today, or very rich people tend to do, the kind of privilege that they have that the rest of the world doesn't have? And could generative AI democratize that? So you can think of a long list. And we did that exercise. If you're a very rich person in New York and your kids go to school and you want to best equip them to go to the best universities, you're going to give them tutors across topics.

And that's something that you can afford because you're rich, but everybody else cannot because they're not. And following that line of thought, there's a whole generation of AI tutors, AI tutor-in-a-box, that offer very AI-driven, personalized experiences to kids so that they can improve on whatever topic. So that's one example. But you could have plenty more. You could have EAs, you could have personal concierges, personal travel agents. Even your high-maintenance spouse could be disintermediated or replaced by an AI girlfriend or boyfriend.

It's a brave new world. So, bottom line, so much to build, and it feels like we're very much in the early innings of all of this. We've seen a lot of interesting things. AI psychiatrists, for example, which, by the way, is already a use case for ChatGPT. It is not supposed to be that, but a lot of people use ChatGPT to just sort of vent and talk. But we've started seeing a generation of psychiatrists-in-a-box trained on actual therapy-to-patient conversations, which is just one more example. It's super interesting.

And we think there's going to be a lot there and a lot to build.

Yeah, applications, agents. Applications.

Yeah, exactly. Unpacking that, that basically means everybody tried, but people sort of struggled to find value. And that's within our group of startups. So by definition, people that are super tech-forward tend to use the latest tools, be very curious to try new things. So the fact that there is that gap is really interesting. And then if you look at the rest of the world, sort of the Global 2000 companies, call it, and government institutions and all the things, it's actually fascinating because I just saw some stats recently that show that there is a big gap between what individuals say in terms of adoption and usage of AI and what organizations, institutions, enterprises say in terms of adoption of AI.

So the stats I saw is that for ChatGPT in particular, 39% of Americans said that they use ChatGPT, including 28% that said that they use it at work, and 11% use it every day. So real usage on a daily basis at work. And then there was another stat that said that only 5% of American businesses say that they are using AI technology to produce goods or services. And that was U.S. Census Bureau stats. So, look, you could argue the data, but I've heard variations of this, which is that people sort of almost secretly use AI.

So it's so interesting to unpack, right? Because there's a little bit of bottoms-up adoption where we've seen in technology, people use the tool first, and then the whole strategy is that once you have a bunch of people that use a tool, then the vendor comes in and tries to sell enterprise contracts and all the things. So we've seen this time and again: people use the product first, enterprises use the product second. But I think on top of that, there's something around behaviors that is profoundly interesting in that people don't know whether they should be bragging about using AI at work or whether they should be hiding the fact that they do.

I mean, does that mean that you're not supposed to do that? Is that cheating? Is that trying to deceive your coworkers by giving work product that you should have been doing, but the AI really did? Does that mean that you should be fired or that you no longer should get your salary?

Absolutely, all those things. So that's a world of people at work. And then the company, the enterprise, is much slower. So what we've seen talking to a bunch of different companies is that there's sort of like two parts to this. There is one part which is like, okay, over the last 12 to 18 months, people have been overwhelmed by questions and pressure from their boards as CEOs and boards said, hey, we really need to do AI. What does AI mean for us?

Right? Nobody wants to be caught flat-footed for the next big revolution. And especially all the CEOs today are people that were probably like 20 or 30 or even 40 in the last dot-com, and they saw the disruption and what it did to companies. So now that they are in command of their organizations, they don't want to be the people that go down in history as not having seen the power of generative AI. Right? So all boards said, hey, we need to do generative AI.

And then I think people that report to those boards and CEOs sort of rushed to figure out a way to do AI and to experiment with AI, which is a great thing. And they've all defaulted to sort of what I would call, and not in a pejorative way, but I would call the low-hanging fruit, sort of the single-use, kind of obvious applications that are mostly easily applicable and deployable.

Yeah, plug and play. Exactly. So that would be search. That would be search either in sort of AI-powered search or like chat-kind-of-search. So that could be the world of ChatGPT on Azure. It could be the world of Glean, which is getting a lot of heat in the market these days. I mean, certainly all the copilots for GitHub Copilot and developer assistant, engineering support, coding kind of things, because that's targeting one very specific use case and it works.

And then even in our portfolio, Synthesia, which is AI for video generation, famous for its avatars, another low-hanging fruit. So it's obviously a fantastic company getting plenty of traction and doing something very important, but also at the same time very easy, right? You sign up for Synthesia and then you can create your video, you import your company's kind of color palettes, and then boom, you're doing generative AI. So that's the world of low-hanging fruit. For anything else that's a little more complex, then it's been tough.

It's been, let's call it, very early. All right, so what does generative AI mean to me if I'm a pharma company, if I'm a bank, if I'm a manufacturing company? How do I use it for my specific use cases? Not chat-search kind of thing, but what does that mean for my supply chain or for production, like analysis of patents?

That is exactly right. That's exactly what happened. So people did what you would assume, which is like, okay, well, let's turn to people that can do the work, that can spend hours and hours interviewing vendors, interviewing our people to understand use cases and sort of match use cases with potential vendors and all the things. So it's been a banner year for the Accentures and the McKinseys and the BCGs. And then famously, Accenture reported $3 billion in annual fees, which made the headlines and all the things.

In some ways, that's the irony of it, right? Like, one of the highest-grossing companies in generative AI today is actually Accenture in terms of revenue. So look, leaving that aside, you could argue that Accenture is like at $80 billion-plus in annual revenue. So actually, three is still a small number. So let's not get completely carried away. But directionally, that was kind of funny.

That's exactly right. That's exactly right. But yeah, so in the enterprise, for that second part, there's a lot of consultants. There's a lot of pilots still. And I think everybody was hoping that 2024 was going to be the year of generative AI in the enterprise. But it's probably taking some more time. I mean, look, it's definitely happening. And we have a nice vantage point in that, again, we work with Dataiku, which is the leading pure-play enterprise AI company, which has a whole generative AI part of their platform called the LLM Mesh.

And through them, we see that people are actually starting to deploy the thing at scale. There's a number of Dataiku customers, some of which are public and on their websites, many of which are not, that do use the generative AI capabilities orchestrated by Dataiku to do things in production. But it's starting to happen. It's early. We're still in the world of pilots. And look, I get it. If you're an enterprise, so many questions. There is: how concerned should we be about errors, hallucinations?

What does that mean in terms of, hey, is there liability exposure? Then there's a whole series of questions around regulation, the EU AI Act. What does that mean for us? How's that going to evolve? And then there's perhaps, most importantly, the realization that to actually do AI, there's a whole amount of work that you need to do around getting your house, your data house, if you want, in order. Because if you're going to do RAG and all the things, sort of feed your data to an AI, or at least create guardrails for AI with your data, your data better be good.

So where does it live? How do you grab it? Who has access to it? So the whole big data, data infrastructure, modern data stack kind of stuff that you and I have been talking about for years, and also on this podcast, comes back in play. And there's a lot of work, and some companies have been doing it for a very long time. Some companies are still kind of early in figuring this out, right? So it's hopefully not a whole reengineering, but this taking your whole data infrastructure to the next level kind of conversation that needs to be had.

Yeah, there is that whole discussion. Look, we think the answer is no. We think that SaaS is not going to be dead. We think that SaaS is just going to have to evolve dramatically. And if you think of SaaS being largely a wrapper around a database, and you could argue that's what Salesforce is, a lot of workflow on top of a database, the next generation of SaaS is going to be a wrapper around intelligence, with workflow, with integrations, and all the things.

And look, that evolution is going to happen. Some existing vendors are going to do it, some net-new vendors are going to do it. But we think that intelligence is going to be everywhere, which is kind of funny, by the way, because we're effectively saying that the future is wrappers. And in the last year, every VC would say, well, I don't want to invest in wrappers because where's the defensibility and all the things? And I think the collective thinking has shifted pretty dramatically.

And now we've seen a bunch of wrappers specialized for this industry or this problem actually grow very fast. So now everybody wants to invest in them.

Yeah, absolutely. And part of the reason why it's interesting is that you could be a direct beneficiary from whatever next breakout in AI happens, right? If ChatGPT goes from GPT-4 to GPT-5, I guess there were rumors at some point that GPT-5 was going to come out in the next few weeks, GPT-5 Orion. But apparently that's not happening. But whenever that comes out, your application will automatically get better. You'll be powered by this. So great for OpenAI, but also great for you as an application vendor.

And as it turns out, there's so much work that needs to be done around integration and workflow and all the things that you can still build a very valuable company. So what you were referring to, the thick wrapper, I think the layer is indeed pretty thick. Yeah, absolutely. So, yeah, we think there's very fertile ground there. And there's a lot of reinvention happening. There's a lot of tech-specific business problems, like horizontal, vertical, per industry. There are AI solutions appearing everywhere, which is going to make our work for the 2025 MAD Landscape even more exciting.

And I know you're looking forward to it, as I do. But yes, this AI—3,000 logos this year.

Exactly. We can do it: AI SDRs, AI HR, AI for finance, and then AI for lawyers, and then AI for accountants, and AI for bankers. We saw a bunch of those, AI for consultants, and some of them are in copilot mode. Many of them are evolving towards being sort of agents in agent mode. So agent meaning, in that case, just completely replacing the workflow, with a view to effectively being autonomous. So plenty to build there, and fascinating implications.

And it looks like we should probably start wrapping up before this gets too long. But fascinating implications around what that means in terms of business model, in terms of what those companies really are. That's like that whole discussion around selling the work, which I think Sarah Tavel at Benchmark started last year, which is very insightful, which is at some point you stop being software and you're basically the result. So why would you be a software vendor when you can be the company doing the thing powered by software?

And at a minimum, that has an impact on business models. And we saw that in our portfolio with Ada, which is one of the leaders in AI chatbots for customer service, where not that long ago they would sell on a per-seat basis and occasionally on a per-conversation basis, sort of selling the effort and the access. And today their business model, the way they price, has evolved to selling on a per-resolution basis. So you sell the work because you sell the result, you sell the resolved conversation as a unit of value and the way you price.

So super interesting. And look, it sort of changes everything in the enterprise and how we work and all sorts of fascinating applications, which could be a conversation for another day. But like we were talking about, those agents that for now are sort of like single-purpose kind of agents. You have your AI SDR, your AI finance agent, or whatever. And then the battle that's already happening, there's all sorts of companies starting to do that and announcements in the press, is, okay, what happens when you start connecting all those agents, like agent networks? And so far, from my perspective, it's the Wild West, and it's sort of interesting, but it's very scary in that, as we discussed a couple of times already, the AI doesn't get it right 100% of the time.

So it's fine for a single purpose, maybe depending on your use case. But when you start adding and combining multiple agents, then the error compounds, which is fine in certain use cases and terrible in other use cases. I think maybe last time we were talking about that, where you don't want an agent to sort of book the wrong flight or book you a meeting at the wrong time kind of thing.

And all of this leads to a profound transformation of the enterprise. Once you have a bunch of agents, what does that mean in terms of how a company is structured? So there's the whole dream of—I don't know if that's a dream or a nightmare—but this whole idea of a company that could generate several hundred million in revenue but have a handful of people. So that was a Sam Altman quote from probably a year ago at this point. But we're starting to see variations of that, including in our own portfolio, where people are actively using a bit of AI to replace certain jobs, or they decide to at least experiment with an AI instead of hiring the next person.

So we're starting to see the reality of this happening. But all those questions about the future, what does that mean when that happens? What will the reality of work be like? Will we be managing people? Will we be managing a couple of people and then hundreds of AI? And what does that mean in terms of our jobs? What does that mean in terms of our kids and how they need to be trained? It's just so fascinating as you keep unpacking.

Yes. All right, so like I said, let's close here. There's so much going on that we could be talking about this for even longer, but hopefully that was somewhat interesting for anybody listening to this that would want a check-in and an overview of how people like us think about the AI space at the end of 2024. But we're excited. There's plenty to build as founders. There's plenty to invest in as investors. And there's going to be ups and downs and all the things, but it always happens that way.

But yeah, the future is very bright, I would say.

All right, thanks, Aman. And we'll do this again soon. In a few months, we'll check back in. Thanks for doing this.

Hi, it's Matt Turck again. Thanks for listening to this episode of The MAD Podcast. If you enjoyed it, we'd be very grateful if you would consider subscribing if you haven't already, or leaving a positive review or comment on whichever platform you're watching this or listening to this episode from. This really helps us build the podcast and get great guests. Thanks, and see you at the next episode.