AI Infrastructure, Distribution, and the Next Wave of Software

**Sophie Buonassisi** (0:00)
Andresen Horowitz raised $15 billion, and $1.7 billion of which was allocated towards infrastructure.

**Jennifer Li** (0:06)
When it comes to storage, compute, and all the tooling, we consider that a big part of infrastructure. Again, how do we store memory and so on? Those are all the opportunities that's emerging to build infrastructure.

**Sophie Buonassisi** (0:19)
Jennifer Li is a general partner at Andresen Horowitz.

**Jennifer Li** (0:22)
More than 90% of the code are being written by agents.

**Sophie Buonassisi** (0:25)
What did you see early in the voice AI space?

**Jennifer Li** (0:28)
So when we first saw the ElevenLabs demo, I remember you're using a Gandalf voice to just narrate a book, or like, holy shit, this is really alike, and also has all the right pauses, stressing, intonation. It's super engaging. I've always been a big fan of the One Piece manga. It's only every couple years for eight episodes. Right, yeah.
Make these products and tools your friends.

**Sophie Buonassisi** (0:52)
The best ideas live in the graveyard.

**SPEAKER_4** (0:55)
What does it take to build the infrastructure behind AI?
This episode, originally aired on the GTMnow Podcast, features Jennifer Li, general partner at a16z, in conversation with Sophie Buonassisi. They discuss the rapid evolution of AI infrastructure and the systems powering the next generation of software. From model development to developer tooling, Jennifer explains why infrastructure is becoming the most important layer of the stack, and how distribution is emerging as the key differentiator in a crowded market.

**Max** (1:28)
All right, we're back with another fun and amazing episode of the GTMnow Podcast, the special edition VC bonus episodes we have with myself and my general partner, Paul Irving. What's up, Paul? How are you doing?

**Paul Irving** (1:44)
Doing well, Max. How are you doing? You've been on the road quite a bit. I'm about to hit a road trip coming up, continuing to be pretty busy and exciting times.

**Max** (1:51)
Road warriors indeed. It feels like there's certain seasonality to it for sure. And this just happens to be one of those extremely busy times. So wouldn't have it any other way. It's fun and exciting. And with that, what better way to kick off the pod than how fun and exciting it is? Are we in a bubble or not? I know it's nuanced. Does this feel like 1998 or 2001? You know, we just had the all-birds pivoting to new birds AI and purchasing what, 50 million worth of GPUs to rent out. Like, there's stuff like that that I read and I'm just like, what?
You know, we're doing that again? But then there's, you know, the growth of Anthropic, where, you know, you're looking at that and you're like, no, this is real. This is a platform shift that we haven't seen. The likes of since, you know, maybe mobile or the Internet in general, right? So what's your take or what's your view on that?

**Paul Irving** (2:48)
I think from a high level, because if you only look at the high level, aspects of it rhyme so nicely, it's really easy to make that comparison. You say, okay, if AI is going to be a shift as transformational and large as the Internet when it first launched, there is going to be a huge build out of infrastructure that exists. Companies are going to grow as fast as you've ever seen them grow from an equity value perspective at least, and a lot of money is going to be thrown at it. And a lot of that stuff from a very high level will rhyme if you look at 1998, 1999, 2000, and what the last sort of three years have been post-Chat GPT launch. But you have to dive into the numbers, like you just mentioned, to really understand that this is completely different. Are there going to be some parallels? Maybe, but the core drivers of is there value today?
What are the economic factors which will influence the success of this over the coming two, three, four, five, ten years or not? So, I'm just going to list off a couple of those and would love to get your reaction to some of them as well. But four years after the Internet's public release, there were 70 million users globally. Chat GPT and AI apps already have one billion monthly active users. A completely different scale and in less time. 90% of this AI build out, so you talk about data centers, GPUs, 90% of it is pre-committed.
When we were running Fiber in the early 2000s and late 90s, it was 3% pre-committed. That's totally different. I think the differential between the two of them, you look at them, is just unused capacity versus used capacity. Every time that we bring more compute online, it's almost a one-to-one dollar creation ratio for the frontier model companies or the infrastructure companies to have that surface. And then you look at the constraint side of things, like Fiber was pretty cheap in relative terms. And you know what, over the ensuing 15, 20 years, we really benefited from an over build out of infrastructure. But right now, the infrastructure build out is not simple. Energy is a big constraint. Land is a big constraint. GPUs, if you dive into NVIDIA's earnings every single quarter or listen to Jensen when he talks, there's still capacity constraint from a supply perspective and meaningfully. And so, I think the core drivers of it is their demand, how pre-committed that demand is, the amount of users, the amount of value being created, the revenue growth of Anthropic, opening AI and even some of the smaller private market startups that we're lucky to meet and invest in is night and day different from what it was like in the late 90s and early 2000s.