Bessemer's Talia Goldberg: top AI companies reaching $100M ARR in 1.5 years vs. 6-7 years for prior cloud era

magic. Adio is the AI native CRM that builds, scales and grows your company to the next level. You can get started for free adio. com. And we have our next guest uh Talia Goldberg from Bessemer Venture Partners uh coming into the studio. Welcome to the stream. How you doing? Welcome to the show. Thanks for having me.

Great to be here. Um why don't you kick us off with a little bit of introduction on yourself uh some of the companies that you've invested in, your career, your position at uh Bessemer and then we can go into the report that dropped today. Awesome. Um so it's great to be here. I'm a partner at Bessemer.

I'm based in our San Francisco office. I've been at the firm for a little over 10 years, which is virtually all or most of my professional experience. Um, and I'm fortunate to be involved with companies like Perplexity, Fall, Deepell, Service Titan, um, and a whole bunch of others. How did you get into venture?

I got into venture really early in my professional life. I got into venture in college. Um, actually first round capital started this thing, dorm room fund. Oh, yeah. Yep. Which I helped found with them. It started in Philly. I went to Penn.

Um, crazy enough, First Round's probably like the only VC that had an office in Philly. I don't know why, but they did. And so they started it there. Um, robotics from was it like Carnegie Melons out there or something? Yeah, but not in Philly. That's in like Pittsburgh. Yeah, I guess you gota I don't know.

It's a foothold. I get them all confused. Uh anyway, um take us through the state of AI. Is artificial intelligence good? It's a thing here. It's happening. You know, it's funny. So, in um in 2015, not long after I joined Bessemer, the firm started this report called the state of the cloud.

And it became a very popular report that dropped every year on the cloud ecosystem. And so 10 years later, we've been doing it every year. And it really morphed this year to the state of AI. And we were debating internally like should it be the state of the cloud? Like should we continue with it this way?

Should it be the state of AI? What's the how do you even define what's AI? What's SAS? Like what what what does that boundary look like? But the reality is the center of gravity has moved. Um and cloud may be the delivery surface for AI, but all the activity is there. Markets are being created and rebuilt.

Um and so this year we released the state of AI.

Um and as part of that we released some new benchmarks as well that looked at hundreds of companies probably more like you know thousand plus companies across the Bessemer ecosystem in the broader industry to look at what the new good better best looks like how different business models are changing um and markets are shifting.

So that's the state of the cloud or state explain yeah explain the difference between the supernovas and the shooting stars. I like that analogy. Um it and and it was something that I think people have been feeling um but uh no one had really coined a phrase around it and I think it'll be useful language going forward.

But break that down for us. Yeah. So the the supernovas are really these seemingly out of nowhere amazing growth stories that you hear about and you see on X and Twitter and you're like, "Holy [ __ ] is this real? " And it turns out like it is real. It's kind of mind-blowing.

Um um of this select kind of like top percentile of AI companies that have just totally accelerated and compressed growth into a very short period of time. Um and they look very different in a lot of different ways. different business models, different gross margin profiles, different retention profiles.

Um but just to put um a comparison, on average the top cloud companies um of this like last generation of cloud and SAS took about six to seven years in the current cohort to get to 100 million of ARR. Um and that was considered and is considered like very good um if not great.

Um, and then this new cohort is here and they're like one and a half years we're there. Um, and they're getting to 100 million and it's real and it's not just one. There's like multiple and many data points. Um, um, and so we're seeing it at a shocking pace.

The top percentile are getting there in about one and a half years and the top decile in about four years. There are some trade-offs. So gross margins look different.

Um, in the report there's like a little asterisk by the supernova which I find very funny which is like actually a lot of these companies are negative gross margin. Yeah, I knew you were gonna say that.

Um, sort of you know there's accounting rules aside like you know how we all think of gross margins being um quite different and so not all revenue is created equal but nonetheless the adoption is just astounding. Yeah.

So uh talk to me about the difference in underwriting an investment in a supernova versus a shooting star. I imagine uh if you're investing in supernova, you're excited about the growth, but you have to have a pretty firm view on the gross margin profile, the decrease in inference cost over time, something like that.

Like what questions are you asking when you're looking at a supernova company versus a shooting star company? Yeah, that's absolutely right. I think um the the two things that we talk a lot about, there's one the gross margin profile and then the second is is revenue durability.

Um I'll hit on both on the gross margin profile. It's funny.

If you had asked me two years ago, I was like, "Hey, if anyone that has like gross margins that are negative today, if you just look at the cost of the models over the past, you know, year or two years, and you play that out, like it's 100x cheaper to run a model of constant quality today than it was, you know, a year and a half ago.

I think those numbers are like roughly accurate. Um, so it's wildly different. And yet when I look in retrospect at our companies, it's not like their margins have changed to be suddenly like 90%. So I'm like, "Oh my gosh, what's happened?

" And um the the reality is that everyone is doing things that require a lot more compute. And to keep up with the status quo requires like the next best models that come out, the reasoning models that are more expensive.

Um we're having queries that take a lot longer, that do a lot more complicated work and complex outputs. Um and so the margins have improved by and large and they do improve with scale. So we are seeing that but not nearly at the rate of model advances.

So um I think we still feel quite optimistic about the potential for margin expansion and in fact we see it happening. Um but it's not as dramatic as one might have hoped. Are you plateau pill and should we should we assume that uh inference costs will decline with Moore's law going forward?

Because I feel like everyone's been saying like, "Oh, no.

We're we're not just going to get 2x more efficient over the next 18 months like Moore's law would imply, but we're going to have AS6 and Cabbrris and we're going to bake it onto a chip and we're going to get this crazy algorithmic enhancement and inference cost is going to drop by 100x.

" And it feels like we might be at this frontier where maybe we're more on what's happening at TSMC is what will define defi like lower costs than just like one weird trick. And and the the other important thing is you know the labs have been focusing on raw intelligence versus efficiency.

Chinese labs have been more focused on efficiency broadly and they've they've had breakthroughs. And so if if we've reached a potential plateau and just intelligence opportunity to focus on efficiency. Yeah. But how do you think about it? Yeah.

Like the harder problem to solve is doing the intelligence and the complex thing. And so I feel like when all the energy starts to shift to efficiency, it's sort of a sad moment. So I'll be I'll be sad if that's what happens. Not for the public markets investors though. They want earnings. I know.

Well, you know, well and and I think a lot of a lot of the darlings of of the last couple years need that efficiency because they can't keep selling. Yeah. you know, dollars for for 80 cents or or on the on the shooting star topic. Uh you have this revenue ramp year 1 3 million, year two, 12 million, then 40, then 103.

How can you be an AI company if you started four years ago? I thought AI was invented two years ago. Yeah.

So that benchmark is really what I think of as like the new generation of SAS companies, some of which may be using AI tools and AI features and functionality, but are not necessarily like the true AI native companies. So, I think this is what it takes to be like a really good best-in-class SAS company today. Yeah.

Um and uh and and we'll see how that shifts. But I just want to say one thing on this last point of efficiency versus uh um compute um costs and and intelligence is that I think there's just two curves that are counterbalancing each other. One is like efficiency.

Sure, there's going to be a lot of investment in improving the efficiency, the potential um um for for each token, but the flip is that we still have what we see happening and the reason that gross margins haven't expanded as much as we hope is that the usage and the complexity of tasks is still is still growing.

And if you look at just a category, let's just take video for a moment. Like I think 2026 is going to be a major breakthrough year for a lot of these video models that are just reaching starting to reach a level of quality that makes them actually like useful. Something like 70% of the internet is video.

Um, it's crazy internet traffic and the co when when generating video becomes easier and a lot of video is generated not rendered suddenly you're going to have enormous demands um on compute and we actually really do need that efficiency because video is really expensive and and complicated.

So I think you're still going to see a lot of spend even if the efficiency per token um increases. I mean, if Google can't give me more than like four V3 queries per day for $500 a month, like clearly like the GPUs really are on fire.

Um, in terms of I wanted to I wanted to talk about uh one of the predictions in here and I know and I know you guys worked on this collectively. Uh, but uh prediction one, the browser will emerge as the dominant interface for Aentic AI and we've been covering the new browser wars.

Obviously, you have uh DIA from the browser company, Perplexities, Comet, uh and then Perplexity was in the news yesterday for their offer.

Uh but in some ways, it feels like Chat GPT and like I'm assuming everyone's expecting OpenAI to launch a browser, but at the same time, it feels like Chad GPT and and other products have really replaced so much browser activity.

And so in some ways it's like OpenAI is already competing as a web browser even though it doesn't look like it can literally browse the web for you and it can yeah it can instantiates it in textic web browser. It's just pulling that information back versus like taking you on that on that journey.

So curious for you to kind of unpack that a little bit more. Yeah. So, I started using Comet a few months ago and it's Perplexes. Comet has like totally replaced my Chrome experience. Um, and it completely opened my eyes to what where I think the browser I think opening I must launch a browser.

I don't think it's just going to be in chat GPT. Um, I think they will.

Uh, and I think it's going to be a very important surface area because using Comet has transformed my workflows and shown me um for a few reasons that it's a much better experience um to and the first product that's really infused AI so naturally in my workflows um when you're just out there in the web in your email in your Salesforce if you're on CRM um if you're shopping and otherwise to have an agent that sees everything that has all of that context for everything that you're doing in the browser which is essentially like an operating system now and can pull all of that information in creates a far more personalized and effective experience than when it's totally siloed which is the status quo today in chat GPT sure it can go out and do things but it doesn't actually have access um and that context across everything you've been doing you know when I spend I don't know 10 12 hours a day sitting in front of a screen so it's quite different and if you believe context is key to performance which I do and to creating a great AI experience.

I think you have to own the browser. Makes sense. Anything else, Jordy? Uh any I I I wanted to dive into the AI native social media giant. We had uh the founder of Pika on yesterday, which is somewhere in between a creative tool and and trying to build social features as well.

I would be uh very excited about a net new social platform. I I think there there I I agree with you guys. is there's an opportunity.

I think people on traditional social media today are a bit frustrated like seeing what they think might be a AI generated content and they're not quite sure and so potentially creating a new space that that people as as all the models get better and uh I I can imagine an uh all that content will go on legacy social media platforms but I would be excited about a place that was really a home for it.

What are you hoping to see there out of, you know, kind of in the next year? I'd be excited to see a new social media that's totally built on new AI native thinking and and content. There in the old world or in the current world, we think of bots as bad, like bots bad, humans good.

Uh, I think there will be a company that totally shifts that and can maybe even crack the chicken and the egg problem by using bots to fill the, you know, empty room.

Um, the company I was most excited about for a while was character in this world because it really felt like they had sort of a chance to be this, you know, very different way of actually interacting with AI in a in a more social experience.

Obviously, they didn't fully get to see that through, but uh, I still think there's a big opportunity there. Awesome. Well, we be a knockout, dragout fight. I think every social media legacy CEO is taking AI very seriously. So, we'll see how it plays out, but uh it'll be fun to watch.

Thank you so much for joining the show.