What Gets You Funded in AI: Insights from Felicis, 500 Global & Mayfield

0:00Okay, well, welcome to the VC panel. Today we have three outstanding investors, Patrick, Tony, Anastasia. We're going to cover what it takes to go from seed to Series A. And to begin, let's start with Patrick. Give a little bit on your background and what you're investing into, and then we'll go around. Thanks, Molly. And hi, everyone. Again, my name is Patrick Salyer. I'm a partner at Mayfield. Mayfield is a 55-year-old early-stage venture capital firm. So we invest in lead and co-lead rounds at Seed, Series A, Series B, checks ranging from a couple million up to 15 million. Personally, I'm focused in AI middleware and applications.

0:39So a lot of the areas that I'm seeing and interested in are AI agents. Everything's AI agents today for various horizontals and verticals. We're seeing a lot of innovation on business models, like services companies that are now AI enabled. We're seeing all sorts of AI builders, like lovable for marketing, lovable for sales. So these are kind of some of the more interesting things that we're seeing right now. Yeah, thanks Molly. I'm Tony Wong, managing partner at 500 Global. We got started in Silicon Valley and now in our 15th year, about 2.1 billion AUM. We are pre-seed and seed, so we're often the first check-in.

1:15We're early investors into companies like Canva, Topdesk, Intercom. And then more recently, our AI companies have really just accelerated. So Sakana AI, we're the first investors there, Deep Infra, and Play AI, which is now part of Meta's super intelligence lab. So we're often talking to founders before they even start their businesses. So we're definitely open to chatting. Hi, everyone. I'm Astasia Myers. I'm an AI-focused general partner at Felicis Ventures. Felicis is an early-stage VC fund that invests across Inception, Seed, Series A, and Series B. We tend to lead and write checks between about 1 and 40 million.

1:58We've been really humbled to be part of the journeys of many of these hyper growers from Mercore to Canva, Notion, Runway, Superbase, among others. We think about AI across the stack. And so I've been looking at infra, security, AI apps, and even the application to frontier technologies like robotics with their investment in skilled and defense tech as well. A heavy bench of investors and experts in this next generation of AI. Not to skip over myself, but I am Molly O'Shea. I run Sorcery. It's a tech and investing newsletter and podcast. And I do know that Sam Altman is related to Jack Altman.

2:44I don't know if you've seen that clip, but it's right? It's not. It was, yeah, it was fake. Anyways, so to begin, I just want to know, so what is the current trend that's happening right now? We've gone through so many different waves in AI. What is happening at the earliest stages? Stassia, maybe do you want to break this down? One theme in AI that we're incredibly excited about is AI voice. We've been investing across the stack with AI voice and audio models at the infrastructure layer, middleware, so the tooling to build the workflows themselves, and AI voice native applications. Actually, yesterday, I announced I had led an investment in a business called Assort Health, which is using AI voice for patient communication and access.

3:37And it's one of these hyper growers in healthcare. One of the reasons being is AI voice is a very powerful technology. Previously, you had people manually picking up calls, either people working internally at companies or in contact centers. It was a very painful process with very little automation. Now with the immeasurable improvements in AI voice models, you can not only automate the task, but often the NPS of the experience is better. And so the ROI for these gen AI app companies that are voice native is incredibly high and customers are just ripping it out of the hands of these early stage offerings.

4:24And so we are super excited about AI voice and thinks it's a technology that's in right now. Yeah, just feel free to take it away. Yeah. And I think like Assort Health is an amazing example of what's happening in general with AI agents and what we're calling AI teammates or this idea that like the last generation of software was really about SaaS, like systems records, systems of engagement. And this next generation is about automating work. It's automating and giving each role and persona kind of their own teammate or assistant to complete their workflows and become anywhere, I think, from like 30 to 70 percent more effective and more efficient.

5:04I think what's exciting about that is I think the TAMs are literally 10x what software was because you can actually think about like every individual on the planet and how do they get 50, 60, 70 percent more productive and efficient. And I think over the last couple of years, we saw a lot of companies around horizontals. So sales, marketing, customer success, and every persona within that. And now this is moving towards things like verticals. I think Assort Health is an amazing example. As an example of that, we're seeing this in real estate. We're seeing this in education. We're seeing this in legal and every iteration of that, which I think is super, super exciting, super interesting.

5:42and there's a lot of enablers for these businesses. So like one of the trends we're seeing is reinforcement learning post-training and actually like every agentic flow before you're having humans say, was this good or bad? Now actually agents are being used to say whether these flows are working or not working. And there's companies that are producing that. So the infra level is being impacted by this as well. So I still think we're actually pretty early in this trend. And what's exciting for builders is like you're not replacing software anymore. These are greenfield markets. It's some of these TAMs don't exist, but we think they're very large.

6:14And you're seeing kind of every iteration. And the companies that are building around this, the founders, tend to have a lot of insight in those areas. They tend to have been building in those areas for the last 10 or 15 years. So you're seeing repeat entrepreneurs, first-time entrepreneurs, but folks that have like a lot of perspective about those areas. Yeah, I think, you know, when we look at a company, what I always want to ask is how fast is the technology moving and how broken is the experience? So maybe a couple of years ago, it's pretty clear that text and images, the winners were becoming more and more clear there, but we didn't know who the winners in voice were, right?

6:48So that's why we invested in voice model companies. But today, as Stasia has mentioned, we probably wouldn't invest in another voice model company, but voice agents is something that I think the experience is still relatively early and still relatively broken, but the surface area is huge. And that's where we look at, okay, well, there's definitely a problem to be solved. If you're building in that space, the question I would have for you is, as a founder, the voice agent approach, which we love, still feels like an interim approach, because in the end, everything will either be API or MCP, but the voice agent is actually a really good interim approach where if you don't have an MCP, if you don't have an API, you get to just pick up the phone and call.

7:44So we really like that. And as a founder, you kind of want to start navigating how do you go from sort of this interim approach to something that would be durable. So that's one thing that we ask founders when they're building in the voice space. Other areas that we like, security. Anytime you have a new technology, there's obviously new adversarial vectors that open up healthcare, fintech, and others, as the folks on the panel mentioned, lots of verticals and surface area there. Tony, before we go too far, could you just share what MCP means for people that don't know the term? Oh, yeah. So basically the model context protocol where, you know, Anthropic came out with that first.

8:26It looks like all the other labs are now starting to adopt it. it's basically a way to interface with, you know, if you don't have an API, if you don't have a specific way for a model to be able to relate to your database, that's a protocol that, you know, OpenAI, Gemini, Anthropic, all these labs that are building agents are now using. So basically, it gives you a way to, you know, interface with anything. So it'll be a future agent to agent interaction as opposed to sort of mouse and keyboard using a UI. So that's a really exciting future. Obviously, there's lots of applications, sites, merchants, brands that don't have it supported yet.

9:13But we believe that that's where it's going. But in the interim, you'll see this messy middle. The hot topic, it's a trending term, have to have to clarify that. Okay, so now as we, I did want to lay out more on the investing side, We have access to some phenomenal investors right here in front of us. So I'd love to understand more about the growth expectations between the seed to Series A, because the graduation rates have kind of dramatically taken a turn, whether or not you're an AI company or you are. And so I'll bring up this slide deck from Carta that we have access to. And feel free to tell me to go to any slide that you want.

9:53But I guess to start off, what level of growth do you now expect between seed to Series A companies? It's become kind of a contentious topic with this recent clip of Hamant on 20BC talking about getting exponential growth much, much faster. But where do you realistically see this kind of growth? Yeah, that was a pretty interesting clip of you have to go from one to 15 to 100 million. And that is interesting versus the one triple, triple, double that many builders have been anchored around for the past few years. growth expectations are higher than ever before. There is a perspective that the time to build an AI is now and there's so much revenue to go after and so much opportunity that founders should be heat seeking missiles for cash in the back, essentially.

10:52And so we tend to see businesses that are in a strong, healthy position to raise their Series A at different levels of growth, depending on the category. I think I do a lot of dev tool and infrastructure investing, and that 1 million RRR number is still a good signal, but the time horizon to get there is no longer a year. Usually there's an expectation of that shrinking to the first two quarters or three quarters. And then for Gen.ai applications, the bar is higher of two to three million. I think this, one of the slides that I think is the most interesting to look at of the data that you pulled is the conversion rate today from seed to series A.

11:44It's hovering below 20%. And that is a function of the expectations and flight to quality for investors. Because when they see these breakout growth rates, they'd rather be part of a journey like that that they think can go the really long distance. The other thing that we are observing is kind of the manipulation of the data to try to demonstrate that you've achieved these milestones. And so we take a very fine tooth comb to looking at the definitions of revenue and ARR that the founders are presenting because often they're trying to use other proxies like credits or pilots or NREs to show that they have this level of growth.

12:41But really, in fact, it's a much lighter weight form of customer engagement. The last thing I'll say about this change in expectations from investors is that one thing that we pay special attention to and one thing that we think is out is kind of this vibe error that may not necessarily be enduring. So even if you are very upfront and direct and using the correct definitions of revenue and ARR, we still want to make sure that that is sticky, there's loyalty to the product, that the migration from your product is hard, so that when we're modeling out the second and third years, we can have confidence in those numbers.

13:29Yeah, just building off that. One thing that I think may muddy the waters a little bit is basically the sort of stages and names have changed a little bit. So like a Series A is kind of like what the Series B used to be in terms of check size and valuation. And the seed is what the A used to be. And it's not exactly apples to apples, but I would say like roughly that's what we've seen. And so you're going to have a shift in graduation rates because you've always had sort of a certain jumping off point. So I think that that's one thing that's worth so founders can maybe have a sigh of relief of like, maybe it's not so different than before.

14:04But what I do think is valid right now, which is these markets are so up for grabs and they're so greenfield and there's so much latent demand when you really hit something that the growth should be faster. because if you think about like at least the application layer in SaaS, you're almost always replacing something that came before. Maybe it was like on-prem to cloud early on, but then it was like cloud version one, cloud version two. And that has a certain adoption cycle, but here you just have like, there's not software in place. So if you hit demand, it should grow quickly. So the advice I've been giving to founders is just like, take early on, take your time, really focus on like deep product market fit, like focus on finding demand because you can always sell and find some customers, but really focus on where there's latent demand and let the customers pull you.

14:50And if they start to pull you, that's where you'll see this fast growth. Otherwise, you still are in the exploration of the PMF phase, whereas you could have like strong armed your way to like a millionaire before and probably got that next round done. It's probably not the right thing. You really want to wait and make sure you have something. And then that's probably the signal where it's the right time to raise. Yeah. You know, you observed the stage shifting earlier, And I'll tell you a little bit about what's behind it, because in addition to stage shifting, we're seeing round compression, right, where people are raising multiple rounds in a very short period of time.

15:26And, you know, you'll have your pre-seed and your seed and then your seed extension and then your pre-A. All that happens now within months. And the other thing that we're seeing is that the compression is happening so fast, we're now seeing round conflation, where you have a$30 million seed that's raised, right? In no world where that would be a real seed round, right? But what in reality is happening is that that's more like a seed and Series A, like as if you raise 5 million seed and a 25 million Series A all at once. So the expectations when you go out and raise your Series A, your real Series A, is that you would look more like a Series B company.

16:08So that's what's causing that stage shift. And, you know, some of the fastest growing companies are experiencing that where they do that 30 million seed and they're able to get to that Series A. But that's certainly, I think, the exception that we're seeing. But it's certainly a dynamic to be observing. So what are the real signals that you're looking for within the product, within the growth, within the revenue? Well, I think, you know, it's always the velocity of growth. I think Heyman got unfairly trolled for his comments that he made. And the reason is that if you look at what people said, it's like, well, that only applies to a certain, you know, the small percentage of companies that can actually achieve that.

16:55But that is the point of venture, to be fair. But I think when people, you know, when investors and founders think about velocity of growth, they fall under three different traps when they see these revenue ramps. So the three different traps that are most common, they can fall into the first category, which is what I call a cost wrapper. Everyone's heard of the model wrapper or GPT wrapper. This is more of like revenue that is actually a cost wrapper because your cost of delivering the good or service is really high. So obviously that's buffeting the revenue as opposed to margin actually buffeting the revenue.

17:37So that could really drive revenue up because your costs are really high. The second is the second sort of trap or category that these hyper growth companies can fall under is that there's a surge in demand or a pull forward of demand that's not necessarily durable. So we saw that during the pandemic with Zoom and others in e-com. And then the last category is more if there's some sort of regulatory shift or some incentive, you know, like renewable energy and had a bunch of incentives. And the real question is, can it still survive or show that acceleration without that regulatory support? And if you can run through that sort of that gamut of pitfalls, you can get to the last category, you know, category four, which is actual real innovation that's happening with lots of demand.

18:26And that's, you know, hopefully, you know, if you can sort of run the gauntlet and get to a category four, you're in that promised land. And there are companies that are like that. We saw that in the, you know, sort of original internet as well as now with AI. right astacia you brought up a really good point between real revenue and by revenue and there's like there's multi-dimensions to this as well because we're seeing like the time from zero to 100 million dollars become like a new benchmark and like this new racing ground but i'd love for you to maybe unpack that a little bit more and like maybe let people know like what are the types of companies that are growing that fast because we do have the ai labs we also have lovable like There's all these different kinds of categories within it, but just to set the expectations between those and then maybe more AI agents and other infrastructure companies that you guys are looking at, could you just describe that a little bit further?

19:18Yeah. So a pattern that we have seen is generally there's two buckets of companies that are some of the fastest growing of all time. This first bucket is a lot of the businesses that are actually selling in to the model labs who have incredible urgency and very large budgets to fulfill their needs. And that is best exemplified by these chain of thought companies from Mercore to Handshake and others where there's almost infinite budget for the products themselves. And with Mercore, this is one of many acts that they're excited to release to the public. The other bucket that we're seeing hyper growth in is in the prosumer market where AI has created a really game-changing, delightful experience that wasn't possible before.

20:12Usually it is the democratization of a skill that was more specialized to a broader audience. We saw that with Canva that lowered the barrier for design away from the Adobe suite to make anyone the designer and the billions of people that they could sell to. And now we're seeing it with software engineering. Anyone can be a software engineer now with the cursors and lovables of the world. So those are the two big buckets, often companies selling into the model businesses themselves and really delightful Gen AI native businesses that unlock a skill for the billions of people out there who are creators.

20:58And where do you see, maybe Patrick, I'm curious from your standpoint, where do you see us right now within this AI super cycle? What trends have already passed and what are you looking forward to get into next that people are overlooking? Yeah. Gosh, isn't that the ultimate question? Where are we at? Is this a bubble or is this not? I have like three text messages to people, my friends asking me this question. if I knew the answer there'd be a lot of things I'd be doing on my own personal bank account I mean my own personal investing account but the you know I still think like I mean the various layers right like the modern sort of data center power cooling all this stuff there's still some innovation to be done there over the next handful of years I believe right and at the infralayer I mean you're seeing obviously stuff around agents MCP some of these layers the thing that you have to be aware of there is this time around the cloud providers are like they're ahead like the the big hyperscalers there were the innovators in this market and so like during the cloud era they were behind and it was easy to innovate there and there was a lot of successful cases we had a bunch of investments there but at the infralayer i think they're ahead this time so you have to kind of like they're moving quickly so you have to be i think very aware of that it's the application layer where all these agents are impacting workflows.

22:22And there's just all, what I see is net new markets that I think are quite exciting for entrepreneurs. And there's a lot of them. I mean, I think there's like literally trillions of dollars up for grab in these markets. And I think it requires context. I think it requires understanding workflows, understanding buyers. And so I think there's a lot of room for innovation. Now the counter to that would be as these models get better and they do reinforcement learning on agentic workflows, like what will they take over? What will they own? And so that would be the counter. And there's certainly proof that they're going to go after things like coding and probably some other large categories.

22:59I sort of believe in a hybrid world personally. Like I think that it requires so much context and understanding of the end customer and how customers buy and kind of all of this stuff that it's going to be hard to go after this long list of use cases. So that would be the area that, if I was starting a new company, I would really focus on because I think it will be defensible. I think there'll be a lot of room for innovation and it's where we're doing a lot of our investments right now. I think that's a great point, Patrick, and something that we think a lot about, which is what is the integration universe that your product can bring to bear and leverage to have the context for the AI agents.

23:38If it is in more esoteric spaces, it's highly unlikely that the model companies or the cloud service providers will be operating there. And so your earned insights from working in the domain, building these technologies, and particularly the integrations which should be very hard if their legacy systems provides more of a technical moat. And if you're automating traditionally manual processes, it's such a sticky product with high ROI, the switching costs will be very high for the buyers as well. So I really agree with the points that you've made. One thing I'll just build off of as an observation I've been trying to think about, which is a lot of these agents, at least to date, have been...

24:26The last mile of building them is actually quite complex and hard, which is a good thing, by the way, if you're a builder, because I think that creates service area. But a lot of how that's being addressed is either through business model innovation, e.g. you don't provide a technology, you provide a service that's human in the loop or forward deployed engineers, which is this terminology for like basically parachuting an engineer into a customer to get that last 10%. And what that's meant basically, I think in terms of like ICP is you either can make this tractable by going after very large customers, which is kind of what like a Sierra or Palantir has done.

25:04And you have big contracts or you're very consumer oriented, longer tail type approach. But there's this like middle ground, which actually like SAS lived off of, which is like this 30 to 50 K contract. It was all hosted in the cloud. It was a 5K setup. You got going in a month. That market is actually like a little bit harder to go after. I think in this right now, by the way, I think it might become more tractable, But I think as entrepreneurs thinking about where to build, it's like being aware of not just like what you're building, but who the ICP is. And is it tractable to deliver this type of solution now versus in the future, I think becomes really important.

25:40Yeah, you're essentially talking about to deliver the service for the mid-market and how if you look at the acquisition costs of CAC and delivery against the LTVs, the numbers just don't really add up at this point. Yeah, I was going to ask, so how have business models majoritively shifted? SaaS era is one thing, but this new AI era is causing different kinds of forms and functions. Astacio, to your point, vibe, experimental, there's that cohort. But maybe, Tony, if you could answer this one, how are you seeing the shift in business models? I think the shifts in business models are both hitting the startups as well as VCs in general.

26:22I think AI is actually disrupting the VC model. And I see that in two ways. One is we're now seeing founders that are, they're not giving as much, you know, the household name, the tier one venture capital firms that have just been around forever. We're now seeing founders that are shifting to a different type of investor, ones that are a little bit more technical. So that was an interesting observation that started happening a couple of years ago. And because we're so early and we have a pretty wide portfolio, we see some of the meetings that founders are taking versus, you know, they're passing on.

27:06And it was surprising to us when we saw that shift happening. The second shift I think that's happening within venture is, you know, I think about like some of the techniques that model builders use in post-training. Like there's a technique called reinforcement learning. And traditionally, the criticism there is that it's very sample inefficient and it's also very costly. So you need, you know, a lot of compute. And now that we have a lot of compute, you know, RL tends to work really, really well. Now, if you were to put that into startup building, if the cost of building, the cost of capital is coming down and now you literally have models that are coding for you.

27:49We are almost seeing this shift to almost like an RL version of a startup building where you have a big rise in solo interpreters that are not raising money. And so you could almost think of them as just a bunch of founder agents going out there, trying things and seeing what works. Now, that wasn't possible before the coding agents. It definitely wasn't possible before cloud. But that's certainly a real trend that we've observed over the last few years. And it'll be interesting because I think the last few decades of venture has been more akin to a post-training technique called supervised learning or supervised fine tuning, where we as venture investors, like, oh, yeah, you know, we've operated and scaled companies before.

28:36Here's how you would do it. That's sort of like supervised learning. And then you go out and raise from downstream capital. And then your next, you know, like these downstream investors will say, hey, we like this company, but we don't like that company. And what does that sound like? That sounds like RLHF, reinforcement learning with human feedback. So our entire venture industry has almost been built. All the playbooks and the best learnings and practices have been built through that. But I'm really excited to see this whole RL surge of founders going out and trying things that may be different.

29:13And we're going to learn from those founders as an industry. Yeah, I mean, this is a, it is an interesting, it's one I think a lot about, because we think about like the construction of on these early stages, what creates a great company that has the best chance of winning. If you look historically, you will find things that like companies that have multiple founders are actually like statistically way more likely to build an outlier company. If you look at things like founders here in the Bay Area, although people can argue a lot about geography, great companies can be built anywhere and they are built anywhere, but statistically they're way more likely to succeed here.