Bessemer's State of AI 2025 Report

Live from Schenectady, New York, it's SaaS Talk with the Metrics Brothers, Growth and CAC.

And I'm Growth, better known as Ray Reich, founder and CEO of Benchmarkit.

And I'm CAC, better known as Dave Kellogg, EIR at Balderton Capital, independent consultant, and the author of Kellogg.

And together we are the Metrix Brothers.

And we go together like black holes and galaxies?

Well, we must be talking about AI then, since the majority of galaxies have a black hole at the center. And it's a little like the black hole of capital-intensive funding at the center of building AI foundation models and the hyperscalers.

Okay, well, we certainly have spoken about the stellar velocity of several AI companies like Lovable and Cursor in a recent episode called months to $100 billion. But today we have a different space metaphor, um, here, which is going to be shooting stars and supernovas. Yep.

So it sounds like we are going to discuss the Bessemer Venture Partners State of AI 2025 report, CAC, where they introduce this concept of shooting stars and supernovas.

Right you are, Ray Growthreich. An interesting part of the Bessemer reports is that they often include predictions, and this one does. It includes a top 5 predictions section. And listeners should note that this report, while there is a downloadable PDF, it contains a fraction of the information on the web page version of the report. Not a best practice in my opinion, but nevertheless true. So if you're listening, definitely go look at the web version. If you just download the PDF, there's not a lot there.

You're exactly right there, Kath. But before we blast off on this episode, let's hear a word from our sponsor. The Metrics Brothers is proudly sponsored by Benchmarkit, the B2B AI and software industry's largest, broadest, and most contextualized benchmarking index. Dave Kellogg and Ray Grothright have put the Metrics Brothers podcast on hiatus for the remainder of 2026. All episodes we have ever recorded are still available on your favorite podcasting app. You can reach Dave @kellogg and Ray on LinkedIn @rayreich. Thank you so much for listening to the Metrics Brothers podcast. Okay, Dave. So we said something during the introduction, I think it's really important regarding this State of AI 2025 report is they do have the PDF, but you can go to the bvp.com and their Atlas Research section. And that's where you're going to see almost 5 times as much information on the benchmarks and the research. At a high level, you know, Bessemer's invested over $1 billion in the last 3 years on AI-native startups, which provides them some pretty good insights. I think it's also important to know that this report was based upon only 20 companies that they did deep research on. And that's how they came up with these 2 AI archetypes. We have supernovas and shooting stars. They're going to talk about. And then they also talked about things like the importance of proof of concepts and pilots, why vertical AI is such an attractive category. And then they give us their 5 predictions and the top 10 takeaways for native AI founders. So you want to hop into it, Dave?

I do, Ray.

Okay. So the first thing at top of the website and the first 3 pages of the PDF, they talk about these supernovas. And supernovas often will hit up to $40 million of ARR in year 1, then soar to over $100 million in year 2. Similar to what we talked about with, you know, the new metric in AI, which is months to $100 million ARR. That's what the supernovas really represent. However, they sacrifice gross margin. We also talked about this in previous episodes.

Yeah.

Average gross margin of one of these supernovas is closer to 25% because they have such high infrastructure and compute costs. And they're okay with kind of using a blitzscale-type technique with lean teams, and they'll exchange burning cash for fast growth. And what's interesting in these supernovas— and they never shared which companies that they used in their research, Dave, also probably not a best practice. But the average ARR per FTE was $1.13 million of ARR per FTE. And if you look at companies that kind of fit into this, you know, Lovable, Cursor, Bolt, those are some companies you could label as these supernovas.

Yeah. So, so my take on this, Ray, is this is not even really benchmark data in my mind. This is what I would just call definition of terms that Bessemer is coming out and saying, Like, just like I believe they actually defined the term centaur, which was $100 million ARR. I can't remember. Is that right? Was it Bessemer or somebody else?

Nope, they did. You're exactly right in their State of Cloud report.

Yeah, there we go. So, just as Bessemer defines centaur and said a centaur is definitionally basically a cloud SaaS company that hits $100 million in revenue as contrasted to a unicorn, which is $1 billion in valuation, right? But just as they defined a term centaur, They're trying to define 2 new terms here. One is supernova, the other shooting star. I think, you know, supernovas were bad. I was in astronomy club in high school and like supernovas explode ferociously. But, and I guess they're trying to get the good part of exploding.

But no, Dave, in what was almost a throwaway line in the report, they said be aware or beware of supernovas because they can Burn out quickly, just like you said.

I was going to say that often you could end up a nebula. I mean, that's the typical outcome if my memory from Explorer Scouts is correct. Um, but in any case, they're trying to define a term here called supernova. This is not benchmark data. They're just defining it. They're saying— and they use words like average, which must lead you to say, hey, $40 million ARR on average in year 1. on a supernova, but by year 1.5, they're at $100 million. But again, if you look at some of the language of the slide here, Ray, this is what they say for CloudCentaur. On average, the top private companies reach $100 million in ARR in less than 7 years. Okay, well, on average, mean or median? And what's most importantly, what's top, right? So, so I don't consider this a kind of a benchmark report. I consider it definition of terms. And Bessemer wants to say, we're going to call you a supernova if you're 1.5 years to $100 million. We're going to call you a shooting star, which we'll cover in a minute, if you're 4 years to $100 million. And by the way, we already call you a cloud centaur if it takes you— well, centaur itself means $100 million in ARR, but on average, it takes 7 years to get there.

Hey Dave, I don't think we should rename our podcast the Astronomer Brothers. Let's keep it with the Metrics Brothers. But let's talk about those shooting stars.

We haven't pulled the trigger on that decision, Reddit. It's out under survey as we speak, whether we should rename the podcast to the Metrics Brothers. But let's talk about the AI shooting stars.

So shooting stars, once again, this is on the first main section of the website, slide 3 in the downloadable PDF, right? The growth is deliberate but very fast. In fact, we're going to introduce you to a new acronym in just a minute. But it's 1 million to over 100 million in 4 years. And they typically are saying, you know, around $3 million in year 1, $12 million in year 2, $40 million in year 3, $103 million in year 4. And even then, Dave, because I think those are averages, it doesn't really fit their new acronym of growth expectations for native AI, which is Q2T3, which is quadruple, quadruple, triple, triple, triple. That's what they're naming this versus the old You know, triple, triple, double, double, double.

Now, got it. And the question with those kind of terms, Ray, and I don't know if they fixed this here or not, is off what, right? Am I quadrupling off $200K in ARR to $800K and then quadrupling again to $3.2 million, right? And then, then tripling enough to, to $9,278,100. Uh, I hate when people define these kind of isms and they're, they're not clear about what, because T2D3 suffered from that. And I think Q2, T3 may suffer from the same thing. But maybe you've read— maybe you found something in the report that I didn't.

No. And when you only have a sample size of 20 and a few of those are supernovas and maybe, maybe it's teens of shooting stars, it's going to be hard to have more empirical data. But the other thing they did say was gross margins perform better in these Slower growth AI-native companies. Around 60% is the— and I don't— they didn't say average or median here, Dave, but I'm guessing its average is around 60%. But what was amazing was they showed these companies only had $164,000 of revenue per FTE, which to me sounds like a more traditional SaaS company.

Yeah. And look, Ray, once again, I view this as terminology definition, not benchmarking. And they call it AI benchmarks on their slide. And the subtitle is what great AI Startups Look Like in 2025. And I actually think the subtitle is more descriptive than the title. This is what they think a great startup, what they as tier 1 investors, tier 1 venture capital investors, think a kind of great startup looks like. So whether that's top quartile, top decile, who knows? Whether it's mean or median, who knows? To me, they're just trying to paint— and the thing that's good about this is they're painting a concrete picture of what they think it looks like. And on the supernovas, I forgot to talk about gross margins there, but my favorite footnote is for AI supernovas, gross margins are often negative. So many of these supernova companies, if you've got negative gross margins, your check to Amazon is bigger than the check you got, right? And you're basically selling dollars for 90 cents. And the question is always, what happens if you could sell a lot of dollars for 90 cents? What happens when you raise the price to $1.10 or $1.20? And volume certainly falls off. So Negative. There is a question on these supernovas of how recurring is that ARR, right? These shooting stars look much more to me like regular, just very high-growth SaaS companies. I agree with you. The gross margins are a little bit on the low side, but they're certainly not insane. And the ARR per FTE is kind of right smack in the middle of most of the benchmarks I've seen.

Yeah. Now, one of the things— so, by the way, if you're on the website for the Bessemer State of AI report, there's a lot of discussion around dark matter, systems of record. But I went to the vertical AI section next day, which is section 4 of the website version. And it's where they really highlighted that vertical AI is one of the fastest growing sectors of these shooting stars. Actually refer to things like, you know, some of these vertical AIs are growing 400% a year. Gross margins are strong. But let me give you some examples. And these are not— some examples are in the report, some I pulled on my own, but they're saying that context, the context of the industry, the industry jargon, the actual data that a company has to help train these generic models is going to be really important to developing a moat. That is topic-based. An example is in healthcare, Abridge. Abridge is being used a lot for clinical notes, et cetera, or legal due diligence. Lexian and EvenUp are a couple examples of that. So they take these horizontal models, they train them on their internal data, and that's what helps build that competitive moat and actually provide answers. And workflow recommendations versus just text responses. I don't know if that makes sense, but they're saying vertical AI is a great way to build a wedge. And you can start with a key pain point that almost every buyer in that industry is really aware of. And they're saying those that have multimodal aspects to it, like there's some voice and audio, are even more compelling than it's just text or just— audio. The multimodal is really important.

Yeah. So, I like— look, I liked vertical SaaS, Ray, and I like vertical AI. I think anytime you have a chance to specialize a software package in the domain, you end up with a better software package. And the question is always, is the market big enough to support it? And the answer is usually yes, in my experience. So, I like this idea. Just basically the extent you can train a software model underlying a software package in a specific vertical, in its domain, in legal. I mean, even back in the day, Ray, at MarkLogic, we used to do legal classification where we just try and take a bunch of legal documents and say, is it a brief? Is it a case? I can't remember what else, but what are the different types of documents that you have? AI, machine learning back then were always good ways to do that. So, so, because basically if you could train it on legal documents, it would get better at recognizing legal documents. So, uh, I like the idea. I like what they have to say and the startups they talk about here. I've talked to a few legal tech startups myself in AI. Um, so I like the prediction. I also like what they had to say about more horizontal stuff, uh, just going back to talk about, you know, seeing similar basically model-driven wedges in HR and recruiting. Right? Enterprise search and FP&A, 3 spaces that are dear to my heart.

Okay. Now, I'm going to jump a little bit because I wanted to just pick the top 5 to 6 things we could discuss today. So, they have several predictions, but the one I just wanted to talk about real quick was prediction number 3 around evaluations, evals, and data lineage will become a critical catalyst for AI product development. And I kind of thought about that and thought about the recent MIT Nanda report that suggests that only 5% of AI proof of concepts are converting to broad commercial deployment. And specifically, when you look at some of the proof of concept success measurements that Bessemer highlighted here, they were talking about having private and use case specific evals. built on proprietary data. So just having a model that was trained and performs well on a horizontal basis is not enough. You're gonna have to train that model specific to your own proprietary data that's use case specific, and that's gonna significantly increase the performance and productivity of that model in a company's environment. They also talked about that will— those proof of concepts will also provide an opportunity to have real business-grounded metrics Things like, what's the accuracy of the recommendation of the LLM in my environment? Are the outcomes— can I measure them? I.e., I actually resolved X% of my support cases with one interaction. Things like that, Dave.

Yeah, I mean, to the extent AI gets us closer to selling results rather than features, you need to be able to prove those results. In many cases, people will be priced on those results. So, Both sides want to prove them, right? The customer wants to prove them to make sure they're getting value, and the vendor wants to prove them to make sure they're getting paid. So, I think there is alignment around this. They talk about kind of eval suites and companies that provide AI measurement infrastructure, like Braintrust, LangChain, BigSpendedAI, Judgment Labs, are basically pioneering an infrastructure stack for this new era, offering harnesses, etc., benchmarking environments. And I think this is actually really important, And I think it's a somewhat subtle observation. So, so I think it's a good one, right?

Yeah. And there was a couple other things about the importance of these evaluations and why it will hopefully lead to a higher percentage of conversions into production. They use this word data lineage. Now, I've heard it, Dave, but quite frankly, I wanted to learn a little bit more about that and why it's so important. And that is with these AI models, where are you getting the data from? Is it a system of record? Right. How did you use it? How did it change over time? So that data lineage is going to be so critical to these agentic AI models actually having confidence, especially in regulated industries.

So, so, so look, I mean, data lineage, as I'm sure you know, the concept's been around for a long time and there have been data lineage tools for decades at least. But, but, you know, AI has made it more relevant than ever, right? Because now people really care when we're training a model on data. If we're just making a report out of data, then you can look at the report and go, that looks wrong, where did you get that? Then we can go debug that real-time. When you're training a model and the model is then executing your business for you, you don't have that opportunity to go, hey, wait a minute, that looks wrong, because you've trained it and then deployed it. Yes, knowing where the data came from, But basically, AI has made metadata much more important than it used to be. And that's why companies like my friends at Elation have focused on this a lot, because ultimately, when you're building a model that's going to run your business for you, you care that it works and you care what you fed it to build the model. Yeah.

You know, in these last 5 minutes, Dave, what I'd like to do is go to page 4 of the PDF. And that's the top 5 predictions, because I always like that Bessemer does this. So I'm just going to highlight all 5, and I'd love to get your feedback on the ones that jump out to you. So first of all, they said that the browser will emerge as that dominant interface for agentic AI. These are agentic browsers like Comet and Dia, and which is going to actually allow to have interactions across tabs and sessions. So addressing some of that memory issue we have with today's AI model. And it'll actually be able to help automate multi-step workflows across different systems of record. So, I thought that was really interesting. They thought the browser will emerge as a dominant interface.

Yeah, this one caught me by surprise because somehow I'd missed this trend. So, I would advise people to spend half an hour researching it. I just found a video, I think her name is Julia Moore, at Andreessen Horowitz, where she does a live demo and review of Idea and Comet. And I was in the middle of watching that when we started the episode. It was not obvious to me why you need to have kind of an AI browser to do AI, like, at all. But that video is making it clear. So I think this prediction is spot on and something is going to happen. And I think it's probably— remember there were rumors a while back that somebody was going to buy Chrome for $30 billion? Some people said it was trolling. I can't remember who it was.

It was actually Either it was either Anthropic or Perplexity offered $32 billion just as a trial balloon out there.

Yeah, borderline trolling, like, like no one thinks they're going to take it seriously. But the— but there's a reason for it, right? The reason it was funny was that, wait a minute, AI might be a massive disruption to browsers. And it looks like I think it's going to be.

Yeah. Then a second prediction was that 2026 will be the year of generative video. I still remember '24. with companies like Midjourney was kind of the year of images, generative images. And then we've been doing a lot of voice and voice context this year. But, you know, these are products like Google's VO3 or OpenAI's Sora. So for those of us who like to do a lot of experimentation with AI and LLM, it's going to be the year of generative video, Dave.

Yeah, agree. And no comments on this one. I'm not a big video guy, but I think the trend is correct.

Hey, we already talked about prediction 3, the why evaluations and data lineage will become so critical for AI product development. They did, and I'm not an expert here, so I didn't bite on this, but they believe that a new AI-native social media giant could emerge. And they gave an example. Just think if your AI assistant can make sure you never miss a local event or a birthday again, because it can look at Hundreds of different sources of information, or if you have birthdays in 3 different places, they could bring it all together. So, a new AI-native social media giant, Dave.

Yeah, I don't know about this one. I guess it's possible, but this one's less obvious to me. On the prior one, you covered it real fast, but I think the one word you left out that matters is explainable, that enterprises want AI they can trust, right? And I always think explainable is a super important variable. We won't drill into it here, but It's very important that these models be able to explain to you how they did what they did. Otherwise, people aren't going to trust them. Basically, one of my beliefs from working with machine learning and data mining back in the day is people don't trust black boxes. So, any black box aspect needs to be kind of made more transparent.

That's a good one. Now, the last one, I don't think they went out on a limb here where they say the incumbents strike back as AI M&A, mergers and acquisitions, heats up where a lot of the legacy software vendors will start buying these AI-native companies to apply their intelligence on top of their data. I think I already see it happening, but that's their big prediction. I'll give you an example. I think about someone like Semrush and Predis.ai where they're combining right now where Predis.ai can help you do organic content creation based upon how you're performing in SEO and where. So, I think we're going to see a hell of a lot of that, Dave.

Yeah, I think that one's— this is a pretty safe prediction in my mind. We're already seeing it in the form of acquihires or reverse acquihires. But you'll see it through either paying up for entire companies or doing acquihires. But part of this is a war for talent. Part of it is a war for data. And part of it is a war for technology. And you can get any one or all 3 of those things when a big company buys an AI company. I've often felt that software looks a little bit like pharma in this respect, that I don't know how much drug creation Big Pharma does. It seems like they let little single drug companies do the innovation, and those that make it through, they then acquire. So, it's kind of an acquisition-driven R&D model. I'm not sure about that because I'm not a pharma expert. But I sometimes feel like software is that way too. How much, you know, SAP, Workday, how much innovation, Salesforce even, do these giants do versus just buying companies, either market leaders as Salesforce often does, or kind of technology tuck-ins in order to get innovation in their suite?

Yeah, I love the strategy of off-balance-sheet R&D and experimentation, Dave. But that's a whole other topic. So the last thing I'd say, they ended their report, and this is on pages 11 and 12 of the PDF and at the end of the online. They had top takeaways for native AI application founders. And don't know if we wanted to go through all of those, Dave, but I think it was really interesting about some of the topics we discussed. The 2 AI startup archetypes and shooting stars might be a better model to follow. That memory and context are critical moats to native AI companies. And that vertical AI can provide you a unique wedge based upon the specific jargon and domain expertise, and that the browser can actually be the canvas for native AI applications.

Yeah, I think these are all really good. I think the most important part about the startup archetype, so this is what investors are shopping for. So if you want to just sell an AI product pitch without AI numbers, as I've said before, you're going to have a lot of trouble getting an AI valuation, right? So it's always really good to know what the VCs are shopping for. Even if you're not necessarily that, it's going to help you understand their reaction to you. I think there is a generation of AI-native apps that are coming, and they're not going to be bolted on. That's kind of their takeaway 3, and I agree with that, that this is probably, Ray, since cloud, the biggest disruption opportunity in enterprise software, right? Not probably, definitely. So, like, in my lifetime, there's been 2 big enterprise software disruptions. One was client-server and the other was cloud/internet. And I think this is going to be the third.

I totally agree. A lot of people are saying it's 10x bigger.

I had to—

I want to end on this though. And that was prediction 10. Taste and judgment are your differentiation. And it says in the world of agents and automation that human insight is the edge and that native AI application founders should really remember that it's their intuition And knowledge that could help define the next year.

Yeah, I thought you were talking about me when you're saying taste and judgment are my differentiators. I was really impressed. You're never nice to me.

Yeah, Dave, I think with that, maybe we should say goodbye for the week.

Okay, well, thanks for coming for this week, and we'll see you all next week.

Okay, bye everyone.

The Metrics Brothers AI, SaaS, and More is a production of The Metrics Brothers Growth and CAC and a member of the Benchmarket Podcast Network. By accessing this podcast, you acknowledge that the Metrics Brothers make no warranty, guarantee, or representation as to the accuracy or sufficiency of the information provided or the humor content of the jokes told.

Ray?

Information, opinions, and recommendations presented are, according to our spouses, probably wrong and provided for information purposes only. This podcast should not be considered professional or, for that matter, unprofessional advice. We disclaim any and all liability for any direct, indirect, undirect, misdirect, incidental, special, ordinary, consequential, inconsequential, or other damages arising out of any use of, or God help you, reliance upon the information presented here. Ray Grothreich is based in New York City and available on X @rayreich. Dave Kellogg is based in Silicon Valley and available on X @kelblog. Schedectady, which is French for unspellable, is not our actual production location. Thanks for listening.