Recorded live at Seattle Tech Week, this episode of Founded & Funded brings together two people who track the AI capital markets from different vantage points: PitchBook EVP of Research & Market Intelligence Nizar Tarhuni, and Madrona Partner Sabrina Albert.
These two dig into where AI venture capital is actually landing across the model, infrastructure, and application layers, why AI “harnesses” are becoming a winner-takes-most category, and where vertical AI applications may hold an edge that horizontal platforms can’t match. They also cover PitchBook’s VC Exit Predictor, the new normal for AI seed valuations, and the tranche-financing structures Nizar says are red flags for founders.
And in a live Q&A, they take on a frequently asked question in the AI market: Is this a bubble? Nizar explains why he thinks it’s still too early to call, what the data says about the value already being created, and why the next few stages of company growth will tell us much more about which valuations prove durable.
For founders raising capital right now, or operators trying to figure out where durable value is emerging in AI, this conversation offers a data-grounded look at what’s happening beneath the headline numbers.
Listen on Spotify, Apple, and Amazon | Watch on YouTube.
This transcript was automatically generated and edited for clarity.
Sabrina: I think the first topic I’d love to explore with you is really understanding where the capital has gone. The market is moving really quickly. The way I like to think about it is there’s three main stacks of how you can think about it broadly. There’s the model and infrastructure layer, there’s a middleware and tooling layer, and then there’s the application layer. And we’ve seen over the last year a lot of capital flow into each of these three different segments. There’s been a ton of funding, OpenAI, Anthropic, et cetera. But I’m also hearing a lot of talk about investments also at the infrastructure energy layer, but maybe we could start by just really understanding what’s happening at the model and infrastructure layer. How have you seen some of the capital shift over the last 12 months or even six months, and what does the data support?
Nizar: If you look writ large across all venture fundraising, you’ve got, call it 80 to 90% of all venture dollars, unsurprisingly, flowing into some sort of AI segment right now. When you then break that down in the layers of the stack, I think the evolution you’ve seen has been immediate investments in the model layers. And so you’ve seen that grown, call it 18 times over a year’s period, where out of 230 or so billion dollars invested across AI, literally it’s about 10 companies represent somewhere around $150 billion of that. So call it two-thirds, if you will, were in model companies, but only around 10 companies. And out of those 10 companies, that’s actually down about 40%. So there is obviously more competition at the model layer than I think we’ve seen that weed out as all of us pay attention to who the main two or three players are and some of the other folks that are on the fringes. So literally 10 companies are making up about two-thirds of all AI investment.
When you then break it down to the infrastructure layer and then the application layer, at the application layer, you’ve got a lot more deals, so close to 1,000 transactions just in Q1of 2026 alone. And they represent about 10% of capital, and then the remaining 20% is sitting in the infrastructure layer. And so I think the shift you’re seeing now is that with any technological shift, many of you have invested or built companies in different technological platform shifts, they move faster in the next cycle than the cycle before, and that trend tends to continue. And I think what you’re seeing now is the evolution of how fast we went from model into infrastructure into the application layer is faster than what we’ve seen before. And I think it’s tough to compete at the model layer, which is why I think you’re seeing dollars concentrate because you have such good proliferation happening with open-weight models.
I think folks also have to be a lot more cost-conscious around how they’re deploying some of the frontier models. I think you’re starting to see the counts grow in the application layer, but the dollars are still concentrated at the model layer. But the opportunity in terms of volume of companies coming to market, it seems that we’re starting to see that show up at the application layer.
Sabrina: How about in the tooling layer? Have you seen anything emerge there? One of the areas that we have seen a lot of activity in, or we’ve been writing a lot about as well, is this idea of the harness, which goes to some of the things you were talking about: understanding how to manage costs, understanding memory context, all of these things to build the application on top of it. Have you started to see some interesting companies emerge in there, or any data that supports that?
Nizar: So, I think in that infrastructure layer, and especially when you think about harnesses, I think some of the most important things have been, can you build the harness that actually takes advantage of the context and memory, understands the workflow, understands the context of any bespoke industry or company, a little bit more vertically oriented? I think what you’re seeing there is more of a winner-takes-most market, and where you might have three or four harnesses that arise in a specific vertical, but the reality is it becomes, then, at that point, a race to gain market share. And so I don’t think you’re seeing the same level of proliferation as you’re seeing at the application layer, which I think is mainly expected, but I think you’re seeing a lot of durability being built with the harnesses to do it well.
And I think one of the things I think about is if you’re hiring a good product manager or anywhere where you’re trying to bring talent on board, one of the things I think a lot about is if you were to ask that same person, what is something you’re interested in outside of work? And you’ll get their hobby, and there’s a level of detail and passion that comes through and all the little details for whatever that hobby is. And I think about a harness sometimes is very similar, which is the level of context you need, it’s all those little details. It’s all the edge cases that you need about whatever vertical you’re owning, whatever context and memory and workflow needs to be incorporated in it.
It’s actually really hard to do. And so for you and I, it might look like this wide surface area that harness is able to unlock, but the reality is, underneath the hood, so much work has been done to understand every use case and edge case and every nook and cranny of what that harness needs to enable. And I actually think that’s really hard to do at scale, which is why you don’t see the same level of proliferation as you get at the application layer. But at the same time, I think you see a lot more durability for the companies that can actually do that well.
Sabrina: Yeah, it’s interesting to see the dynamic shift from the model layer to, hey, do you have the right harness to make sure that you understand how to route to the best models, to think about memory, to think about the context? Because if you’re building an application, to your point, you need all of these things. You need all of these ingredients to build a really successful and durable application company.
From what you’ve seen in the data as well, has there been a white space as it relates to opportunities within the application layer? There’s obviously, to your point, been a lot of funding dollars that have flowed there. You can think about it in the vertical context or the horizontal context. What are you seeing?
Nizar: From the counts of deals perspective, you’re seeing that move now towards the vertical AI component. From a horizontal perspective, it accounts for more dollars; you’ve got to accomplish a lot more. It’s more context that you’re managing. But I think the white space lives in the vertical components. I think what folks are trying to figure out is, can you be very specific to a specific domain? Can you be vertically oriented? And I think that is where you’re seeing. I don’t know if it’s “white space” yet, but it’s where you’re starting to see a lot of activity. And I think the cost to build some of the applications in a vertically oriented space is a lot easier than building something in a horizontal space where you have to manage a lot of different contexts and a lot of different stakeholders and understand a lot of different client types.
And so I think TBD on if it’s white space that translates into recurring revenue and durable market share that grows, but it’s certainly where you’re seeing the most activity is moving into the vertically oriented application space.
Sabrina: I definitely observed that as well. I think some of the verticals we’ve seen a lot of activity in is legal. I think the models are really great at understanding a large corpus of information, specifically as it relates to text. We’ve seen a lot of activity in finance as well historically. I know PitchBook has a point of view there. And then healthcare is another vertical. I’ve seen a lot of pickup in activity on both the clinical AI side as well as the more revenue cycle management side. But I imagine there’ll be other verticals too that emerge in insurance, maybe one. I’m not sure if there’s any in particular that you’ve seen more pick up in activity over the last few months?
Nizar: I think some of the ones you just mentioned are where you’re seeing the most. Finance certainly has a tremendous amount of applications that are being built. And I think within finance, you also find your sub-verticals, everything from personal finance into corporate M&A transaction activity into running VDRs and deal rooms. I think across every one of these verticals, you’re also seeing the sub-vertical niche create. And I think the pitch you see is somebody starts in one component and then thinks that they can apply it to the next. And I think some of that will ring true; some of it won’t. And I think some of the spaces where you have a lot more regulatory activity, so inside of insurance or in healthcare, I think you’re going to see more and more applications that live inside very sub-niches of that segment, where transitioning from say document and record keeping an Epic type competitor into actual clinical trial management, those things don’t actually live the same and they have very different contexts. And I think you’re going to start to see more point solutions show up for those specific sub-segments.
Sabrina: The other thing that I like to think about is, do you think these vertical AI companies are more defensible vis-a-vis OpenAI or an Anthropic? One of the big questions that I get a lot as an investor is how do you build a durable company in today’s day and age when Anthropic and OpenAI have raised all of this capital? Will they eat your lunch over time? From your perspective, do you think that vertical focus has certain advantages over maybe some of the horizontal ones you were alluding to it a little bit earlier?
Nizar: I think the evolution’s interesting right now where, for example, it depends on who you’re selling into and then the context that your end customer is managing. So I think a good example is to think about an applicant tracking system. When you’re Google, when you’re Microsoft, you’re constantly building your own ATS. You’re going to run your own ATS. You have very specific conditions. You have hypersensitive data and workflows that you want to manage this out of your own context. You’re going to build your own.
When you’re PitchBook with 3,500 people, we use Greenhouse, or we use other ATSs. And so I think what you’ll find is that for some companies, they will likely build their own harnesses. They might use external tools to help build it, whether it’s different Cloudflare objects or whatever it is, but they’ll build their own and they’ll build their own application layer on top. It’ll be bespoke to their own context, and they’ll care a lot about that, and they can do it on their own. For other companies, if you’re selling into the middle market, for example, those companies will be a lot more open to using a purpose-built, vertically oriented AI solution that accomplishes the bulk of their needs. And I think from that perspective, the vertical orientation creates opportunity, and there’s a big market for that.
So I think part of it just comes down to the human nature of who you’re selling to. I think there’s a tremendous amount of defensibility when you’re building something into the middle market that can accomplish a very specific workflow. I think it’s probably going to be a bit harder when you’re attacking the higher end of enterprises, where I think the wherewithal, the capability, and the sensitivity of data is a lot higher, and it becomes harder, I think, to build for that end market.
Sabrina: Yeah, it’s interesting because we also do this list every year. It’s called the Intelligent Applications 40 list (IA40), which we partner with PitchBook on. 2026 will be our sixth year doing the list. One of the interesting stats from that list is that only one company, Databricks, has been on the list each of the 5 years we’ve released so far. And on the 2025 list, there was 50% turnover in the companies. So this question of what’s in versus out and over six months to a year continues to be top of mind. One of the things that we use to determine the list is actually the PitchBook Exit Predictor score, which takes into account a bunch of different data points, understanding how the company is doing over time. Maybe share a little bit more about what that is and how you guys think about what the key data points are that you have to capture to understand what creates long-term defensibility? What are the things that you guys are thinking about to predict that exit score?
Nizar: The VC Exit Predictor score is a set of IP that we created, which effectively takes a commingling of data sets that allows us to effectively forecast whether we think a company will have a successful exit. And then we’ve got some other modifications coming that allow us to try to time when we think that exit will come. And so part of it comes from how much capital you’ve raised, the pace at which you’re raising that capital, but we also track other things in terms of employee growth, your hiring; we track your valuation spreads and your step-ups. It’s a number of financial metrics, and then we overlay that with some macro conditions as well to give us a sense of how a company is performing relative to their comp set.
And I think when you think about the IA40, when you look at Databricks, which has stayed on that list, one of the things that we’ve seen drives companies to stay on a list where I think half of them maybe rotate out every year is this concept of durability. And the concept of durability I think comes from staying power and going concern. One of the things that makes a company more durable, especially in today’s day and era, is how high the switching costs are for somebody to move off of that. And I think one of the things that’s really easy to miss in the AI era is you can get speed to market and you can get revenue really quick and fairly faster than you could in the past. And I think you’re also in an environment where everybody wants to try different things.
I think the hard thing to do is to really be honest with yourself about if somebody’s using this, is it easy or is it hard for them to switch off and use something that’s comparable? And so there’s a ton of products that you see. I can think about PitchBook when we look at different products where something looks incredible and it is an incredible product. But if you talk to the engineering teams underneath, they’re like, This is how they build the stack, this is how the architecture looks. Anyone can go out and do that. It’s really cool. It’s interesting, but it’s actually a lot of folks will probably build that way or that’s the new way of building, the new architecture is being built.
The thing we think a lot about with the Exit Predictor right now is how do you think about durability, and how do you think about switching costs? And are there real switching costs that make it very hard for somebody to switch off that product? And I think that gets masked quite a bit in the AI era today, where things are easy to use, and you can use them quickly, but just because they’re easy to use and you can use them quickly doesn’t necessarily give them the same staying power you might expect.
Sabrina: It’s easier today more than ever to try different products and to get to a level of productivity that people had historically never seen before. And I think one example of this that captures it best is that Anthropic announced that they added $11 billion of run-rate revenue back in April, which is incredible. No software public company had done that in the history of software public companies, which is really interesting to see. And six months ago, the discussion was all about frontier models, Anthropic, OpenAI, et cetera. And now, fast-forward to today, a lot of the discussion is around open source. And to your point about switching costs, how hard is it to actually swap out some of these models? And a lot of the companies that I work with are talking about, “Hey, we want to actually use these open-source models so we can control cost, so we can manage where data goes and not give that all to the model providers.”
Have you started to see movement in the open-source activity or inference layer? What have you seen maybe in the data that supports that some of these companies are starting to accelerate more so than perhaps the frontier models?
Nizar: So we don’t track specifically maybe where different models are being used and the pace of movement there. But I think what we could tell you is when our analysts or our reporters are covering the space, I think it’s likely no different than what you’re seeing, which is folks are looking to experiment a lot more with the open-weight. I also think that, again, it’s human nature. At some point you feel like you use the tools that you have. And so, if OpenAI and Anthropic are the only show in town, or the ones that you’ve been told are the ones that you can trust to use, over time, especially if I think about PitchBook as a net buyer of AI products, at a certain point we don’t have an infinite budget. We have an OpEx number we have to manage to. And you start to get a lot smarter about which models I need to run for which tasks.
At some point I might use 4.8 for certain tasks, even if it’s “More expensive” by tokens, because it’ll get the job done faster, and that’s better for me than running something on a cheaper model that’s going to rerun it three or four or five times. At the same time, I might feel fully comfortable with an open-weigwhich tasks. model if I can adjust the weights myself for a specific task that sits in my specific context. And I think you’re going to see more and more companies that are taking that approach, which will put some level of pressure on the frontier models or, in some way, shape or form, temper some of the growth trajectories. But I think some of that just comes down to the nature of the margin expectations that every company has, which I think will be slower to be realized, but the OpEx realities, especially for public companies, are real, and you can’t escape them.
And your investors ultimately hold you to a bar of, “you need to drive your top line with a level of margin that we expect relative to your comp set.” I think that’s forcing a lot of different behavior for folks to want to be a lot more sophisticated. Where I think it creates a certain interesting opportunity is at the inference level or at the measurement level or at the cost management level, we use a middle layer to help us figure out what should our cost productivity be. And so when our engineers or folks are accessing models, they’re accessing them through a middle layer, we know exactly what they’re spending. We can allocate tokens to them. We know when we can shut it off. And so I think that’s a different vendor that we’re spending money on in order for us to get better cost efficiency out of how we use AI.
There’s things like that from that measurement layer and the token management layer. I think you’ll see interesting infrastructure show up there. But I think that’s a symptom of what you were bringing up at the beginning of the question, which folks are starting to think about using a lot of other components. And I think almost reverting all the way back away from AI. And so some of the things that we do is you try to figure out what can we do deterministically? If we know exactly we need X interaction to lead to this outcome, and instead of running it through a set of models and having them run loops on their own, if we can program it with Python scripts and then keep data sets sitting in Airtable and then flow them into the outputs that we need, we’re likely going to do that. And we can get work that took us weeks down to seconds rather than spending the money to run through different models.
So I think the automation component is a lot more top of mind for folks, whether it’s with AI or not, I think, in some ways, because of the cost implications of the token costs.
Sabrina: I think this question of ROI on AI is so real today. Before maybe six months ago, there was this idea of token maxing, and everyone was trying to spend as many tokens as they possibly could to show that they could use these AI models. But I think what we’re starting to see is that CFOs care deeply about how much companies are spending. And then there’s also this whole, to your point, the component of how do you bring in the right models for the right tasks? And sometimes you should use a deterministic AI model. And sometimes you should actually think about maybe routing to a cheaper quad model, for example, based on what it is that I’m actually trying to do. And I do think that some of the companies that are going to succeed, or are short-term succeeding right now at least, is the inference players and the Base 10s and the Fireworks of the world.
Have you seen specifically more capital flowing into some of those inference provider companies?
Nizar: I think we have, but I think it’s also, I don’t feel like that was relatively new. I think they’re the early beneficiaries of folks realizing that you were going to have an inference problem. So I don’t think we’ve seen anything different. I think we saw the same, the uplift that we saw in those spaces and they’ve maintained.
Sabrina: I want to shift a little bit into something that I think a lot of the audience are probably curious about. I think the macro story has been really interesting in terms of what’s happening. SpaceX’s IPO recently. We’re seeing a lot of stories in the market where companies are raising pretty big rounds and then some that are more modest. I’d love to just really understand what the funding environment looks like today from pre-seed all the way through Series A, Series B? What are you guys actually seeing in the numbers?
Nizar: If you went back 10, 12, 15 years, every year, it seems like you’re met with a new record of where funding rounds come in.
Sabrina: Yeah.
Nizar: I think it’s definitely bifurcated when you sit at the model layer, you’re just in a different game. You have different capital needs, you have different revenue bases, you have different capital availability, whether it’s in the private credit space, whether it’s through debt capital. You have more sophisticated, believe it or not, for such young companies like capital structures of how they’re raising capital to fund. And so I think those numbers tend to flag a lot of what you see in the media and the news. And the reality is I think if you look at seed stage AI companies today, I think the average seed level fundraising route is actually $5 million. I think you see the average seed-level pre-money valuation around 78 million, which is really high relative to what you would see even just two or three years ago. And I think the thing that’s interesting is if you compound that relative to non-AI companies, then you see a big delta there.
And so for the consumer products or the AI-enabled products, you obviously see a lot fewer dollars flowing there. And then I think that’s interesting is as you move that up the stack, I feel like historically you would see double, double, triple, triple when you would get good growth. I think what you’re seeing today is you’re going from seed; you’re almost 4Xing in your Series A if you’ve got some traction. And then from there you’re seeing things grow two and a half times, another two and a half times. And at that point, you’re in a different hemisphere where things can double from there. So I think we’re seeing numbers like we’ve never seen before. And I think we’re seeing the step-ups between the rounds for those who have success.
Sabrina: Yeah, it’s a really interesting dynamic in the market. I think it’s a pretty barbell approach where we still see some more traditional, for today’s day and age, seed rounds that are getting done, four to $5 million median, maybe 30 to $70 million post-money valuations. And then on the flip, you see these companies that are raising at a hundred million plus dollar first rounds that are getting done in a two-step tranche where it’s coming in at one price that’s a lower price, and then immediately after coming in at a billion dollar valuation. I’m sure you’re seeing that in a lot of the companies that PitchBook has data on all the companies. What’s your perspective on that in the market? Do you think it’s a point in time? Do you think this will continue on? Or have you seen that historically happen in the past?
Nizar: If I give you my personal perspective, so at PitchBook, just like any other company, we look at a lot of different companies from a strategic investment minority or a majority acquisition perspective. And I think when you think, as you can imagine for us, what we take a look at is primarily financial services, data businesses, AI harnesses, things like that that are focused on our industries. And I think that’s exactly what you’re seeing. I think you’re seeing rounds that have 10 or 11 different tranches in them and they get really convoluted and it’s hard to understand exactly who’s paying what and for how long. You see rounds come at you where somebody says, “We think you’d be a great partner. This is the round right now, but don’t worry, we’ve got a next round coming in. Here’s a valuation that we’re already raising at.”
And I think in general, my personal take when you see that from a founder is that you are playing a capital-raising game and you are ultimately less interested in the outcome that you’re driving for your customers. And especially when you see that at the Series A or the Series B level. To me, it feels personally like you’re playing that game way too early, and your brain’s not thinking about the right things. And so for us, we tend to, we walk away and those aren’t exactly the deals that we spend a lot of time on.
That being said, it’s fairly common right now. It’s a pitch. I think sometimes you can say maybe it’s a founder trying to take care of existing investors. Sometimes it’s a way to fill out a round and bridge some documents before you think the terms might get a little harder on the next round. But in general, I think sometimes that’s the signal of exuberance and when a market has maybe gotten a bit over its skis. And I also think from a founder’s perspective, and I’m curious to get your take here, I think you should be cautious about setting that precedent with certain investors, whether they’re financial or strategic. I think it certainly makes people think twice. And I think it certainly makes people question if you have this incredible product that’s growing and the numbers back it; I think it could turn a lot of people off. When it feels like you’re somewhere in your journey and the things you’re prioritizing are not necessarily the value that you’re driving for your customer.
Sabrina: Yeah, I think there are always edge cases where maybe it’s that very special founder that has that very special background or has done this before and has a very specific point of view and is able to maybe command that market price, or the market is willing to give them the valuations that they’re looking for, which we definitely do see. And we’ve seen a lot of these kinds of rounds come together. But I do think that it comes back to the fundamentals of making sure that you pick the right long-term partner, especially if you’re doing this at the seed stage, partnering with somebody that you think will be there for you in the long run so that through the ups and downs, they’re really actually going to have your back. And I think that’s incredibly important, especially in today’s day and age where the market can have some of these really interesting swings in terms of where rounds get priced. But it’ll be interesting to see where the market shakes out.
Nizar: I think it’s okay to know your worth.
Sabrina: Yeah.
Nizar: If you’re sitting in a comp set, you’re at the Series A, Series B level, and you’re growing, and you can back it with the fundamentals where your business is at, and you know relative to your comp set, you’re a better asset, I think it’s okay to command a premium relative to where the market’s at. I think the thing that gets a little interesting is when you start to see more games get played with the tranches, or you’re leaving tranches open, or you’re promising next rounds at a higher price point than where the last set of investors are. I don’t think it’s fair to the group of investors that are there. And so I think there’s a dichotomy between trying to maneuver the paper and the terms in a certain way versus just simply commanding a higher price because you feel like you’ve got a premium product. I think the latter is totally okay relative to how your business is performing.
Sabrina: All right, I want to pause here and open it up for questions.
Todd Bishop Question: Hi, thanks for the great conversation. Todd Bishop from GeekWire. We follow closely the PitchBook and NVCA numbers, and we’re always shocked given that we’re here in Seattle and we see all the activity, that there are markets like Philadelphia and Austin and New York even that are above Seattle in terms of the total amount raised and the number of deals. What’s your overall take on how much that matters anymore in this world, how much you think about geography, and especially since we’re here at Seattle Tech Week, what this room of founders and investors and others involved in this community should think about when it comes to the numbers as it relates to Seattle in particular?
Nizar: Do you have a take on that as a local VC?
Sabrina: I’m curious how PitchBook specifically tracks that information, especially as it relates to the fact that, as you mentioned, the world is becoming more distributed and there are a lot of engineers or maybe a co-founder of sorts that is based in Seattle or you can have a big engineering group here, but technically it could be classified as a Bay Area or maybe somewhere else opportunity.
And so I definitely think that there continues to be a lot of innovation here in Seattle. I think that the engineering and AI talent continue to be extremely strong. You see a lot of people coming out of some of these large tech companies. That’s always been the historical trend. But look at OpenAI and Anthropic investing in really large offices here in the region, which will produce a lot of great talent. And often what we see is some people will go down to the Bay Area, live there for a little bit, and decide that they want to come back to Seattle because of all the great things that it has to offer, both from a company-building perspective and access to talking to some of the largest enterprise companies in the world. All of those things are definitely supported here.
Nizar: I think one of the things that gets maybe misunderstood a little bit is the point you made, which is that just because the companies might not be founded here, there’s a tremendous amount of talent that sits here and a tremendous amount of these companies that are hiring talent in Seattle. So the company founding might be in San Francisco in a different place, but when you walk down South Lake Union or you jump into Bellevue and you take a look at whether you’re Snowflake or whether you’re Anthropic’s new office or whether you’re a Statsig, which is now OpenAI, there’s so much talent in some of the biggest unicorns that have grown over time that are actually working out of Seattle.
And so it might get masked by not seeing a pre-seed company founded here. But I think from an economic activity perspective, there’s a lot more here that doesn’t make its way into those numbers. And I also think another point is obviously you can get capital in a lot of different places. And I think the proximity between here and the Bay, I would say that corridor in general, you’re going to see a lot of activity fly both ways.
Live Question: Hi, thank you so much for your time today, I’m curious, outside of the tech circles, we start hearing people saying that we should brace for the bubble to bust. From a data standpoint, do you have something to say to support that the AI trend that’s happening is more cyclical and not a bubble, or the fact of saying that there is something that we need to brace for from a pure investment standpoint?
Nizar: I think it’s hard to call it a bubble or not a bubble. A lot of people have made a lot of money doing that. A lot of people have lost a lot of money doing that. I think it’s probably a little bit too early. I think it’s where value accrues. So the way I think about it is you’ve got this stepladder growth that’s, honestly, we talked about earlier, it’s 10 companies driving two-thirds of capital. And so the bulk of dollars has gone into the frontier labs. I think if you were to ask anyone around, are the frontier labs a bubble? I don’t think you’d hear the answer yes. I think it’s changed all of our lives, whether you are approaching it from an enterprise using the underlying technology to build your net new products today, whether you’re an individual using it for your own personal needs.
And so I think you’re seeing value accrue. Also, you’ve never seen revenue growth and ARR growth at that clip, and you’ve also never seen it done on a revenue-per-employee basis as we have with some of those companies. And so I think there’s some justification you can make for the value that’s accrued there. When you then break down some of the valuations we’re seeing, I do think it’s a lot more barbell-shaped. And I’d say for some of these companies that can get to 10, 20, $30 million in ARR in, call it, two to three years, that’s also relatively new. That’s not necessarily a normal phenomenon where five or six years ago, in most industries, you’re seeing a company get to that scale that quickly. And so from a percentage of total AI dollars, it’s actually not that much that’s flowing into those companies. And even the results of many of them are actually pretty impressive right now.
So I would say, for where we’re at in this exact moment, it doesn’t feel, I could be completely wrong, that you’re in full bubble territory, but I think it’s about where do you see durability grow? I think once you get to the Series B, C level, which is honestly not that unlike venture and other industries and historically, you’ll start to get a sense of who can compound their growth and grow into that valuation and who you’ll see kind of taper off. And we haven’t had time to really go through that cycle yet. I think it’s over half of AI deals that are even got done are still at the seed stage. So I think we need a little bit more time to see where things play out.
Sabrina: Yeah. I think we’re just at the start of this AI reasoning revolution, so to speak, and the capabilities and the power of the models can unlock so much more than historically software could. And so I think we’re going to start to see a lot of new things that were unlocked that previously software couldn’t actually do because these reasoning machines actually can think a lot like humans. It doesn’t mean that we’re going to be displacing all these jobs tomorrow, but it does mean that it’s increasing the surface area of opportunity across a wide variety of different industries. So I think that it’s very hard to call if you’re at the top. That would be impossible. We’re just at the start of this AI transition.
Live Question: I’m curious how Service as Software fits within your layered rubric and where you would consider a service that is built as software versus software as a service, and if you’re seeing that or if you think that that trend is something that is picking up or whether or not it’s mostly smoke and mirrors?
Nizar: I think it’s a big component of what you’re seeing, the capabilities of AI that can do for you. So there’s a lot of services industries. Many of us here worked at corporations, we pay a lot of money for a lot of different services that take work off our plate where you feel like it’s primarily knowledge work or reasoning work that should be automated, but it needs a specific context. Maybe it’s in a specific vertical or in a specific horizontal. And so we were talking a little bit before this about marketing. A lot of us use different agencies to do a lot of work for us. I would say we bought a small startup that is running its own marketing stack last year, and we’ve basically taken most services that you would go to a third party for and have tried to bring them in-house through a tool stack.
And it’s actually been very effective for us in a much more cost-effective way than it would’ve been going through a traditional agency. And then we can allocate our spend to the agency we use in a much more efficient way for things that we think we can’t get out of some sort of reasoning tool.
Sabrina: Yeah. I think about it also in the sense that you can actually charge for outcomes now, or the sense that before, when you were thinking about traditional software, you would charge for a seat or a unit of software. But now you can really fundamentally change it because before it was like you’re delivering a service? And that’s the idea of service as software, is that if I deliver this outcome for you, then you can actually pay me for it.
And so that’s a new terminology, so to speak. And I think a lot of it is at the application layer of what we’re seeing, and it’s across these different industries. Recruiting is one; marketing is one where you can actually go and execute and change the content pages for you. Historically, that might’ve been a human that was doing it or an agency that was doing it. So it’s taking over a lot of that grunt work, so to speak, that people don’t really like to do. So I definitely think it’s a term that we’re seeing and hearing a lot of across a wide variety of different industries.
Live Question: Hi, thanks for all of this commentary. I have a question. I love this model of the three different layers. When you think about companies like OpenAI launching advertising and consumer-facing advertising soon, how do you think that starts to impact, like at which layer? And do you think that will bring, I guess, more interest in how these companies are using human data and using their own consumer data to improve models?
Nizar: I haven’t had the same kind of visceral reaction as others have had to OpenAI wanting to bring advertising in. I think we all got just accustomed to using Google. And even with their AI overviews, when you see something is sponsored or whatnot. I also think all of us have had a lot of tremendous business outcomes, leveraging the networks and the rails that Google created to be able to advertise our products and depend on keywords. And I think they created a free market where we could compete to get our brands in places. And I think it also ultimately leads to when you have a good product that you’re advertising and you’re willing to pay more for it, you get the results for it.
So when I think about OpenAI specifically down that path, I think you’re transitioning the surface area from where that’s happening. And so if you’re not doing it in Google and your first search is in OpenAI, I think there will be reasonable ways to create an experience where you can have branded searches or keyword searches where you’re seeing logos or brands show up in a way that doesn’t obstruct and take away from your user experience.
I think it’s interesting because you’re right in the sense of there’s a lot more concern around how these firms are using others’ data sets, but ultimately how does Google run that network? They know a tremendous amount about what people are doing and how they’re searching. And I think time will tell how guardrails are put in place in order to protect folks’ datasets, but I don’t view it very differently than what we’ve become accustomed to right now. And I also think it’ll give folks, there will be opportunities, I think, with paid tiers to opt out of having that even happen, which is no different than what they do today with you can choose to believe them or not, but if you have enterprise-level contracts, technically they don’t use your data and they don’t share it. TBD, if that’s true, but technically, ostensibly, it’s supposed to be how it works.
Sabrina: I think there are different tiers. I think they try to roll it out for the free tier to understand how the market would react to it. But also because you’re collecting a lot of information on the consumer, could you do real-time advertising of a product based off what you’re searching? And oftentimes that’s the highest level intent. If you’re talking about something semantically, maybe in the flow, you can recommend a couple of different items. And some consumers don’t quite mind that. Obviously, if you’re at the enterprise level and you’re doing something in your secure environment, that data should never leave the environment, that shouldn’t be used for ad-driven purposes. But I think on the consumer side, it is definitely one business model that could end up working really well for them. And they’ve tested it within a really confined environment.
Nizar: And even if it’s not paid, folks are still spending a ton of time and energy figuring out how to get their brands to show up inside of LLMs.
Sabrina: It’s a whole segment.
Nizar: Exactly. And so we use a product called Gauge that allows us to get a sense of what people are searching for, what our competitors are looking for. When somebody has a specific keyword string that we might use in Google, we try to get a sense of like, okay, where are we ranking relative to our competitors? And then these tools literally help you create the content, the articles, you vet them, you edit them, you put them in place, and then you can track over time how your share of voice is looking relative to your competitors. And so from that in and of itself, you’re spending time figuring out how do we elevate our brands through, I think another tier where you can just pay to have yourself branded. I just think it’s a very natural transition.
Sabrina: Awesome. I think we may be at time. I’m getting the nod that we are. Well, this has been a lot of fun, Nizar.
Nizar: Yeah, thank you.
Sabrina: Thank you for doing this with me and appreciate everybody for joining us today. And I think we’ll stick around for a little bit in case people have questions, but thanks.
Nizar: Cool. Thanks everyone.