The Exchanges

Q in some other cases, There is perhaps a more correct solution to a problem, and there's a little less what you might call fault tolerance in the system. And so as we think about maybe workflow applications or enterprise applications, are there any that come to mind with similar potential, maybe a little more tolerance in terms of getting wrong answers or there not being maybe a specifically correct solution?

A Yeah, first of all, I love how you put that, uh, fault tolerant. Um, You know, I think that's notably different from other areas where the bar is incredibly high. Um, autonomous vehicles would be a good example of that. Uh, the stakes are, you know, the highest, right? Um, people die if the product isn't right. Um, and, and as a result, the process of getting there may take a bit longer. On the other hand, I believe that the potential workflow applications of AI are enormous and incredibly near term. And a lot of that is due to this, again, more tolerance for, for faults. And right. I think a lot of these workflows are actually oriented around AI assist. So in other words, where AI is making the human more productive, but humans still stay in the loop. And as a result, they can be deployed immediately. Um, I think coding has been pretty salient recently. I've even heard developers say that they can't work without it. You know, that's not to say that AI is, you know, developers are saying, Hey, you know, create this app soup to nuts, right? That, that may be in the future in the very near future. But, um, today what developers are doing are, um, using it to code complete, right. And, and figuring out how they can make their process, especially the, the repetitive parts of their job. Just way more efficient and faster. And, um, you know, we always talk about finding new solutions…

AI assessment note: “I think coding has been pretty salient recently... Video editing with companies like Runway”

Q Um, someone tweeted that venture capital is just a wrapper around LP capital. Yeah, to speak to the point. How should we think about AI native companies relative to traditional SaaS companies?

A Yeah, um, this is an interesting one, and actually I think I'll go back to the, the charts, uh, just cause they illustrate the point better than words could, um, but the first thing I'd call out that seems, that sort of just jumps off the page is that the AI native companies are far outpacing their SaaS counterparts, um, and you can see it in terms of new companies blowing past this golden metric of time to a hundred million of ARR, um, which is truly an incredible feat to accomplish, um, On the, on the left, but what's amazing is that it's not just the best companies that are doing really well, um, it's actually on average, you can see this from the aggregated Stripe data, um, where AI native companies are growing faster and, um, than, than sort of the, the SAS generate, SAS two point O generation, if you call it, um, and we'd sort of attribute this to a number of factors. Um, one is just looking at the compelling ROI out of the box, um, and With AI improvements and capabilities, you're seeing this 10 X plus improvement in the customer experience as well. Um, whereas SAS two point O, you, you generally saw a little bit more of an incremental improvement, you know, call it 25, 50%. Um, and then the second piece is it's early days for this, but, um, and it's related, uh, but you're starting to see replacement of some of the services budgets, uh, versus just software. And, uh, th…

AI assessment note: “AI native companies are far outpacing their SaaS counterparts”

Q Um, in closing, what are the, what are the key message, key messages we want to leave this audience with? Sarah, take it away. Great.

A Um, so we kept this slide simple, uh, and you, you've already heard these messages over and over again, um, but they really are the key ones that we want to impart from, from our work and, uh, the way that we think about investing, um, and that is the market is growing faster, and it's larger than anyone anticipated, most of all us. You've got to be really thoughtful about where you're placing the bets. The stakes are higher than ever. And this is a market where heat cannot be confused with momentum. Um, and then finally, as Martine said, you're on the field or you're irrelevant. We're incredibly bullish on the opportunity ahead, and we think this is just the beginning.

AI assessment note: “they really are the key ones that we want to impart”

Q It's funny because there was all this talk, um, a year ago about GPT wrappers. Every app was a GPT wrapper, you know, foundation models gonna win everything. What changed? And, and why aren't the foundation models winning everything?

A Yeah, um, I, I think this is a fun one for a lot of the investors in the room because, ah, we don't all invest in just foundation models. Um, and the fact that we're seeing this gangbusters growth in AI apps, um, is, is, is really exciting to see. Um, so I think two parts to your question. Um, one, you know, it makes sense that a lot of the focus and investment early on in this cycle was on the infrastructure side. Um, and now what we're seeing is Apps are benefiting from that massive investment as intelligence effectively has become free, and you have this aligning of the stars where the fierce competition on the state-of-the-art model market is driving both non-stop continued capability improvement paired with continuous price decreases. Um, and in fact, I think model inference costs have gone down 10 X year over year. Um, so the stars, you know, as, as we say, are aligning on that front. Um, And then on your question of why don't the foundation models just take over everything, I think this is a question, I mean, it's a very reasonable question. It's one that we sort of ask ourselves for every new investment that we make, and we certainly saw this with the early marketing copy AI apps. Um, but, you know, I think in ask, in talking to the founders of these app companies themselves, and then also the model providers, the answer to this question has become increasingly consiste…

AI assessment note: “where You have complex workflows and a ton of customer data where deep integrations actually are necessary”

Q It's funny because there was all this talk, um, a year ago about GPT wrappers. Every app was a GPT wrapper, you know, foundation models gonna win everything. What changed? And, and why aren't the foundation models winning everything?

A Yeah, um, I, I think this is a fun one for a lot of the investors in the room because, ah, we don't all invest in just foundation models. Um, and the fact that we're seeing this gangbusters growth in AI apps, um, is, is, is really exciting to see. Um, so I think two parts to your question. Um, one, you know, it makes sense that a lot of the focus and investment early on in this cycle was on the infrastructure side. Um, and now what we're seeing is Apps are benefiting from that massive investment as intelligence effectively has become free, and you have this aligning of the stars where the fierce competition on the state-of-the-art model market is driving both non-stop continued capability improvement paired with continuous price decreases. Um, and in fact, I think model inference costs have gone down 10 X year over year. Um, so the stars, you know, as, as we say, are aligning on that front. Um, And then on your question of why don't the foundation models just take over everything, I think this is a question, I mean, it's a very reasonable question. It's one that we sort of ask ourselves for every new investment that we make, and we certainly saw this with the early marketing copy AI apps. Um, but, you know, I think in ask, in talking to the founders of these app companies themselves, and then also the model providers, the answer to this question has become increasingly consiste…

AI assessment note: “where You have complex workflows and a ton of customer data where deep integrations”

Q Um, someone tweeted that venture capital is just a wrapper around LP capital. Yeah, to speak to the point. How should we think about AI native companies relative to traditional SaaS companies?

A Yeah, um, this is an interesting one, and actually I think I'll go back to the, the charts, uh, just cause they illustrate the point better than words could, um, but the first thing I'd call out that seems, that sort of just jumps off the page is that the AI native companies are far outpacing their SaaS counterparts, um, and you can see it in terms of new companies blowing past this golden metric of time to a hundred million of ARR, um, which is truly an incredible feat to accomplish, um, On the, on the left, but what's amazing is that it's not just the best companies that are doing really well, um, it's actually on average, you can see this from the aggregated Stripe data, um, where AI native companies are growing faster and, um, than, than sort of the, the SAS generate, SAS two point O generation, if you call it, um, and we'd sort of attribute this to a number of factors. Um, one is just looking at the compelling ROI out of the box, um, and With AI improvements and capabilities, you're seeing this 10 X plus improvement in the customer experience as well. Um, whereas SAS two point O, you, you generally saw a little bit more of an incremental improvement, you know, call it 25, 50%. Um, and then the second piece is it's early days for this, but, um, and it's related, uh, but you're starting to see replacement of some of the services budgets, uh, versus just software. And, uh, th…

AI assessment note: “AI native companies are far outpacing their SaaS counterparts”

Q layers changing, if at all? Because right now we are relying on a few models like GPT-III, which influence Copilot or some of these other layers. Do you see that changing where we then have many, many more kind of Fine-tuned models to the use cases that people need, or do you see it all kind of layering back to the few foundational models that we have today like GPT?

A Yeah, this is a really fascinating question and one that we've been discussing internally a lot, especially as to your point, a lot of the focus more recently is on these horizontal large language models and the sheer capabilities of these models has, has really been astounding. But to your other point on fine tuning or, you know, the process of using additional data to further train a pre-trained language model has a really big impact on how accurate and useful these models are. Especially if you think about B to B use cases where a lot of the truly useful data is actually proprietary enterprise data and not available widely on the internet to train on. So if you put this together, what does it actually look like in the future? Um, Honestly, the, the real answer is likely that it's use case dependent. Um, you know, I think one of the exciting aspects of the LLMs is that they really democratized access to large and expensive models via their API. You don't need tens or even hundreds of millions of dollars to deploy them in your product. And if companies are then further able to fine tune these models to their particular industry, there's an argument that specific use cases Don't need expensive models. And we're seeing some companies start to make the decision to own the model, which affords them more control, uh, but also maybe more cost effective in the long run. And so it's, …

AI assessment note: “Honestly, the, the real answer is likely that it's use case dependent.”

Q solidifying their existing features. But to that end, it feels like almost like every white collar professional is going to be reshaped in some way by AI to differing degrees. But I'm curious to hear from you in 20, 23 specifically, because that's what we're talking about with these big ideas. Are there any kind of short term, near term companies, products, industries that you think opportunity lies the most?

A Yeah, I think there's a lot of talk or attention on some of the things that we mentioned, right? Copy editing, coding, um, a lot of the consumer use cases. Um, so maybe I'll throw out one that's a little bit less than news, but that I'm really excited for. And I think it will be a 20, 23, uh, change. Um, and that's just the potential for AI and the data workflow. So I'm no data scientist, but as an Excel junkie, you know, I know there's a lot of repetitive workflow involved in getting to the insight or the answer that. AI can help you wade through. Um, and from the demos that we've seen that have, frankly, I think the 10 X potential that you see in a lot of these other arenas can be applied to data workflows. And that's incredibly exciting because data-driven decision-making has become the norm. It's, uh, the backbone of our data infrastructure investing thesis. But for AI to tackle that as an arena next, I think is incredibly exciting and a very, very large end market. And again, you know, I think a lot of the companies that own the workflows already will come out with some of these breakthroughs and AI features. They really do change the use case from novelty to something that's actually useful, that will save you hours of your day, that will take you to an outcome that is greater than you would have gotten to had you not used the AI. And I think that's what That's what I'm m…

AI assessment note: “and that's just the potential for AI and the data workflow.”

Q in some other cases, There is perhaps a more correct solution to a problem, and there's a little less what you might call fault tolerance in the system. And so as we think about maybe workflow applications or enterprise applications, are there any that come to mind with similar potential, maybe a little more tolerance in terms of getting wrong answers or there not being maybe a specifically correct solution?

A Yeah, first of all, I love how you put that, uh, fault tolerant. Um, You know, I think that's notably different from other areas where the bar is incredibly high. Um, autonomous vehicles would be a good example of that. Uh, the stakes are, you know, the highest, right? Um, people die if the product isn't right. Um, and, and as a result, the process of getting there may take a bit longer. On the other hand, I believe that the potential workflow applications of AI are enormous and incredibly near term. And a lot of that is due to this, again, more tolerance for, for faults. And right. I think a lot of these workflows are actually oriented around AI assist. So in other words, where AI is making the human more productive, but humans still stay in the loop. And as a result, they can be deployed immediately. Um, I think coding has been pretty salient recently. I've even heard developers say that they can't work without it. You know, that's not to say that AI is, you know, developers are saying, Hey, you know, create this app soup to nuts, right? That, that may be in the future in the very near future. But, um, today what developers are doing are, um, using it to code complete, right. And, and figuring out how they can make their process, especially the, the repetitive parts of their job. Just way more efficient and faster. And, um, you know, we always talk about finding new solutions…

AI assessment note: “Video editing with companies like Runway really come to mind.”

Q layers changing, if at all? Because right now we are relying on a few models like GPT-III, which influence Copilot or some of these other layers. Do you see that changing where we then have many, many more kind of Fine-tuned models to the use cases that people need, or do you see it all kind of layering back to the few foundational models that we have today like GPT?

A Yeah, this is a really fascinating question and one that we've been discussing internally a lot, especially as to your point, a lot of the focus more recently is on these horizontal large language models and the sheer capabilities of these models has, has really been astounding. But to your other point on fine tuning or, you know, the process of using additional data to further train a pre-trained language model has a really big impact on how accurate and useful these models are. Especially if you think about B to B use cases where a lot of the truly useful data is actually proprietary enterprise data and not available widely on the internet to train on. So if you put this together, what does it actually look like in the future? Um, Honestly, the, the real answer is likely that it's use case dependent. Um, you know, I think one of the exciting aspects of the LLMs is that they really democratized access to large and expensive models via their API. You don't need tens or even hundreds of millions of dollars to deploy them in your product. And if companies are then further able to fine tune these models to their particular industry, there's an argument that specific use cases Don't need expensive models. And we're seeing some companies start to make the decision to own the model, which affords them more control, uh, but also maybe more cost effective in the long run. And so it's, …

AI assessment note: “Honestly, the, the real answer is likely that it's use case dependent.”

Q solidifying their existing features. But to that end, it feels like almost like every white collar professional is going to be reshaped in some way by AI to differing degrees. But I'm curious to hear from you in 20, 23 specifically, because that's what we're talking about with these big ideas. Are there any kind of short term, near term companies, products, industries that you think opportunity lies the most?

A Yeah, I think there's a lot of talk or attention on some of the things that we mentioned, right? Copy editing, coding, um, a lot of the consumer use cases. Um, so maybe I'll throw out one that's a little bit less than news, but that I'm really excited for. And I think it will be a 20, 23, uh, change. Um, and that's just the potential for AI and the data workflow. So I'm no data scientist, but as an Excel junkie, you know, I know there's a lot of repetitive workflow involved in getting to the insight or the answer that. AI can help you wade through. Um, and from the demos that we've seen that have, frankly, I think the 10 X potential that you see in a lot of these other arenas can be applied to data workflows. And that's incredibly exciting because data-driven decision-making has become the norm. It's, uh, the backbone of our data infrastructure investing thesis. But for AI to tackle that as an arena next, I think is incredibly exciting and a very, very large end market. And again, you know, I think a lot of the companies that own the workflows already will come out with some of these breakthroughs and AI features. They really do change the use case from novelty to something that's actually useful, that will save you hours of your day, that will take you to an outcome that is greater than you would have gotten to had you not used the AI. And I think that's what That's what I'm m…

AI assessment note: “and that's just the potential for AI and the data workflow.”

Q Hey everybody, great to see you all. Martine, Sarah, we just went through the state of the firm. What's the state of play right now in AI?

A The last two and a half years have really, really felt like a blur. Um, maybe just to set the table, given that the AI landscape is changing so quickly, Martina and I thought it would be valuable for our internal team, actually, to reflect and take stock of where AI, or where value, I should say, is accruing in the AI ecosystem. Um, so the slides that you're about to see here are, um, actually a product of our internal presentation to the GPs, where even we were surprised by some of the findings and wanted to share those today. Um, So we're gonna go through a lot of data and slides. I think it really distills into a couple of key takeaways, and that's one, AI companies are growing faster and are larger than even we expected. There's value accruing across every layer of the stack, models, infra, apps. Um, all that being said, there's this paradox that we're seeing where more value creation is occurring in a shorter amount of time, paired with more wipeout potential happening over a shorter period of time. We'll definitely dive into that dynamic more. Um, and then finally, our conclusion is that you've got to be on the field, but you have to be smarter about where you're taking those bets than ever before because the stakes are higher.

AI assessment note: “There's value accruing across every layer of the stack, models, infra, apps.”

Q Yeah. Let's, uh, let's put that first point you made more into perspective. What's the scale that we're talking about when we say foundation models are going faster than expected?

A Yeah, so this is a point that we thought was better answered actually, um, through two charts rather than just chatting through it. Uh, and you know, two years ago, I would have actually said we were probably the firm, maybe the most bullish on where the market could go. Um, and I've got to say, even we are surprised by how large and fast growing this market is. Um, and if you look at the revenue of just two of the top tier frontier labs, um, You know, if you look at the chart on the left, not only have they surpassed the early revenue ramps of some of the best SaaS companies in history, um, I'll, I'll point you over to the right, uh, they're actually starting to pass the early ramps of some of the hyperscalers, and I think these stats are even more breadth, or these charts, I should say, are even more breathtaking when you think about just the timing of when their products launched. What we find even more exciting in this space is that it's not the case that just two companies are growing very quickly in AI, and in fact, Um, that's probably a good segue to the next slide that shows markets are not only growing faster and much larger than expected, they're also fragmenting.

AI assessment note: “they surpassed the early revenue ramps of some of the best SaaS companies”

Q Yeah. Let's, uh, let's put that first point you made more into perspective. What's the scale that we're talking about when we say foundation models are going faster than expected?

A Yeah, so this is a point that we thought was better answered actually, um, through two charts rather than just chatting through it. Uh, and you know, two years ago, I would have actually said we were probably the firm, maybe the most bullish on where the market could go. Um, and I've got to say, even we are surprised by how large and fast growing this market is. Um, and if you look at the revenue of just two of the top tier frontier labs, um, You know, if you look at the chart on the left, not only have they surpassed the early revenue ramps of some of the best SaaS companies in history, um, I'll, I'll point you over to the right, uh, they're actually starting to pass the early ramps of some of the hyperscalers, and I think these stats are even more breadth, or these charts, I should say, are even more breathtaking when you think about just the timing of when their products launched. What we find even more exciting in this space is that it's not the case that just two companies are growing very quickly in AI, and in fact, Um, that's probably a good segue to the next slide that shows markets are not only growing faster and much larger than expected, they're also fragmenting.

AI assessment note: “they're actually starting to pass the early ramps of some of the hyperscalers”

Q So we've, um, we've talked about the big winners, but there are also some big wipeouts, as we mentioned. What have we learned about the commonalities between ones that win and ones that don't?

A I think what's interesting about this funding, funding cycle in particular, um, and I, I know, uh, history repeats itself sometimes, but, um, in this case, especially on the foundation model layer side, I think what's remarkable is just the massive size of the rounds that folks are raising before any traction, um, and, you know, we're participating in some of these rounds, uh, we'll, we'll talk more about what our thesis is when, when we do, um, but as we all know, the more money you raise early on, the more pressure you have Um, to really show performance. Um, DG likes to call this transition, ah, going from a, you know, a tell the story company to a show not tell company. Um, so you really need to show not tell when you've raised hundreds of millions of dollars.

AI assessment note: “the more money you raise early on, the more pressure you have”

Q So we've, um, we've talked about the big winners, but there are also some big wipeouts, as we mentioned. What have we learned about the commonalities between ones that win and ones that don't?

A I think what's interesting about this funding, funding cycle in particular, um, and I, I know, uh, history repeats itself sometimes, but, um, in this case, especially on the foundation model layer side, I think what's remarkable is just the massive size of the rounds that folks are raising before any traction, um, and, you know, we're participating in some of these rounds, uh, we'll, we'll talk more about what our thesis is when, when we do, um, but as we all know, the more money you raise early on, the more pressure you have Um, to really show performance. Um, DG likes to call this transition, ah, going from a, you know, a tell the story company to a show not tell company. Um, so you really need to show not tell when you've raised hundreds of millions of dollars.

AI assessment note: “the more money you raise early on, the more pressure you have”