So, thank you very much for inviting me. My name is Martin and I run the infrastructure fund at Andreessen Horowitz. In this talk, I'm going to discuss how we see GenAI's broad-ranging impact on the software industry from the perspective of early-stage venture capital.
AI has been around since the beginning of compute. We've seen a number of AI hype cycles. We've seen a number of AI winters. And the one thing that's been consistent is that until recently, we've never seen the creation of generational AI companies. AI has remained largely in the domain as a tool primarily used by large companies. So to understand the current wave, it's important to understand why it hasn't worked before. So let's start there.
So the answer is actually quite simple. Historically, there's never been a strong business case for AI. There have been multiple issues. For example, many of the markets it tackles have not been large. Take for example game playing or chess playing. There's no market there. There's also been issues where while games have been good, like handwriting detection, they've not been significantly better than say a human being. And so it's hard to adopt it as new technology. And in many cases, AI requires large upfront investment before you can deploy it, and this is very hard for startup companies to do.
So let me give you a specific example. I did my PhD at Stanford in 2003, and when I joined there was a professor named Sebastian Thrun, and in 2006 he ended up winning the DARPA Grand Challenge where they drove 212 kilometers off-road fully autonomously. And when that happened, we were so excited because we thought self-driving cars are here. In two years, we will all be in robo-taxis. Unfortunately, as we all know, it took a lot more investment and a lot more time before this happened. And that caused a lot of startups to try and do autonomous vehicles. And by now all of them have either been acquired or were killed.
So robo-taxis will happen. But a startup could never wait 20 years and invest a hundred billion dollars, which is what we had to do. So this is just an example of one of the many cases where AI is just not suitable historically for startup companies.
But now we're seeing a bunch of generational AI companies, right? OpenAI, Anthropic, and Cursor. And so if we've never been able to create these AI companies in the past, why is it that they're working now? And the answer again is very simple. For the first time, there's huge demand for the product and the business case really works. Actually, as I'm going to show, this wave has similar economics to what we saw with the microchip and then the internet.
So before I go and try and explain the economics, let me just be very clear: when I'm referring to this wave of GenAI, I am referring to having a generative model that takes in an input and creates an output that is useful. That could be text, that could be an image, it could be the answer to a question. What I'm not referring to are older versions of AI like predictive models or computer vision or detection. This is strictly generative AI.
So let me give a simple example of why the economics are so strong for generative AI. This is going to seem silly, but I will make it more complex later. So let's imagine I, Martin, want a picture of myself as a Pixar character. Traditionally, to do that, I would go to an artist and they would charge me say $100 and it would take an hour. It'd probably take days, but let's say an hour.
Let's compare that to actually using generative AI. If you use a text-to-image model that was conditioned on my likeness, the inference cost for that is about 1/100th of a penny and it takes about a second. So if you look at the economic benefit, it is 10,000 times cheaper and 1,000 times faster. This is a massive economic shift.
Now, I know that looks like a silly example, so let me do one that's kind of closer to how this is being used in the industry. Let's imagine I wanted my lawyer to read a legal document for me. My lawyer costs about $700 an hour, tries to get back to me within 24 hours. So if I needed this being done, I would contact my lawyer and it would cost $700.
Now, if I instead used ChatGPT deep research, again, it would take a fraction of a penny and a few seconds. And we're actually at the point that if you do use AI for deep research, you end up with them citing the responses so you can double-check that it's okay. So again, we're in a case where it's 10,000 times cheaper and a thousand times faster.
So these are just two examples, but as you can see, GenAI is dramatically cheaper and faster. And these gains are in many areas. It's not just legal. It's not just images. It's for the creation of prose, the creation of software, the creation of music, the creation of video, etc. And if we look historically, the biggest, most disruptive waves in computation had this similar economic feel. So let me describe them.
So prior to computers, prior to the microchip, we would compute numbers by hand. That's where the word computer comes from. And then we created the stored-program computer ENIAC, and it was about 5,000 times faster than a human. So that brought the marginal cost of compute to zero. Now everybody could have compute. You didn't have to hire someone to do it. And this brought on the microchip wave. Like we're still riding the digital revolution. IBM, HP, DEC, Dell, all came out of this. Fujitsu.
The internet was similarly disruptive. So prior to the internet, in order to send information across the world that was complex, you'd have to use a ship or a boat or a courier. And then with the internet, it was instantaneous and the price per bit converged on zero. So the internet brought the marginal cost of distribution to zero, and this created the information age. Right? Out of this came Amazon and Google and Meta and Cisco and so forth. The economic growth driver of the last 30 years came from this.
So like I said, the microchip brought the marginal cost of compute to zero. The internet brought the marginal cost of distribution to zero. And now AI is bringing the marginal cost of creation and language reasoning to zero. And this has resulted in the most historic growth we have seen in the history of the industry.
So for example, I sit on a board of a company called Cursor. Their first real year in market, they went from zero to a 100 million run rate. The following year they went to over a billion. We've never seen anything like this before. If you look at the growth of the large labs like OpenAI and Anthropic, they look more like the public clouds which were heavily subsidized than normal software companies. So this is like truly historic traction.
It's also clear that this is not just all of software growing up. This isn't a macro movement in software. This really is about AI companies. So how do we know that? The data shows that AI companies are outpacing non-AI companies in revenue growth by a large margin. So if you look at the graph to your right, the outliers are massively bigger. So a top-decile growth in AI is over three times faster than non-AI. Three times.
It's very tempting to say, well, there's lots of growth, but maybe we don't have good margins like we saw in the share economy. Maybe we're just using money to get growth. But if you look at the numbers, that's not really true. In fact, non-AI gross margins are a little higher, but only modestly.
So another way to look at this problem is the amount of ARR, recurring revenue, per full-time employee. And again, the amount of revenue per full-time employee is quite similar between AI and non-AI companies, but if you look at the leader AI companies, they have much, much higher ARR per FTE. So in short, AI companies have tremendous growth, far faster than growth in traditional non-AI companies, but the business metrics are strong, including margin and especially ARR per FTE.
So this is also being reflected in the public markets. It's hard to overstate how fast, how furious, and what significant of a disruption this is happening. I mean, right now, AI stocks in the S&P accounted for about 78% of full returns for last year. 78% were AI stocks. So, I would argue this is the largest creation of new budget and the largest shift in existing budget we've seen in recent memory.
Let me share a few observations we've learned about the adoption of AI in the enterprise. So the first one, it's a secular phenomenon similar to the internet. So what does that mean? It means individuals will use it before the organization does. So keep in mind when you hear reports like 90% of deployment of AI in the enterprise fails, right now AI is not being used centrally nearly as much as it's being used individually. So for example, an organization may not use it as part of their platform team, but all of the individuals are using it for coding or they use ChatGPT to brainstorm or something like that.
Second, and this one's so important, many people don't realize that we've already hit a point where AI is as good or better than top professionals. This is for coding. This has actually happened just in the last few months, as well as writing prose.
So an anecdote for you. I was at a dinner with some top journalists from the top magazine in the world, which all of you would know, and the journalist said, okay, so I wanted to see how good AI was, and again, this is a very famous top journal. So I created two stories. I made the story, and then I asked AI to make the story, and I submitted that story to the editors and said which one was AI. And the editors didn't know; they got it wrong. So when you have the top journalists in the world, you can't tell the difference between them and AI. We know that we're in a very different technology wave with very different assumptions.
However, this last point is important. In almost every case we see that is economically useful, AI enables humans. It does not replace them. The reality is this journalist knew the story that needed to be written and had done the research and had told the AI to write it. Right? So whether you're doing code, you're doing art, you're writing stories, you normally have human beings that are driving these things. We have not seen massive displacement in jobs. We have not seen AI operate independently or autonomously at this point. So our assumption is this is a key enabler for individuals to improve productivity, not a replacement.
So, even though a lot of individuals use AI, you are seeing adoption in the enterprise. I don't think we've seen a technology this general in maybe forever, maybe since the beginning of software. So, these are all solvent enterprise use cases that are good businesses. Everything from software to legal. I'm not going to go through them specifically, but I will show some anecdotes.
So, let me just touch on a few areas. The first one, Klarna is a large fintech company, and announced earlier this year that they're handling two-thirds of all customer service chats via GenAI. So, this creates tons of savings, but it also increases the throughput of a single customer service representative. In general, when we see the deployment of AI in support, we see deflections, meaning the amount that AI can serve, go from 30% before to 80%. So today, AI is servicing 80% of customer support for AI-enabled businesses.
Amazon in the recent earnings call revealed that they did a porting of a very large legacy codebase, you know, saved 4,500 developer years, right, which is almost $300 million. So the point is less that AI can port a codebase. What's interesting is you don't want your developers porting codebases. There's not a lot of value to the business to port a legacy codebase. You want your developers working on new things. So this is freeing up developers for new development.
And of course Walmart also talked about how they did a massive data catalog creation with AI. Again, it was a massive project. It required 100 times less headcount. So these are just a small example that are public from earnings calls that show there is strong adoption, but the gains are very material and very, very real.
So up to now I've been discussing the macro disruption of AI to the software industry, and now I want to focus on software engineering. And the reason is software engineering is the largest vertical for AI in the enterprise. So 46 to 60% of enterprise spending on AI is for software engineering today.
Historically it was software that disrupted other industries, right? Software disrupted transportation, lodging, it disrupted the back office, and software itself has remained largely undisrupted. In fact, you can argue that software engineering hasn't changed since compilers were introduced in the 1970s. And at that time there wasn't enough of an industry to disrupt.
Yet here we are, and we've all seen the headlines. We're approaching a majority of code being written by AI in large tech companies. A staggering 92% of US-based developers are already using AI coding tools. 92%, and this is in legacy companies as well as new companies in the United States. And just this morning I read that if you look at GitHub, the repository, 4% of all commits are coming from just one AI tool, Claude Code. So I would say AI development is likely the fastest-adopted technology in the history of software engineering for tooling.
So unfortunately at this point it's not a stretch to say that coding as we know it is dying and will be dead soon. And for me this is very sad because I've spent my entire life programming. I've been programming for 30 years. I did my PhD in computer science. My company was a software company. In fact, today I work with many startups where the engineers actually no longer look at code, and I work with some of the most technical startups in Silicon Valley.
And this isn't just some narrow phenomenon. Some of the most technical, most famous people in the industry are sounding the alarm. Like Ryan Dahl, for example, this is the creator of Node.js. It's one of the most important, popular open-source projects ever. Basically said, we're done programming.
So where does that leave us as an industry? What happens to software development? Let's explore that. So to do that, I want to dig into the numbers of what this means to improving the performance of software. So how much productivity are we getting from AI? Now the numbers I'm showing here are largely from my experience. This is not the result of some research. But I've asked and spoken to many people and they're relatively consistent.
So, if I invest in a new team of senior developers and they use AI, it makes them super fast. And let's imagine they're going from just starting to the first prototype. They will be 10 times faster, guaranteed. And the reason is when you're starting a company and you're starting a new software program, you're building a lot of stuff that's been built before.
Now, if you take that startup and you wait for say 2 years and you check back in, now there's 200 people, 50 engineers, they've got customers, they've got a deployed system, the software is much larger. That 10 times improvement reduces to about 30 to 50% improvement. So you've lost the ability to go fast. And then if you look at very large legacy companies with legacy codebases that have millions or billions of customers, that improvement drops again from 20 to 30%.
So it's very clear that AI is not a silver bullet. The gains are not infinite and they diminish over time. So it's worth asking the question: why is that? And it turns out that writing code is only a small fraction of software engineering.
And if you look at what people spend their time in, it tends to be other parts of the software engineering process. So I'll just use two examples. One of them is a lot of software engineering is just understanding what the customer needs and then writing it down. I think the average change to a mature codebase is only 10 lines of code. And that 10 lines is codifying some learning that came from the market. So the 10 lines is not the hard part. That's quite easy. And so even if that's infinite, it doesn't matter. The hard part is the learnings.
It also turns out that today writing code is running code. And much of dealing with software is managing it in production. This is dealing with performance, cost, reliability, and AI doesn't help much there because you don't know what you're going to need until you do it, right? So, you still need people that understand running systems and performance systems.
So, as software projects get more mature, the issues dominate with or without AI: learning from the customer and operating. In fact, I've had one software engineer, very senior, tell me that writing software has always effectively been cut and paste, and these other things like architecture, operations, and learnings are the challenges of creating software.
So while coding is very much dead, programming really isn't. I think we will not be looking at software in a year. It's quite clear that software engineering is very much not. You still need professionals that understand how computers work. You need them for architecture, for design, for operations. This is still very much a technical discipline.
And so, despite many decrying the end of software engineers, it's worth noting that the most active AI companies I know are all hiring software engineers a lot. And in fact, in Silicon Valley, we have a massive negative unemployment for software engineers. And we are actually hiring technologists in areas that we've not hired for in years: chip design, network design, operating systems development.
So I would say coding is dead, but software engineering is far from being dead. I think we're in a golden age of software. This right now is the best time to be in technology and software. And there are many reasons for this, some that I alluded to, but I would like to go over them one more time.
So to begin with, there's much more software being created now than ever before in the history of compute. So I have been a programmer for a long time, but I stopped because it was too complicated to get back in and learn the frameworks. But now I program most days because AI has made it so easy.
I run an investment team, some of whom are non-technical. Now people in my team who have never programmed before are programming, and they're doing it for personal productivity tools, work productivity tools, hobbies, etc. So we're really entering this era of bespoke software where every idea can become a software app.
Now if there's more software being written, then you need infrastructure to support it. And so now we have another opportunity to redo the full compute stack. Every 10 years to 20 years we get the opportunity to recreate computer science, computer systems. From the mainframe to client-server, we had to redo everything. The internet, we had to redo everything. The movement from the wiring closet to the cloud, we had to redo everything. We had to redo chips, we had to redo networking, we had to do operating system software.
And the same thing goes for AI. The workload is very, very different. So for the first time in Silicon Valley, we're seeing companies come and do hardware, do switching chips, build operating systems, build kernels for inference, build middleware layers. And so again, we're seeing this massive amount of energy and a massive amount of innovation in areas where previously there had been no investment.
So I think the last one is the most important one, and that's software can now tackle problems that it's never been able to tackle before. So it actually increases the reach of software. Right? So, it's not just writing code, but it's things like creativity. It's things like science, emotions and therapy, right? And so now software can go and tackle a number of problems that it never has. And so, I just view the TAM growing so much. So, it's just a really incredible time to be in software.
So to wrap this up, I just want to say that in my lifetime, I never expected to see a new super cycle like the internet. I was there. I mean, I'm old enough to remember what it was like. And there was so much energy and there was so much innovation and it was so inspiring and so many new companies got created and we created an entire industry and it solved problems for humans. And you know, that frankly has slowed down, and I feel like we're all very blessed to experience another one. And this one is almost certainly going to be bigger and more impactful.
So unlike the internet, for example, the economic case, like I told you, is very clear. The demand is the strongest any of us has ever seen. There's no limit in demand and there's a clear willingness to pay and a clear economic case. And this is already being reflected in the market even though it's just been a few years.
So the simplest way to think about it for me is that software has gotten much easier to make, more accessible for everyone. It's much, much, much more powerful, and now it's addressing a much, much larger market. So I think it really is the golden age of software. And thank you very much.
