The Future of AI in B2B SaaS: Insights from Synthesia and Theory Ventures. Hosted by Jason Lemkin

Is this going to make B2B apps an order of magnitude cheaper? What's the impact?

I think the impact is your gross margins are so much better than they used to be, right? Like the second biggest cost for AI-enabled SaaS applications has been inference. And when we were looking, like a typical software company is probably at 75% gross margin. A lot of AI software application companies are like 0 to 30%. And now their single biggest cost has been reduced by 95% overnight and they didn't have to do a thing. So now they, now, now they have AI and you, Jason, you and I both know this, like the growth rates of AI companies are much faster than say the classic SaaS businesses just because there's so much in them. And so now you have companies that are growing 1 to 7 or some massive growth rate before they used to have challenging gross margins and now they'll probably look the same as SaaS companies, which means you have pretty significant multiple and valuation expansion.

Yeah.

So I think it's at the application layer. This is like an unbelievable Christmas present that, that came on January 22nd.

Thank you. Thanks for joining our Digital AI Day for 2025. We did this last year and with the goal of really doing a deep dive at the intersection of AI and B2B. And luckily for our kickoff today, we've got two folks joining us, just to have a discussion on what's literally going on today. And I think it's some ways it's so— I can't even keep up. We'll talk a little bit about Deep— I can't even keep up after Deepseek and everything. It felt like a year ago was insane. We've got Victor Ripabelli from CEO of Synthesia. We'll talk, we'll first do a little bit of a deep dive there. Synthesia just raised a big round at $2.5 billion or something like that. You gotta keep up with, with OpenAI and, uh, no, that's like a seed round these days, but pretty impressive. We'll dig in a second, but Victor has been at the intersection of AI and B2B and enterprise doing AI-based video and training for enterprises. So no one's been thinking about the intersection of the two probably more. And Victor was with us in SaaS Europe in London last summer, and I'm going to ask him what's changed. And obviously we all know Tomáš Tunges from now Theory Ventures. In the last 14 months, he's raised just only $900 million or something across two funds in 14 months or something like that. Is that about, right?

A little less than that, closer to $700 million.

But okay, the pace has picked up for AI investors. It used to be that like $60 million was a lot for a first-time fund, $50 million. Now guys like to— now granted, not a new investor, but a new fund can raise like approaching a billion dollars in 24 months to invest in AI and data. Honestly, it is a derivative of what's changing, isn't it?

But check sizes are up, Lead check sizes are up 42% in 18 months.

Yeah. So I want to dig into that, but I can't think of anyone to talk about right now the intersection. And I want to go into some of the data with Tomas. But before we get there, Victor, get us up to speed. So you've been doing— give me a deep dive on Synthesia and where it wins. But if roughly speaking, you've got hundreds, of relatively larger customers using your AI training. Here's you yourself on the screen, right? You think using your own product, doing in many cases nuts and bolts B2B activities, right? Corporate training, sales training, external on websites, internal SKOs, which we thought we wouldn't need any salespeople with, with AI. It turns out we need more than ever, but you're at the, at the intersection of an area that is competitive, right? There's a there's more than one vendor doing video in enterprise. You're on one side or the other of 9-figure in revenue monetizing this. So get us up to speed on what you do today. And actually, what would be interesting to me is like with all the change, what are you most excited about that you're going to release this year?

Yeah, thanks for having me. Today we're the largest AI video platform for the enterprise, right? And 4 years ago, that all started with avatars. So Most of those who are familiar with Synthesia probably think of Synthesia as being like the avatar company. We invented this technology where instead of recording yourself with a camera, you simply create an avatar of yourself or use one of our kind of off-the-shelf avatars. You type out something and boom, magically a video comes out, which obviously is 100 times more efficient than having to record yourself and gives you all the added benefits. You can go back and edit it after the fact. You can translate it into different languages and a whole bunch of other really cool things. And where we've gotten to today is That is still for sure a very important headline feature of Synthesia. But really, we are an AI video platform today, and we help our customers predominantly in the enterprise. We also have a freemium model, self-service motion, but the enterprise is what we're focusing on, on communicating better. And that could be with our employees, could be with customers, could be with partners. In 2024, people want to consume their content in video or audio format. They do not want to read anymore. Right. There's a lot of tech in enterprise.

It's true.

Very successful at doing is building a video tool, right, which has avatars as a component, has AI voices as a component, has an AI screen recorder as a component. But everything is put into this Canva PowerPoint style flow, right? And so the reason we're growing so fast is because what we essentially have done is we've enabled any white-collar office worker who knows how to make a PowerPoint to make a video, right? Before, yeah, you have to go to the video production department with your idea. They would stack rank it against 10 other ideas. They'd say, no, Jason M. Lemkin, your idea is not important. Probably enough. We'll do what the marketing wants to do. And you have to go through all that stuff, right? If you did get managed to get a video through, and you're doing customer support, then one month after you've made 20 beautiful customer support videos, they're all out of date, because you changed the product name, the policies, they're all out of date, different, everything's out of date, right? And that's a horrible thing to change. But with this new platform that we've built, right, you can just go back and edit it after the fact, which is incredibly powerful. And so I think what we've graduated from over the last 4 years, right, is being the training company with avatars and now being the enterprise video platform. That's what we're building. And the way we think of that is if you look at the video value chain, no one has ever managed to build a really big company about enterprise video. And we think the reason for that is that the value chain has been completely disjointed, right? You'd have a camera where you shoot some content with, you go to Adobe to edit it, you go to Frame.io to collaborate on it, you'd go to Vimeo to then publish that content. Or maybe YouTube, depending on what your goals are. And what we've done is actually we've collapsed all that into one platform, give you all the tools to make the videos. And it's so easy that anyone can do it. It's a full modern collaborative platform, right? It's a content management system where you can have teams and organizations and your CAO is happy because we have audit logs, all the sort of stuff you need, right? It's collaborative and you can translate, update the content. Then we also give you a publishing platform, which is a relatively new product for us. Essentially, it gives you a video player where you can deliver the content to your end users. And that's exciting because when we own that entire video value chain, we can this year begin to do a lot of things you couldn't do if we hadn't bundled it. For example, we can begin to use the engagement data of the videos to automatically create new versions of the video that perform better than the previous one. We can make your videos interactive, bidirectional communication, which of course is— think of it as AI agents, right? Where you can— it's not just the video that goes, plays from A to B. It's like an experience. It's like a small app almost you can interact with. So really what's exciting here is, is the future of communication, which I think you look at people's private lives, everybody listens to podcasts and watches TikTok videos and YouTube videos. But when we go to work, the buying process in as a customer, right? When you're receiving support somewhere, all that stuff's texted. We're helping customers and we're helping enterprise essentially uplevel their communications, which, which drives all sorts of different business outcomes depending on our customers.

So you're saying that folks that have blogged for a while, like Tomasz and I, we're just, it's dead. Like there's, like, we just, we should just, it's, it's over. It's, it's distant. Time to hang it up. Distant past. Time to hang it up. There actually aren't many bloggers left. Okay. One quick question on this. And then, then we can dig in a little bit into, into today. And I want to, let's use your round in a second as a case study. Let's use it with Tomasz. But what is some, with the latest models, latest everything, we can talk about DeepSeek for a moment in the next. What's something you're excited that you can do this year in an upcoming release that maybe you couldn't even do last year? What is something literally that's at the bleeding edge that you're doing for real in the enterprise?

So we're launching our new AI video models within the next 3 to 6 months, and they're going to be a massive upgrade, both in terms of quality and fidelity of the avatar. So those of you who've tried the product will probably have experienced the avatars are great. They're a lot better the last 12 to 18 months than they were previous to that. But there's still some way to go before it's completely photorealistic, right? We've got to pass through that barrier, and we're also going to enable you to basically prompt your avatar into a new environment. For example, if you create an avatar of yourself today, you're stuck wherever you record your avatar in your office, for example. What you can do with this new tech is you could say, hey, put me in the jungle, put me in the supermarket, put me like I'm playing tennis or something like that. That opens up an entirely new creative dimension in the product. And so those two things together have always been one of the core mechanisms of Synthesia, right? The better the quality of the avatars and voices, the more TAM we unlock. So a lot of people laughed at us 4 years ago for going after the training market. And because it's not a very sexy market, VCs hate learning and development.

I get it, but I get it. It didn't, maybe didn't sound like the hottest space, right? We get it. But a mid-sized market in all fairness, right? Even it's mid-sized, right? Classically.

I think more importantly than that, you could get to double-digit ARR just serving that market. And the thing that we, I think, saw that are, there's so many graveyard of competitors that have died around us over the last 8 years was that we focused on a use case where the technology was actually good enough, not where we could dream it to be one day, but doing all sorts of cool marketing and sales content. It worked well for training because the alternative to training was a boring PDF document. And so what we've seen over the last couple of years, right, is that the better the avatar quality becomes, the bigger the TAM gets. So the last 12 to 18 months, we've seen a massive shift towards product marketing and customer support, which collectively those two are now actually the biggest segments that we serve. And that's purely because the quality of the advertisers have gotten better and better. We're not making like, we're not making Super Bowl ads and key brand moments, but we make a lot of tactical marketing content, right? You see the cool ad on Meta, YouTube, whatever, you click to the website. And then once you're on the website, then you learn more about the product with some easier videos. And we could just move further and further up that funnel, the better the quality of the advertisers.

So the further you leave the uncanny valley, the customers can see it and you're seeing more market pull, right?

Exactly.

Yeah. Cool.

All right.

I want to dig into that a little bit more, but before we do, just because we— the agenda changed in a week. Deepseek, right? I saw a lot of folks talking about this on social media and X. I don't know if you guys did. I want to get both of your opinions, but I listen, I'm not an— I'm the least smart person probably of all the folks that are watching, but I have an AI test and I want to know what you got. It's called the Sopranos test and I've been running this for a a while.

Okay.

I'm a basic guy. And, but I enjoy writing in the word and I go in and I ask what happens after Victor. Do you watch The Sopranos over there in Europe? Have you seen the show? Probably don't know. Anyhow, famous show in the United States. And it's crazy at the end of the, now I, there is a rule. I think after 20 years there's no spoilers, right? I think that's the rule on Reddit. Okay. So in the very end, the screen goes black and people have debated this for 20 years. What happens in the last? And so I asked, I ask ChatGPT and I ask Claude and I ask Deepsea and Deepsea won. It's the best ending. It was the best ending. It doesn't mean that it gives me my best diagnosis for why my cholesterol's high or others, but that was my moment where it was pretty cool. For this one use case, it seemed like instant parity. First, Tomas, what's your take in general and Is this real? It's only been a couple days, right? Is this going to make B2B apps an order of magnitude cheaper? What's the impact?

I think the impact is your gross margins are so much better than they used to be, right? Like the second biggest cost for AI-enabled SaaS applications has been inference. And when we were looking, like a typical software company is probably at 75% gross margin. A lot of AI software application companies are like 0 to 30%. And now their single biggest cost has been reduced by 95% overnight and they didn't have to do a thing. So now they, now, now they have AI and you, Jason, you and I both know this, like the growth rates of AI companies are much faster than, say, the classic SaaS businesses just because there's so much— them. And so now you have companies that are growing 1 to 7 or some massive growth rate. Before they used to have challenging gross margins and now they'll probably look the same as SaaS companies, which means you have pretty significant multiple and valuation expansion.

Yeah.

So I think it's at the application layer. This is like an unbelievable Christmas present that came on January 22nd.

I want to share a story, but Victor, will this— is your COGS going to— is your COGS a huge issue and is it going to plummet? Or just from the perspective of Synthesia, is this not as big a deal as it sounds from COGS and the rest? What's your learnings in 48 hours?

We have great software margins already, so I wouldn't say it's not something we'd like to put a lot of focus and emphasis towards, but obviously as we scale up, It's great. I would say though, I think I totally agree with Thomas. It's a massive technical breakthrough. I think it's interesting when you look at the market. I think if anything, to me, this is just your confirmation that humans and markets are driven by stories and emotions, not fundamentals, right? It's absolutely hilarious to see all the world's tech analysts who know nothing about AI just panicking because there's like a new model out. I think people have a tendency to overestimate the impact of this in the short term and maybe underestimate in the long term. I think the long-term tail effects, as Thomas just said, for the application layer is huge. I think that's really massive. But I think in 2 months, I think 99.9% of the people who are using Anthropic and OpenAI today is going to go back to using their products, not Deepseek's model. And there's a big differentiation between those two, right? I tweeted the other day that smart people talk about models and benchmarks and customers talk about app flows, apps and workflows, right? And there's such a big difference of a model that performs better and is interesting technically, right? They'll probably have a big again, long-term effect in the market. But, but, but yeah, I think this kind of the panic that's going on right now, I think is pretty funny. I don't think that most people are using OpenAI or Froppy today is going to switch over to Deepseek magically within the next couple of weeks.

I'll tell you just my— and again, I'm the least smart, but my gut, just from my little portfolio and little slice of the world is I think cost is less important in B2B in some cases. Than that you can use more. Because when I look at a couple companies, one, one I invested in the contact center, which is all about AI agents called Gorgeous Free Commerce. They're at almost $100 million. Okay. At first, the AI was a million bucks a month, right? To deploy these agents. Now it's fine. They've optimized it. And then I was a small investment at the low end of video in a company called Opus Pro, Opus Clip, that Victor probably knows. And when I invested, I'm like, this company's got to be burning. You're uploading billions of videos. You're creating clips for free. Like, this business model can't work. And I look and when I met— oh, it's actually cash flow positive, right? And it was wildly expensive, but as soon as they made huge commits and optimized what they're doing, and at the corner, maybe even cut a few corners, not literally, it's a great product, but maybe didn't have to do everything right. They don't render all the videos in real time. They adapted. So my thought is a lot of B2B folks I know, I don't think Salesforce's gross margins are going to plummet to 20% because of AgentForce, but we might be able to make these products in one year like 10 times better. Whether it's the getting rid of the last bit of the candy valley or giving agents out for free. When I, you guys probably know this, but when I talked with Brian Halligan the other day, I didn't realize Dharmesh, right now HubSpot isn't charging for their agents. And this is a related point. I don't know what you think, Tomasz. I think what if everything gets better and cheaper, we just might not be able to justify $2 for a resolution. It might have to be included.

You think it'll be bundled because it's so inexpensive?

Gosh, HubSpot's not sure they can charge. Mark Benioff is, but it's early. It's early. Customers aren't stupid. And I think we've all gotten ripped off the last 2 and a half years with endless price increases, endless price increases. And I think another one, now you're crappy. If your agent's mediocre, I ain't gonna pay for it. I think it might be a VC fantasy, just like the fact, going to Victor's point, just like Deepseek has changed the world in 48 hours. It might be a VC fantasy that every B2B company can charge 50 to 70% more because of these additional per-resolution charges. I just don't know, but I actually think it might be deflationary in SaaS for the first time.

It's much more difficult for incumbents who are adding on AI functionalities that they need to charge for to not erode their margins, essentially. Much easier for a startup who just builds a product where it is the product, right? Because I think ultimately you don't pay for the models, right? You, you pay for the product. And I think it's just far too early for some of these big things. I think frankly, I've also said a lot of them, I don't really, I'm not sure they've figured out exactly where this stuff is valuable. There's a lot of, hey, we're doing like cool things and everyone agrees that it's cool, but does it actually impact your business? Whereas I think some of those smaller AI native companies, I would put ourselves in that category, we actually found a very real problem, we deliver very real business value. And then, you know, it's built into the, it's built into our business model already, right? But I, everything is Copilot, right? Are you really going to pay a lot more for Copilot for all your employees? Like, clearly that seems to not be the case for a lot of big companies because they just can't pinpoint the value.

Yeah, I don't know. What do you think about— I just don't— I'm just trying to learn in real time. I don't know. I just don't know in a year if everyone will be able to charge another 50% for enterprise applications for their AI add-ons and agents. They may be forced to, especially if the cost goes down another— I don't know exactly how the economics really work at Deepseek. But if it's really an order of magnitude cheaper, right? How long, how often can you charge $2 for what costs a fraction of a cent to deliver?

Yeah, maybe, but like, software's never been cost-based pricing, right? Twilio, okay, fine. They stack a margin on top of the SMS cost to the underlying customer. Stripe, okay, but like—

A lot of APIs are cost-based, but keep going. I'm with you. I'm with you.

Like, Gainsight, they're not marking up AWS storage, right? Or like, RDBMS. Compute. So I hear what you're saying. I think, but I do think the point that you made at the beginning, I completely agree with, which is if Deepseek-like models with great reasoning are 95% cheaper, yeah, software companies will just like 10x or 100x the use and probably keep the margins the same to get to a much better user experience as a way of differentiation, right? So I think that the UX, our expectations as users for what these things are going to do are going to go way up. Because they become increasingly sophisticated. And if you play with Deepseek, I was trying to create a flowchart. I took the Deepseek paper, the academic paper, and I put it in a Deepseek and I said, create a flowchart for me of what they did. Because it's hard to read an academic paper. It spent 2 and a half minutes thinking, and I was watching it at first. It said, okay, like the user said, create a single chart, but the experiments actually have 3 branches. So now I have to decide, do I create one single chart or do I create a chart with 3 different branches? And I'm like, what? Back and forth. And now I understand, okay, well, next time I ask it to make a chart, I need to be clear on this point. Just the same way that we learn how to use Google, we're learning how to use these models. And this is one of the first models that narrates the way that it thinks. And I think that way we'll start to get to more outcomes. And then I think the other question that's really relevant to this conversation is, are we selling outcomes now in software as opposed to selling seats? Yeah. And I don't know the answer, right? But the mantra has been, we're selling outcomes, we're selling outcomes, we're selling leads, not, not SDR seats, we're selling resolved customer support cases or videos or clicks. And if we do get to that place, then I think it doesn't matter how cheap these things become because there's some break-even for a customer to use the AI rather than this.

Yes. It's funny. I, I dug up, I have a post coming out up in my queue, but I dug up, I went back through a bunch of those old KeyBank SaaS studies, right? And it turns out going back to the beginning, per seat was less than 50% of SaaS revenue. When you add in API pricing, other pricing, just because we think everyone talks about blah, blah, blah, Salesforce, it doesn't mean it's most of software. And so it's, it was like 47% in like 2020 or 2021. Okay. So I'm not saying it won't continue to decline, but VC shooting from the hip on social media saying the seat is dead. The seat was never, was at best a plurality of revenue. And, but I wonder, I have just one investment I've made where the customer was playing in kind of a hybrid model, right? And listen, this is not like the most profound insight, but it crossed a million bucks that they were paying very quickly. And of course, what happened the next day? Hey, Victor, we love some feature, but A billion bucks. We, we need a flat rate. Do like, no, no more of this like variable cost. It destroys the enterprise. No enterprise, no matter what. And maybe you guys both have insights. Enterprises still want visibility. There is the ability to move some budget around. There is ability to take budget a piece, but everyone wants certainty, right? If everyone want, everyone's, everyone wants Snowflake as much as they can in their Snowflake spend. They want Snowflake here, right? So I don't know if this is as radical as the internet wants it to be. How do your customers want to pay, Victor? Do they want to pay in brand new ways or do they want to pay in the same way their grandfather paid for business software?

We still sell seats predominantly. I think it depends just a lot on the product, right? I think the way to think about Synthesia right now is PowerPoint 2.0, right? And like PowerPoint is one of those things where we have thousands of people using this inside big enterprises. And I think that's easy to understand why we're selling seats. I think we also want to sell outcomes. We are building kind of more specific workflows that solve, you know, we do deflecting customer support calls to the call center or increasing, decreasing your sales prospecting, which we build separately. But I think it just, I think it really just depends a lot on the product. If it's, I think there's probably a lot of companies who over-indexed on seats where it doesn't really make sense. In a creator tool, I think seats make sense, right? But whole bunch of other types of tools where seats maybe isn't like the right, isn't right now.

Yeah. I don't know. I want to move on to fundraising. I just, my only thought is for founders and others. I just think, just got it. You just, I would just be a little cautious on social media think, because I don't think most business customers want pricing that is radically different than their last product. Right. If I paid this much, if I'm paying $300 per month for Salesforce, and I'm moving to this new version, like $200 seems cheap, but you want $0.83 per contact resolution. It's just, it's, that may add friction to the sales process. Right. And there are certain ways. And if I'm, I think of everyone that's like Stripe-esque prices, Stripe-esque that I've seen. Right. And we may, if you're, if, and I think we may be playing, spending too many mental calories and playing too many games on pricing when we should focus more on closing customers in a low friction manner, right? I think this is over, overdone, this issue, right?

I think the deeper insight here, which I have confirmed so many times over the last years, is just to not over-index on social media, I think, right? Or what VCs want, or like what I want as a founder. It's like, just talk to your customers, listen to what they want, help them deliver awesome value, and the rest will take care of itself. But yeah, it's a bit like, you know, I think with all this like model stuff, it's like most customers don't really care about models, right? They buy products. And I think in the tech world, we love talking about technologies and whatever, it's like 8G's latest like hype cycle. But when the rubber hits the road, eventually you can't just live on like buzzwords and grand ideas.

Yeah.

I think that, that, that's always, that's a big lesson for me the last 8 years for sure.

Okay. I want to spend just a couple minutes on overall VC fundraising because I have both of you, right? Victor's been at this since 2017 before the buzzwords were cool, just raised. What I think is a real round, right? $2.3 billion or whatever it was. That sounds like a lot, but you're not at $2.3 million in ARR, right? You haven't disclosed. I won't ask you, but you're on one side or the other of 9 figures in revenue. Like, I think they got a good deal.

Great deal.

Great. Seriously. So, Thomas, I stole this chart from E&Y, which stole half their data from Crunchbase, which probably got it all out of Deepseek and OpenAI anyway. So we're all stealing each other's data, but this isn't mine. But I like this version of what we've seen before. Even though, which is what happened at the end of last year is money went way up, right? As we can see on the left, which we all saw on at least on TechCrunch, right? But deals didn't. Deals went down, right? A lot of it was growth focused and half of it was AI focused. But this says to me, if you're in the hot zone of Synthesia and we, you, Victor, in a minute you can tell us what hot is. It's the good, it's as good as almost as good as 2021, right? Maybe it will be better, I think this year, but overall deals are down. Is this what you're seeing? And so does this mean we're in a real world where the folks that are saying fundraising's harder are right? And the folks that are saying it's froth are right? Is everybody right?

I think everybody's right. You look at Databricks raising $15 billion.

Yeah.

That's as big as the Facebook IPO was, which was the largest technology IPO at the time in 2014 or 2015, whenever it was. And they did that in the private markets, right? Raising $5 billion. So there's, you have $20 billion, let's say, of that $60 billion in Q4 going to 2 companies, a third of the dollars. And so it's really like what you're seeing here is the early-stage deal counts, I think, probably the same. The growth-stage deal counts are down in a pretty meaningful way. After 2022, the Series B through the D market was effectively dead. And it's starting to come back. If you're— if your check sizes are measured in the billions now to lead to financing, you can't invest in many companies.

So for— and what's your pulse for your average? And even this is top 1% or even higher, but what's your pulse for the classic B2B triple, triple, double company, which was what everyone wanted to fund in 2021? What's your pulse on that in 2025?

I think it's a, it's like a lower multiple A. I think it's like a very reasonably priced investment, but the deal gets done, but they all get done.

All the triple double doubles get done.

No, I don't. They're not even maybe on the margin, depending on the valuation and depending on the market. I think the hard part is those triple doubles in classic software. There's a lot more competition than there once was. Now all of a sudden you look at a company in a particular space and the last 10 years, you and I have funded a lot of those companies and many of our friends have funded those companies. And so if you're tripling, doubling without AI, it's a hard sell.

And then that's the thing with AI, like you used to have 2 competitors. Now you have 400. I can't even keep track. Like I can't even meet with a startup that's pre-seed, whatever, because there's 400. How many other AI legal in-house litigators could there be? Like this was the most boring category when I invested in legal in 2015. Now there's 1,000 startups, right?

Kind of like the Groupon era for B2B. Remember that era of Groupon where Groupon came out and then there were 1,000 Chinese competitors within 6 weeks?

Yeah.

Or a few hundred in the US. It's kind of like the Groupon era for AI where one company takes off and then there are 15 to 20 others because you're wrapping the API, you have a lot of capability out of the gate, and then long-term product differentiation probably matters less than go-to-market execution.

Yeah, the products are superficially great now, right? Versus they used to— all our products in the old days, the first 18 months were terrible. Like they had one cool thing they did, one cool thing. Everything else was terrible. Now these products, at least superficially, like you, you talk about AI agents, how, you know, people can't get ROI from them, but superficially all the, the AI SDRs, superficially all the products are great. Like in a quick drive-by, every AI SDR product's great, isn't it?

The demo's amazing.

That was amazing.

That was amazing.

That was amazing. All right, Victor, what did you— on this slide, was this the easiest round to raise?

It wasn't the easiest round to raise. I think at the valuations that we are in now, it's, it's, and especially when you have a really good business, which I think we do, not just from an amazing vision, great start, but actually numbers, customers, spend their contracts, like the more data and numbers you have, the more you will get evaluated in those things, right? I think there's these like very outlier deals where you're like Devin AI or something and you're like, these 6 engineers is absolutely insane and they're going to change the entire world coding. It's like boom, $2 billion valuation. You can do that maybe if you have a really strong story and there's like a team and whatever. But I think those are still pretty outlier. My sense from like also speaking to peers is that when you're raising growth rounds now, and you know, we're not a foundation model company, we're very applied. We're definitely evaluated like a SaaS company, definitely with some premium because I think it's a fair price. I think it's a good investment. I think the risk reward in this route is the best it's ever been. I think we raised the Series C, that was coming out a lot of like hype cycle. ChatGPT had just dropped. Everyone was like freaking out and we were growing super fast and we were closing a lot of contracts, but it was still to be determined, like, can we get to million-dollar contracts? Will people actually get consistent value out of this? Are they just trying it out? Is it innovation budgets? All of those questions were still open. But then we raised this round. What I think stood out to the market here was that not only are we good at closing new logos and closing new business, but we are very good at the renewals and the expansion, right? And ultimately that's where your customers vote if your product is just a cool demo, of which there's a lot right now, or is it actually something that drives real business value? And my sense from, from, from spending time with a lot of investors, right, is that a lot of, there's a bit of fatigue. Very great. You grew really quickly. You closed a lot of contracts.

Yes.

But that doesn't necessarily mean that translates into lasting value. There's this I think there's this weird thing in AI right now where it's very fear-driven and people are very willing to part ways with their money, both on consumer side but also businesses, to try out new things, right? Because everyone is scared that their business is going to completely disappear because OpenAI is going to kill them next month. They're afraid to lose their job if they don't have an AI strategy, if they don't appear to be an AI expert. Reality, I think most enterprise buyers don't really know what they need. They don't really understand the technology, so they try out lots of different things. Which may be a sound tactic, right? But it means that there's a lot of this demo budget that's sloshing around. And I think that's one of the things that stood out to investors in this round.

It's interesting how important durable revenue was for the round, right? That it's durable. Is that— Thomas, that's another thing on the VC social media. Are folks still anxious that so much of this revenue is discretionary, that it's experiment, that it won't stick, that it's non-durable, to Victor's point, or is that fading as a concern?

I still think it's a concern. It's probably not as much of a concern as it was last year. Last year, buyer preferences were changing so fast. The underlying architectures were changing. The models were changing in a really meaningful way. Are we doing— are we doing RAG, different kinds of databases? We're trying to figure it out. And as preferences change, you had whole categories come and go within the span of a year or 18 months. I don't think we're quite there yet. I think where we are, particularly at the application layer, is trying to figure out which AI agents actually produce value.

Like, Yeah.

How much am I willing to pay for a note-taker?

Right?

Leslie's note-taker just joined the waiting room. I wonder how much you're paying for it. It's like $5, $10.

How many do we need? 100 note-takers?

A friend of mine tweeted, he said this is the, uh, he attended a meeting this week where there were more robots than there were humans. And one didn't attend and just sent their note-taker. And I don't know if you've done this, but like, sometimes I want to listen to—

Look at this call. Look at this call. There's at least 20, 30 note-takers. I'm just like that.

Emilia tries to boot them, but they keep coming.

Yeah. And then you, so, you know, you can attend a meeting virtually, whatever, bionically, so to speak. So I think the question is, okay, how much are you willing to pay for that? Will Zoom bundle it in? Will Google Meet bundle it in? And so is there price erosion there? And so maybe that kind of creates churn, but I don't think there's this like category cataclysm that maybe we saw in '24.

Yeah. Let me tie this to the last point I want to give on it, but this ties to a meta question I have. So one of the questions last year was, is this all experimentation budget?

Will it fade away?

Is it durable to Victor's point? Right. Victor had to prove that he has triple-digit NRR and all the things. Eventually all these companies have to converge to real software companies, probably an IPO and everything. Then where I'm really getting confused is on agents. I'm not confused about the hype and what they can do and whether our agents can all talk to our agents and our agents. That's going to be— I think it's exciting, right? I'm just wondering how many of these we can pay for, how many of these can be. And I just looked this morning. Dharmesh from HubSpot has a little AI agent. He's got a bunch of things on his network, but he's got a little marketplace, which I'm sure in 6 months will be an order of magnitude larger, right? He's got 629 agents, right? There's Salesforce charging $2. But this was— there was one on the bottom right is the latest data from Zylo, and they're all fairly consistent, which is basically over the last year, people didn't add any NetApp. They didn't add any NetApp to their stack, right? Whether that means they're actually adding some, but consolidating on their core apps, right? I think I should have known this, but Victor, you can do it. I think Synthesia is a swap, right? It's going to be a swap in some cases, right? You're going to swap. It's not a total, always a new app. You may be swapping out a training app. You may be swapping out part of the Adobe empire as you go, right? It may not, but how many can we, I believe that the, all the energy in venture, in startups, in YC, in SF is somehow based on this idea we're going to pay for 3 to 4 times more agents. And so we're going to like triple or quadruple our app stack. Can we— is it true? And can we process as many paid agents? Or is this all a big VC hype cycle that's going to implode because we can't pay for 500 of these agents? What do you think first, Tomas? Does my question make sense? Is where I'm confused.

So I think, I think we see, we're seeing consolidation across the SaaS stack where you just have bundling everywhere. Like you look at the sales stack, massive consolidation across—

Yes.

—ETR automation. So I, I, I think you'll see a reduction there. I think people will end up buying agents for different reasons than they bought classic software. For example, we're invested— if you're a large enterprise, you have 641 apps, you probably have some pretty significant security apparatus. You have another 75 security products. Those produce 8,000 alerts. Less than 1% of those alerts are reviewed by humans. Now, it's a law that you have to review all of them. People aren't going to buy additional 5 other pieces of software. They'll probably buy 100 agents. And those agents will do the job of the humans.

They're going to pay for each agent $10 or $20 a month, and they're going to manage them individually. The IT team or whoever is going to have a whole other set of 200 or 300 agents they have to manage. Yeah, I think— I'm not saying you're wrong. I'm just trying to learn, actually.

Yeah, yeah, no, I think that's fair. You think about like the BDR of the future, right? BDR today manages one Gmail inbox with max 40 outbound emails. Maybe they manage 2 or 3. BDR of the future will be managing 40 or 50 Gmail inboxes. And I think if you think about there's a deflation in the number of seats because now all of a sudden BDR acceleration, I don't need nearly as many BDRs, but the value per seat goes up in a pretty meaningful way. And so I think maybe what we see is application count stays flat or maybe even decreases. but the total amount of software spend increases because it's capturing some fraction of that labor spend.

But are, so that, that's a, I, I get that and I can't predict the future, but aren't, if you just think about it from an old-fashioned enterprise, gotta manage X vendors, gotta manage rogue renewals, rogue increases. Do I really want an explosion of 5 times as many apps? Whether you call them agents, they're still apps from a procurement IT process, aren't they? Yeah, but another goddamn thing I got to evaluate, do a security audit, do a cost-benefit analysis, make sure it doesn't go rogue. Like everyone went rogue in 2021 and bought 50 sales apps. I just feel that the cognitive load is high. The cost, not for the end user, but for the buyer.

Yeah, I think the successful software applications, like, whatever, as a salesperson, I might have 500 agents, each of which is managing one lead.

Yeah, you could.

And, and so what is that agent doing? It's gathering the news and the social media and the podcast and transcribing it. And that's pretty easy to review. I just go through my list. What's the latest with Databricks? What's the latest with Snowflake? And each one of those is an agent. So I don't think the cognitive overhead is huge if the UX is correct.

But who pays for it? That, I don't think for the, I think the bionic, I call it the mech AE, but the bionic SDR, I believe is happening fast, right? But who's— I've never found a single AE on all of planet Earth or even our galaxy who will put anything on their own credit card. I've never found an AE. So the company has to pay. I've never found a single sales rep, maybe in McMurdoch Station, that will pay for their own sales, but they won't. They got some. So who's going to pay for their AD agents if they're discrete, if they're not bundled into another app, right? If they're not part of a platform, who's going to pay?

I think it comes from the sales. You've seen the sales budget. I don't know, 10 years ago, what was the sales budget per AE? Maybe $250, $300. Today?

Exactly. I think, yeah.

And, and then 2022, what do you think it was? $750, $800, maybe $1,000? Yeah, it doubled. Yeah. And so Salesforce is half of that with Einstein, right? It's $175 base sales seat plus $500 to get with Einstein. And so I think you see some consolidation compression within the existing sales stack and some of that budget shifts to AI. Yeah, I don't know.

I want to try it. I, I, You think about this much more than I do. This is just a meta worry for me is that no one's gonna— enterprises are going to be fatigued reviewing 200 paid agents. Not if they're part of my stack, not if they're part of Synthesia for content creation, not if they're part of HubSpot, but 200 new vendors is exhausting. What my limited experience, both with my portfolio as a CEO and my brief experience in a big enterprise, is there's There's just, there's only so many vendors you can bring in. And there are moments of time, like 2020, when the world ended. And there are moments in time, and there are always experimentation budgets if it's isolated, if it's isolated from risk. But when there's risk, I just worry there's that, that we're not thinking through how we're going to get these through the CIO's budget, right? The CIO's budget. And I also don't think SMBs want to pay for 100 agents on their own. I see SMBs being cheaper than ever today. They're cheaper than ever, right? The SMEs do not want 200 recurring things on their personal credit card.

But why do we— I don't understand this agents talk here because I want to clarify, is this like just features that we're now calling agents or is it platforms or— I don't think that we'll think of— we have 200 agents like researching social media on behalf of our customers. Like you'll have a— you buy a product that does that for you.

Maybe it's part of your existing stack. Well, maybe that's the question I'm asking you too. I think if I'm on— if I'm on Darmesh's agent network—

I think it's a made-up Yeah, I'm, there's 629.

I'm going to find, I'm going to deploy 100 of them.

I don't know. I just think it's such a weird, it is a weird way to talk about software and features.

Like, I understand it's like, it's, you're at the cutting edge. You're supposed to tell us this is what's happening. You're the funny daddy on the, on the team.

Jason M. Lemkin. I got here by not listening to the buzzwords. I just like to build great product that people love. I'm just trying to love. It just weirds me out. All this stuff. It's like, it's features, man. It's like, It's an LLM-driven feature that does something for you. Why are we like trying to anthropomorphize it to be like an agent? Like, I understand, like, it sells in the marketplace and it's like, it's cool. And in some instances, maybe I can understand why we call it an agent. But I think majority of this is just weird. I think 5 years' time, like, I think everyone's like, why the fuck did we spend 5 years talking about algorithms as being like a coworker? It's just like software that does things for us. And sometimes maybe you want to chat with it. That's great. That's a new interface. But I really don't understand this, like, obsessed with agents. And that said, you know, we're probably launching an AI agent strategy soon because everybody likes to talk about this stuff, but nobody wants to talk about features. So I get it from a go-to-market and sales perspective, but it strikes me as one of the—

But it'll be part of your platform. You have thousands of customers. So I don't have to—

That's like, uh, weird, really.

No, no, I don't have the answers. I'm challenged. Second half of 2020 to 2021, somehow we magically found budget for all these new apps and the pandemic did it, right? And if you saw, and the craziest one in my portfolio was Talkdesk, which is call center. And forget about how AI changed it. In 2020, you couldn't go to the office for the call center, right? So everyone in their space, like, they, they grew 5x in 6 months because IT changed the rules, right? The IT changed the rules. And I think there are some rule changing happening, but I don't think in 2026 IT can double. I just think an agent counts as an app. Potentially if it's a vendor. Let me put it, maybe I'm, I'm rambling too much. Every new vendor in the enterprise's cognitive load, a new vendor to put into the procurement system to get approved, to track. And every enterprise knows that if I bring in one, I, it's not just one year for Synthesia. Now I'm stuck for 8 years. That's how enterprises think. They don't think one year. They're like, okay, it's gonna take me the first year I'm learning, the second year I'm rolling it out more, the third year I'm gonna buy more stuff. And they always like double, we, at, Back at Adobe, like when you put any vendor in the system, you doubled the budget because you knew it was going to grow with add-ons and seats. And it was. And so how many of those can you add each year? Even at Adobe size, you just run out of vendors you can— and I don't know if AI changed, if agents changed that rule. They could change a lot of other rules. And so I don't know who's going to pay for that. I just don't know who's going to pay for the discrete vendors, not the agents themselves, but so many vendors, so many vendors.

I do think there are like, there are fundamentally new product categories though. I would put us in one of them. Like we're never going to replace someone or take budget from someone. We probably will over time.

You don't come from the training budget or the LMS budget if you're for an internal use case?

No, we'll, we'll sometimes take some of the video production budget that they'll go that you use with agencies. Yeah. The key point here is that people are not evaluating us against should we switch out this thing with Synthesia, right? It's a net new thing. It's also why we're growing so fast. I think there are new product categories like that, and then there'll be a lot of these agents or features or whatever that'll exist better within the platforms people already use and There'll be add-ons and extra modules.

I believe you that you're not directly stealing someone's creaky old budget, but what I, my limited experience tells me behind the scenes, you're stealing someone else's budget. They're canceling someone that they don't quite need that was on the bubble. And they, this is, I don't think this has changed since I worked at a Fortune 500 tech company. You get together each year. This is why budget's annual. And then Tomas, you want to add that cool product. What are you going to cut? This is why so many things got cut.

Maybe some of the budget, maybe some of the budget does not come from software, right? It comes from like services-based things. So for our case, a lot of the budget we eat into is like video production agencies, right? That's not software vendors or translation budgets or it's like more service-driven things. And I think especially if we're talking about agents, that, that, that would probably be a lot of what that can eat into. So it's, you may add another software vendor, but maybe the budget doesn't come from like other software platforms, but it comes from like services.

That's a lot of the hope on X is that we can turn humans into software budget.

Well, Morgan Stanley ran a study in Q3, I think, and they said 50% of the AI budget was, uh, cannibalizing existing software and 50% was new. So 50/50. Yeah.

So that's probably right. They got the data.

Yeah. And so I bet Victor is right. Like it's translation or other kinds of services, consulting, customization, legal, like all that stuff, like legal outsourcing. That's where the software budget's coming from. It's eating away at the labor costs.

Yeah. So that's the news. So the answer, like a lot of these things is yes-ish. It's both. We are. And where— and Tomas, just maybe to wrap it up, if so, 50% is cannibalization. Maybe Victor doesn't see it, but they're cutting videographer, they're cutting something else. The 50% that's net new budget, where's like the fortune? Like the GDP is still somewhat fixed in the United States. What's it growing? 3% a year, even under our new regime. So for real, where's the money coming from to, to pay? Where do they get— where does Morgan Stanley or Theory Ventures say, where are we getting this money and how long will it last?

I think it's labor. I'm seeing it in our portfolio, right? Like labor. I look at R&D spend and like year over year R&D growth compared to 10 years ago. We have one company growing really fast. Engineering headcount year over year will be flat. Yeah. Because they deploy some AI-enabled IDE and they can write a lot more software. And so instead of hiring another, say, 10 software engineers at $250,000, $300,000 a year, $3 million in spend, they're starting to spend, I don't know, $100,000 a year on AI software.

Do you think that's true from what you think, Victor? Do you think we can do more with fewer engineers with AI in the new world, or is that a bit of a myth?

No, I think we can, but I think there'll be many different shapes of companies. I don't think there'll be one type of company where you really want to go for it. I still think if you want to build a $100 billion company, you're going to need a lot of salespeople, you're going to need a lot of engineers, but I think there's no doubt that productivity is going through the roof. It feels like we are seeing a lot of these types of companies, as Thomas is saying, where you can, and also I feel like the mindset has changed a little bit where there's a reluctance to like hiring more people and like just being more efficient, which I think is super awesome. But yeah, I think it's never black and white, right? But I think that's a definite new type of company, which I think is super cool. But if you want to tackle the enterprise like we are, that's still very difficult to do without adding significant headcount.

Yeah, you might even, if we were on VC Twitter, the way to take it to the extreme is the new vanity metric is ARR per employee.

ARR per engineer is more interesting than ARR per employee. I could tell you why. It's more interesting per engineer, per developer.

And that's like the sexiest metric. I think it's one of the coolest. We're 10 people and we're generating $20 million in revenue. Like those kinds of ratios we wouldn't have seen a decade ago. And now people are striving for it. And we hear it from founders. Like we We only have this many people. We don't want to hire people. We're raising capital, but we're not hiring. That's unusual. That's a very different pitch.

Do you think I'm making this number up, but I think it's about right. And Victor could probably for a case study, and then when maybe we can break that makes it, that's what I see too. Mike. And listen, the world's changing so quickly today, even to the whole deep sea thing. It feels like everything's faster than it was a week ago, but it feels to me there's still the issue around 200, 300 employees. Where you still end up needing 10 people to do everything. And what I feel like this 10 people can get to $10 million ARR thing, it just shifts the time. Like instead of me needing, but it's still at 2, at, at $100 million, at $50 million ARR, I need so many humans and honestly so many mediocre humans. You start having to hire, okay, maybe that's the wrong term. You start after hiring A-minus tier and B-tier, no matter, even though you can't say it out loud and it just takes so many people. To do some of the basic stuff and you need so many salespeople and so many mediocre renewals people. And I just see everything still gumming up today at that size. Maybe if AI gets rid of that roadblock, that'd be awesome. But I wonder if it just kicks the can down the road a year or two.

Yeah, it's like Salesforce hiring 1,000 salespeople to sell their SDR salespeople, right? The agents. I think there's still a lag. I think we'll— I think it'll be definitely more efficient. But I still think that we're far from you being able to You know, at least when you're creating like true enterprise software, I do think that's a bit different than, than like more consumer or like SMBs, but you don't need the human touch, but it still feels difficult to close. It'd be very interesting if a company with 20 million, 20 employees, whatever, gets to close like million dollar deals with enterprise. I think that would be like a really interesting kind of tipping point. But for so far, for most of the ones that I've seen, at least it's still a lot of self-service credit card revenue. Kind of smaller sales team because it's still just very difficult managing a dual desk, right? You just need people for that.

Maybe let me just wrap. Let me ask one last version of this to you, Tomas, then we could wrap. And this was great. So when we start, when you and I started, Tomas, a lot of VCs would yell at founders and say, you've got to offshore to save money. You got to— everything told that the magic solution when we started, maybe you too, Victor, the magic solution with VCs was have 50 or 60% of your India offshore. That would magically cut your costs, right? Magically, because developers are fungible, they're all the same. So just hire a bunch of folks in some country you've never personally visited. And that's my VC solution. Is the new version today, the version of today is if you're good at AI, you only need 10 developers. It almost feels like it's the new version of offshore. AI is the new offshore. Like you, you got, you, why do you need more than 10 guys? I don't know about that.

Yeah, that's a great, that's a great, that's tomorrow's blog post for you, I hope, Jason. I think like, Some people did offshoring really well. They hired a great country manager. They really spent the time to transfer the culture and the cost efficiencies. And then some people basically just mailed it in and you got what you paid for. I think it's the same with AI. If you change your recruitment processes and you look for people who really understand how to prompt these systems really well and think about building with AI, I think you can drive a ton of efficiency. I think if your CEO has been told by VCs, hey, go employ AI within your engineering team and you buy a bunch of people some subscriptions, Yeah, Cursor is the answer. It's not a— you have to change the way that you work and change the way the teams work, and that's really hard.

I also really still think that what I see across, no matter like what task we're talking about, where AI really shines is still augmenting humans, right? In the case of us, we augment people to make big videos. Cursor works so well because it puts the human in the driving seat and just makes you more efficient. I think this like pipe dream of agents, like just doing things completely autonomously, in my view, is still a while away.

Yeah, we'll see. We can break— my limited experience is that an S-tier agent is better than a C-tier employee. This is what you see across the contact center is you usually see that the agents are like in the— they can't do everything, but they're in the top 10% of CSAT because what they do They get done without complaining and they don't argue, but they can't be below. I think the whole world's going to that. Agents will replace the C tier like very quickly and deepen the B tier as they can. And I don't know. All right, guys, this was great. We ran out of time. Victor, so great to have you back. We'll see you in London again. And Tomas, we'll see you at the next one, my friend. Thank you.

Thanks. Thanks, guys.