David George on the State of AI Markets

Let me just start with what I think the big takeaways are from this piece because this is the first time we've ever done this style piece. We produce so much work and so much analysis inside of our team and we thought we have so many different thoughts and points of view, why don't we put them on paper and share them out with the world? So that was the genesis of this.

My big takeaways from doing this one, AI demand side is crazy. The actual uptake growth quality of companies in AI is extremely encouraging from our standpoint. Companies are starting to run themselves better. I'm going to show you some stats on that. There's been some buzz including this morning, kind of debating what's going on there.

But this crop of companies I would say is more impressive than prior crops of companies partially because the demand for their products is so high. That's demand side. Supply side is healthy right now, but we are starting to see some signs of things that are stretched a little bit. I'll talk about what we see and what we're looking out for.

We've been fortunate to be a part of a lot of these great companies. The most exciting action that is happening in the private markets, it's AI and it's happening in the private markets. We're going to show some slides about that.

Lastly, my big conclusion, what has me so excited about where we are now is just how early we are in this product cycle. Product cycles drive our business. These are 10, 15 year cycles and we're just at the very beginning of it right now.

So, let's dive in. We invest across all private stages. This is a chart that just shows our activity. We're very busy. It's across all verticals. We on the growth side have been most active in AI and infer apps and then in AD, but also very active in our other verticals as well. And I'm going to zoom through some of these. I hate to do the A16Z commercial, but I really like this slide.

I think we have the chance to work with some of the best models and apps and infra companies. Obviously, I'm going to do that on that side. Gong it. I do like that slide a lot. Soundboard effect here. I'm happy. We debated how early to put that slide on the deck and I said put it further back and I was overruled. So thankfully anyway, here's some data.

So we collect tons and tons of data as a growth team because we're basically seeing every growth stage company in the market as a either portfolio company or as a prospect. And so we have a great data analysis team. We did some data analysis. I think this stuff is just super interesting. We geek out on it.

To me, the big conclusion from this is 2025 was a year for accelerated revenue growth. Revenue obviously slowed in 2022, 23, 24 following the rate hikes and the pullback in some of the tech stuff. But 2025 reversed that trend and it accelerated across different types of companies as we rank them by decile and cortile, but especially among the outlier companies it really accelerated.

You've probably seen us put this slide on a page before but the fastest growing AI companies are reaching 100 million bucks of revenue significantly faster than the fastest growing SAS companies in their era. And there's a really important thing I want to call out about why that is the case and that is because end customer demand is so strong and the products are so compelling.

It's not because they spend more money on sales and marketing. It's actually the opposite. The best AI companies that are growing the fastest are not the ones spending the most amount of money on sales and marketing and they're spending less money on sales and marketing than their SAS counterparts and yet they're growing much much faster.

So this was a slide showing just the growth of the AI companies versus the nonAI companies. Roughly speaking, the AI companies are growing two and a half times plus faster than the non-AI companies. And that shouldn't be a huge surprise. The best of the AI companies are growing very very fast.

We had to triple check this data when we saw the AI top performers growing 693% year-over-year, but it matches up our experience and anecdotes that we see from the portfolio companies. So that's growth.

This is the margin profile that we're seeing in the data set and again these are internal data sets that we have of portfolio companies and companies that we look at as potential investments. Gross margins are a little bit worse for AI companies. You've probably heard us talk about this before, but in a way we feel like low gross margins for AI companies are sort of a badge of honor in the sense that we want to see if low gross margins are a result of high Inference costs.

One that means people are using AI features and two, we have a belief that those Inference costs over time are going to come down. So in an odd way, if we see an AI pitch and the gross margins are super high, we're a little bit skeptical because that may mean that the AI features are not actually what is being bought or used by the customers.

We're going to talk about ARR per FTE, but this is a new thing that we've started focusing on and this is one of the things that got a lot of pickup and discussion on X in the last few days. ARR per FTE is sort of a measure of the efficiency of how you run your company in general. So it encapsulates all of your costs.

It encapsulates, not just your sales and marketing, which is an efficiency measure that we've always kind of looked at when we do analysis in the past, but it also captures your overhead. It captures your R&D. And so for the best AI companies, they're running at like 500,000 to a million dollars per FTE.

And the rule of thumb for previous software businesses in the SAS era was like $400,000 in the last generation. Again, I'm going to talk about this a little bit more, but the reason why this is the case is mostly because demand is very very strong for their products. You know, and so they need a less resource to go take it to market.

David, maybe a quick clarifying just before we go to this slide here. So, how do we how do you define AI companies? Is that defined as postjack GBT versus historical AI ML companies founded by a certain time period?

Yeah. Yeah, it's sort of post postg and some of them have were founded like right around that time. We'd give a little bit of grace but if they're their first product in market was an AI you know native product then that's how we define it.

Got it. And then maybe this is a good point but where you can punt till later but like one of the questions I think a lot of folks are trying to understand is the magnitude of change and expected revenue and growth from companies from the SAS era to AI era companies and you've talked a little bit about the magnitude of revenue etc but what happens to those that are not AI native will they have a hard time competing against AI native companies are they all shifting will we see more fallout how should people be thinking about their historical portfolio?

Yeah. So, the way that we're approaching this with our portfolio is, you need to adapt to the AI era or die. And so that's both on the front end and the back end.

So on the front end, you need to think about how you can incorporate AI into your product natively and not just, attach a chatbot app into your existing workflow, but reimagine what it can mean with AI and be aggressive about disrupting yourself and changing.

And then on the back end, I shared some of the stats around the efficiency that the companies are running at. This is going to change too. And so you need to be fully rolled out with the latest coding models for all of your developers and all of the latest tools across every different function inside your organization.

The biggest uptake has been in coding so far and that's where we've seen the biggest leaps. There have been major major changes like in the last two months on this like month and a half in this. Andre Carpathy has written about this. I was on a catchup with one of our, sort of pre-AII companies.

This is a founder who's very AI, like he's very AI deep and so he's adapting his company. We were talking this week and he told me that he was frustrated with one of their products and so he just took two engineers that are very deep in AI and assigned them to build it from scratch with cloud code and codeex and cursor and just they had unlimited budget on coding tools.

He said he thinks it's going somewhere between 10 and 20x faster than progress that they had before. And the bills that they have associated with that is actually they're high enough that it will cause him to rethink what his entire organization will look like.

The conclusion was basically I need my entire product and engineering organization working this way and I think it's going to happen within the next 12 months. But what does that mean for what the team design actually is and where does product start and where does start you know and even where does design start in that process.

So it feels like December was sort of a turning point on code and the next 12 months it's going to kind of hit it's it's either going to hit and take hold in companies or those companies I think are going to be moving much slower than their peers.

So, as it relates to the preAI companies, adapt, we have we have another example of a company that is a pre-ai software company and the CEO has gotten totally AI pill and he's like, we're going to become an AI product. Like, we're going to ship, you know, your employees are now your AI Agents. How many Agents do you have? Like those are the things that he's talking about.

We have another one that was very extreme about it and he said I now ask the question for every task that we now need to complete can I do it with electricity or do I need to do it with blood like this is like the extreme mindset shift that's happening with our companies and so I'm happy to see that our PAI companies are moving very fast and trying to adapt but they very much need to adapt to this new era both front end product wise and back end how they run their companies.

Totally. Yeah. Maybe tactically almost every portfolio you have to go line by line on the company to understand where the founder is on that journey and how much they are implementing from the ground up and and you know what you said in terms of blowing up existing operations. That's also happening in post AI companies too and and increasingly people are just looking every six months.

It's like the things we built six months ago could be vastly improved by based on what is available today. So that if that rate is continually happening, the preAI companies are needing to to increasingly 10x catch up to that point.

Yeah. The good news for the prei companies is the business model evolution is still early days. So the most disruptive thing that can happen to you is a technology and product shift and also a business model shift at the same time.

There's really one I I think of the business models as like a spectrum and I'm talking about like enterprise like B2B just to keep it simple but the spectrum is basically licenses and this was like the pre-SAS you know license and maintenance business models then you had SAS and subscription and that was typically seatbased and that was a big innovation and it was very disruptive like the architecture and cloud delivery was disruptive but the business model change was very disruptive like just go look at what happened to Adobe as they went through that transition.

Then you have this transition to consumption based so usage based and this is how the clouds charge and so many of the sort of volume based like taskbased type businesses have already adapted that and shifted to that from you know seat based to consumption.

And then the next iteration will be outcome based. So, when you when you do a task, you know, and ideally when you successfully complete a task, you get paid based on the successful completion of that task.

The only area where that's really possible today to pull off is is probably customer support, customer success, because you can kind of objectively measure the resolution of of something. But we'll see what happens with the capabilities of the models to the extent that other functions besides customer support can measure those kinds of outcomes that would be a huge disruptive force for incumbents and and honestly seats to consumption might be a big disruption if the composition of companies changes as well.

But that next one is the is the really big one for sure. Speaking of blood versus electricity we should go to AR over FTE. This next slide here.

Yeah. Yeah. So the big the big debate that was going on on this one on the next slide was like oh my gosh look at the AI efficiency gains that are happening in the market. Now there's a little bit of that in this like companies running themselves a little bit differently and you take the example that I gave about you know the two engineers who are rebuilding the product like sure I would say my observation from our companies even the AI native ones is they run leaner partially because they've just grown so quickly and the demand is so strong.

I wouldn't say yet we're at the point where companies have fully reimagined the way they run themselves. I think this is a little bit the result of our data set being the best of the best companies and demand signals for those being extremely high. So they you know they have less resources to serve that demand and frankly you know efficient general efficiency gains that have happened in the technology market you know out of the kind of 2021 most bloated era.

So, we're starting to see some early signs of that efficiency, but the wholesale run your company totally differently. I think, we're kind of early in that in that journey. I'd say the coolest one that I've seen is in the in the public markets that anyone can go read about is probably Shopify where they, Toby's awesome. Like he's a CEO that's that's close. He's in a bunch of our groups and stuff.

He does a great job and he fully embraced this a couple years ago. And then there one of our staff writers actually wrote this whole big deep dive on how Shopify AI itself you know in terms of you know employee direction process etc. And that's just probably scratching the surface of what's going to happen over the next 5 years.

A good seg to the next section on what are these companies actually doing in our favorite topic which is lawyers have only increased in this new world of AI's meeting lawyers not the opposite I I love the tweet I don't know if you saw it earlier this week that a corporate lawyer was quoted saying LLM have actually increased my workload because every client thinks they're a lawyer now it's a good seg to Harvey which is the next slide that's that's very good that's very good.

Harvey's so great I so okay this is a real test for me because you know I love talking about our portfolio companies and I'm supposed to go through this section quickly because you know I think people people know these companies hopefully the takeaway on this one you know one of the big things that we look for and one of the questions I think that came in was how do you know that revenue is going to be sustainable like these companies they all grew really really fast but is it fleeting and the big thing that we push ourselves to do is make sure we go super super deep on revenue venue retention, renewals, and product engagement, actually time spent, how often are people logging into the platform, when they're in the platform, what does their activity look like?

What you see on this page is with the onset of much better product that they've built over the last couple of years, plus the improvement of reasoning models, it turns out lawyering and reasoning go go hand in hand. Users are spending about double the amount in the product as they had before. So it turns out that AI is is really good at lawyering.

Again there's not fewer lawyers. But I think AI is very very good at this and I think lawyers are getting a lot more efficient. The most important thing as it relates to Harvey is they're just spending a lot of time in the product and getting a lot of value out of it which is great. Let's go to a bridge. Oh unless you want to keep talking about lawyer.

Oh, I was just going to make a comment. In all the seven years that I've known you, I wouldn't have ever discerned that you're from Kentucky other than this moment now. By the way, you say lawyer. That was a tell. My there's a there's a couple of those words in my vocabulary. I can't I that I I don't you know, my my wife always jokes. She's like, you know, you go home, you have like one bourbon, and then you you talk like you probably did when you were 18. the Kentucky came out when it came to lawyers. It's it's it's 10:25 a.m. I have not had any bourbons today. So, important distinction. It's important distinctions. Yes, exactly.

So, a bridge a bridge is another one that's super super exciting. I mean, this is like the doctors rave about getting to to have access to a bridge and how much time it saves them and how much, you know, better it makes their lives. So you know one of the customers that we talked to described it like a trusted deputy.

The chart on the right shows something we look for, which is the blue line shows the growth in users and the green line shows the engagement of those users. And so as they have massively grown the number of users, you'd be a little worried if engagement of those incremental users that they were adding was going down, but instead they have extremely high usage among the people who use the product and that has actually held steady and grown a little bit even as they've added tons and tons of more users.

So the these are just examples of the kind of data that we look for to make sure that we feel confident that the revenue these companies are generating is sustainable and again these companies are growing faster than you know any of the predecessor companies but but it's very sustainable. It's, you know, it's high engagement, it's high retention. And that's critically important for us.

Same thing with 11 Labs. Voice is the centerpiece of so many of the new AI tools. I talked about customer support on the B2B side, but, you know, so much, you know, other personal tools, business tools, you know, start start with voice. The usage growth is the thing that I love to look at on this chart. It's just staggering. And this company is growing very fast and is a great example of one of these companies that runs extremely efficiently. So 11 Labs is is really is really a great one.

Non is the next one. So this is another this is a different example. So this is actually a good example of what I was describing earlier. So they were early to this, you know, AI shift and and they spent a lot of effort making sure that they could take the most of the AI capabilities and make their business better.

The biggest way you can see it in their business today is in the handling of resolutions. So part part of what they have is, you know, Agents that have to handle travel bookings or travel changes. AI is now handling 50% of those user interactions. And this is hard stuff like this is travel bookings. This is changes to travel.

So this is not you know complex like tell me the balance of my bank. You know this is like complex workflow that that AI is now able to handle. The way you see that in the business is a 20 percentage point expansion of gross margins over the last 3 years. And that's just exceptional impact.

So you need to adapt or die. Well their competitors are not adapting. They're very old school and while you know they've been sitting still and and doing things the old way, Non now has 20 percentage point higher gross margins than those incumbents.

Then you know Flock Flock is doing absolutely incredible work. I've talked about them so much. It's it's the most compelling customer value proposition that we see in our portfolio because what their ROI is is solving crime. The 10% stat we've covered before. Each year Flock is solving 700,000 crimes.

The the the data point on the right also is a data point that just shows per officer that where there's flock they're clearing almost 10% you know more crimes. So huge impact on the community. Obviously they have a great you know they have a great business and financial model that goes along with it. But the but the impact on their product or from their product is is exceptional.

Okay. By the way, I don't know if you see the chat lighting up of people saying that they're three bourbons deep. Oh, I didn't see it. For what it's worth, there is one question about how do you think about the the benchmark? Like if you were to think about traditional industries like finance for example and using JP Morgan as a benchmark, what would you calibrate the Fortune 500 in terms of AI adoption? And then maybe I'll overlay that that question that that Xavier mentioned as well with you know there was that study about enterprise adoption from MIT at the early outset of last year and they were measuring all sorts of wonky things. Maybe say a little bit more about how and what you're hearing from Fortune 500 CEOs.

What we're hearing from Fortune 500 CEOs I would say is and maybe this is the key sort of link between those two points. What we're hearing from Fortune 500 CEOs is we have to adapt. We're dying to understand what AI tools we need. We're ready to change. We, you know, our businesses are going to fully roll things out and, you know, we're we're ready. We're going to become AI companies.

That's quite different than what is actually happening. And I think the biggest disconnect of sort of, you know, that mindset compared to actual change in the businesses is just change management is hard. It's hard enough to get people to just use an AI assistant to help them do their jobs better.

Coding is probably the easiest one to get people's minds wrapped around. Customer support. It's such a better, faster, cheaper, obvious thing. But in terms of actually, you know, general management of businesses, changing business processes, change management, it's extremely hard to do.

So I'm not surprised that there are anecdotes out there that suggest, oh, you know, things are moving slower than expected, but for the best companies that are fully embracing it and actually know what to do, it has tremendous business impact already. So I think there's going to be a sort of reckoning over the next five years of who can actually embrace change push through change management you know adopt all the best products and those that don't and I think there'll be major differences in productivity you know we have some charts later in the slides you know which I can talk to but you know the expectations around productivity enhancements and you know and growth and all that stuff you know the expectations are high and I think a bunch of companies will achieve those and the ones that don't are going to be at a huge disadvantage.

Chime said they reduced their support costs by 60%. Rocket Mortgage said that they saved 1.1 million hours in underwriting, up 6x year-over-year, and that was 40 million bucks of run rate annual savings. So we're seeing pockets of it in nonAI businesses and I think this is going to be a really interesting year to watch over the next 12 months.

I think you're going to see a ton more anecdotes, but there will be companies that can figure it out and there are going to be companies that don't. Totally. And also, they've a lot of these corporations have had to orient their business to be ready for AI as well. Like there's one version of just like using a chatbot, right? And how much productivity gained that actually gets you? Probably not a lot, right? But if you have to actually completely upend your systems information and backend to be ready for AI, a lot of that is probably latent and and being built up now into actually seeing the outcomes associated with it.

AI winners are driving the public markets. They account for almost 80% of the S&P 500's return. So this is sort of the major thing driving the economy and the stock market. Public markets are doing very well, but the fundamentals are sound. So the prices are going up or you know there's some blips like the last couple of days but they're generally doing well, but the fundamentals are very sound.

I would say the evidence of froth is minimal. So recent performance is driven by EPS growth. Multiples have contracted slightly maybe more than slightly if you're a SAS company over the last few days or a couple weeks, but I would say the market is priced on in general earnings earnings and earnings growth.

So the earnings multiples are higher than average but nowhere near the dot. And so you can just look at the charts and see where we are and you know that that gives me some comfort. And again the earnings of the companies that are the biggest drivers of the market in general I feel like are pretty sound. The companies are good.

So, you know, the the health of these companies, I would say, is pretty good and and the valuations are higher than average in the past, but they don't feel super alarming. I often say the leading tech companies that I was I was just talking about are the best businesses in the history of the world.

If you just look over a long period of time, they have shown margin improvement that suggests that is probably true. And that's, you know, that's on the left side of the page. So, investors are paying for profits, not lossmaking growth, and that's a big contrast from 2122 era, sort of 21 era, and obviously a big contrast from a dot adjusted for margins.

Multiples are are not that high. And so again, I like summarize, you know, five slides worth of materials. The market's higher than it has been in the past, but I think, you know, there's high expectations for a reason and and we're optimistic about the impact of AI flowing through to earnings, you know, overall in the public markets in the coming years.

Maybe I'd focus your attention on the right side, which is if you just took a fourbox of like low growth, high growth, low margin, high margin, and paired up those types of companies. This is a chart that shows how they trade. There's a premium for the best companies.

What you see on the the two columns on the right is high growth, high margin companies and then high growth and low margin companies. Your bad box is obviously low growth, low margin. And those companies shouldn't be rewarded. They they they should trade low. But the companies that are high growth and high margin and you know the high growth and low margin, as long as they have good unit economics and they're scaling into their margins, they should be rewarded.

So I think this is good. If you're not high growth, even if you're high margin, it's tough out there. And that's not surprising. Again, I've talked about this in the past in many different forms. But ultimately, growth is the biggest thing that drives returns over 5 to 10 years. And so, it's nice for me to see high growth is rewarded more than low growth. But if you have high growth and high margin, you're one of those great businesses, it's being very rewarded.

This is just like we're going to talk about supply side of the capex buildout. So the buildout's massive, the size and the concentration of the investment is inherently risky just given how big it is. While it has some bubbly features, the underlying fundamentals I would say bear little resemblance to previous bubbles.

The investment is financed primarily by historically profitable companies like very profitable companies that I had talked about. Debt has started to enter the picture. Cycle times have accelerated which is good but you know model we're closely monitoring the sort of cost of training and the economics of that whole equation right now it seems pretty good the paybacks for the big model companies that spend money on training models is pretty good but we're monitoring that closely most importantly we think that AI is going to be you know the biggest model buster that I've seen in my career certainly I've written about model busters so I won't spend too much time on them but they're companies that grow faster and longer than anyone would have would have modeled in any scenario.

Like iPhone is the classic case of this. If you if you take consensus models from pre iPhone to 5 years later, four years later, consensus models were off for Apple's performance by a factor of 3x over four years. And this is like the most covered company in the world at the time. So, I think that the same thing is going to happen in many pockets of AI where the performance just massively exceeds, you know, what any expectations in a spreadsheet would would show you.

So, tech in general is itself a model buster, but since 2010, tech has delivered high margin revenue at unprecedented speed and scale. So it often looks expensive early, but repeatedly surprises to the upside, I would say, and creates value, I would say, far in excess of the capital that's required to grow. And I I have no reason to think it'll be different, you know, this time around.

So relative to the dot, capex is actually supported by cash flows, and capex as a percentage of revenue is considerably lower. So that's simple headline. We can zoom to the next slide but you know I feel much better about this capex you know dynamic than than do obviously hyperscalers are the ones who are bearing the biggest brunt of the capex and this is a very good thing you know for our portfolio companies this is great like I am all for it get you know get as much capacity in the ground get as much supply as you as you possibly can on the ground for training and inference this is a very good thing and Again, the companies that are bearing most of the brunt of this are the best businesses of all time that I had talked about before.

One thing that we're starting to monitor is the introduction of debt into the equation. So, you can't finance all of the forecast capex that's to come with cash flow, and we're starting to see some debt. So, we're following this closely. We're generally not invested heavily in companies with exposure to debt.

Do I feel comfortable with a bunch of the companies on the page financing with cash flow, continuing to produce cash flow, and using debt even, you know, Meta, Microsoft, AWS, Nvidia as counterparties? Of course, I I feel great about that. I mentioned the ones I feel great about. I don't feel great about all of them. So, not all counterparties are the same.

We're starting to see Private Credit get a little bit more involved in the data center buildout. And you know again the company that's very well covered that is kind of making a bet the company move into becoming a cloud is is Oracle and they've you know they've been profitable forever and reducing their shares forever. But the amount of capital that they are committing is very large. It's a big bet. They're going to go cash flow negative for many years to come.

If you follow some of the buzz around it, like the the cost of their credit default swaps has gone up to like 2% over the last three months. And so, we're watching stuff like this. Again, this is all generally good stuff for our portfolio companies, but we want to make sure that the market overall is healthy as well.

This is just a slide that shows the magnitude of the pace of change of AI. So, comparing AI buildout and AI revenue to what happened with Azure. So the AI revenue is coming along relative to the cloud. It took Azure 7 years to reach one year of AI revenue. So this this is just Microsoft reporting data which I think is a a cool way to to frame how quickly this has happened.

The build's taken a very long time. Again this this AI buildout is happening much faster. But it took 10 years for Azure revenue to surpass their capex. And I think it's I think that sort of ratio or equation is going to happen much faster with AI.

We don't need to geek out too much on depreciation, but this is one of the topics that gets a lot of buzz in finance circles, you know, just what are your assumptions around depreciation of chips in particular. I would say the pricing for older GPUs is very solid.

Early users stick with models a bit longer, but later users quickly switch to the new thing. So, that's the right side. That's like kind of the model side. On the chip side, 7 to 8y old TPUs, Google actually disclosed this, 7 to 8y old TPUs actually have 100% utilization.

We very closely monitor the price of chips in the secondary market and the price to rent A100s and H100s has actually held up very very well. So older generations of chips are still still getting fully utilized. So this is not something I worry about yet but it gets a lot of buzz and you know sort of alarmists who like to to talk about risk in the system.

All right some positive stuff. So the big thing that we talk about all the time is is this paradox right like as tokens get cheaper consumption goes up. All the hyperscalers report demand is well in excess of supply. I believe them when they say that.

I interviewed Gavin Baker, friend of mine on our at our AI summit, and he was comparing the buildout of the internet and and laying all the fiber to the buildout of data centers here. His big line was there is no dark GPU. There are no dark GPUs. There was a dark fiber. You had to lay fiber and then, you know, it laid there dark and it wasn't used. If you put a GPU in the system in a data center, it gets fully utilized immediately.

So that's a very good sign, you know, in terms of, you know, demand meeting supply immediately. I mentioned this earlier, earnest growth should come for these companies like this is our expectation. And if it doesn't, then they will probably be disrupted if they can't change. So change management again is the biggest reason why we see things you know that that haven't sort of dramatically shifted yet.

It's honestly to to me it's not the readiness of the technology itself. It's probably you know product buildout that needs to get built around the technologies and then change management and and putting it in production. So revenue growth has scaled at a staggering clip relative to other categories.

This just shows how quickly generative AI in app revenue has grown from 23 where it was basically you can barely