00:00Joining us, one of the key backers of that ecosystem, Sarah Guo, founder of Conviction, an AI-native venture firm.
00:05And I think that's true, right?
00:08I think the best place, I think, to start with you is, think about all your portfolio companies.
00:12Many of them come on the show, and we say, okay, what's the biggest constraint on you right now?
00:17Where are you most focused?
00:18And oftentimes, it is compute.
00:21And they're sometimes super proud that they are working on NVIDIA gear.
00:25But are they accessing that compute through the cloud provider, or are they able to get the cluster themselves, direct
00:31control?
00:32It depends on what they're trying to do, right?
00:34We have companies that are building at the infrastructure layer, building at the model layer, building at the application layer.
00:41One of the first things we did when we started the fund was go buy compute for our companies, just
00:46because we can take the timing risk as a venture fund.
00:48And we concluded exactly what you just described, which is they're all going to need it, right?
00:53What did that look like?
00:54Sarah, sorry to interrupt you, but literally, what did it look like?
00:57We went and bought H100s at the time, a bunch of nodes, cloud.
01:03But the point was, the market has gone through different shortages.
01:08And now people are predicting a lot of ongoing constrained supply.
01:12And we were worried about access for startups in that period of time, which is why we did it.
01:17But our companies, to your actual question, if almost all of them want to do experimentation, and they all start
01:26with frontier performance, which is NVIDIA chips today, right?
01:29Because that allows you to do new things.
01:31And then, as these companies mature, they tend to post-train smaller models, think about cost, think about how to
01:39be able to, even just being able to use more tokens on the same task allows you to change the
01:46user experience.
01:46But everybody wants to start with current generation chips.
01:50The state of play that Jensen describes is demand vastly outpacing NVIDIA, but also, in aggregate, the industry's ability to
01:58supply.
01:59On the other side of the table, would you say that that state of play is true for the founders
02:04that you're trying to help grow?
02:05It's probably been two quarters of increasing stress in the ecosystem about access to supply at different scales.
02:16Wow.
02:16And so I have spent a good amount of time with the leaders of companies that serve NVIDIA chips in
02:24cloud asking to buy $100 million at a time of compute.
02:30And I've never been in that scenario before where we're trying very hard to pay somebody a lot of money
02:35with a multi-year commitment.
02:37Yeah.
02:37And it just speaks to the shortage dynamics.
02:39I think it's very hard to get on-demand, small-scale compute right now, which is what many startups are
02:45starting with, as you were describing.
02:49And so I think the shortage dynamic is real.
02:52And what is more interesting to me is, you know, NVIDIA have this amazing outperformance on the earnings side.
03:02And Jensen will repeatedly tell people demand is parabolic.
03:07Humans struggle to consume that.
03:09And, you know, it's hard to feel bad for Jensen Huang, Jensen's winner in all of this.
03:14But, you know, he's like...
03:15Thank you for being honest.
03:16Yeah, yeah.
03:16But, like, and he's an amazing leader, but he's like, I'm telling you, honestly, the reality of what I project
03:24repeatedly, and you just keep not believing me, right?
03:27And I'd say, like, I believe him from the demand side because what I expect will happen is this, you
03:34know, massive exponential growth that shows up in things like cloud code revenue is really because you have long-horizon
03:41agents that people are using in this one use case already.
03:44But the world is not just code, right?
03:47And so if there's zero, not zero, but small scale, a few billion to 40 billion of revenue in a
03:53very small number of months because we figured out how to use the models more productively on long-horizon agent
04:00tasks, we're going to do that for lots of different functions in the knowledge economy.
04:04While you're here, if you don't mind, like, SpaceX S1, wow.
04:09Yes, wow, it's inspiring.
04:10You have a big event today, right, where I believe, like, some of the cursor guys will be there.
04:14And so we learn a little bit about that, but there's this number, total addressable market for AI, $26.5
04:21trillion, super focused on enterprise.
04:25You know, they literally label it as such in the deck.
04:29Go ahead, I mean.
04:31Yeah, I think it's a funny number and then, like, a funny, you know, diagram when you visualize the tan
04:37because it's like Starlink and then enterprise AI.
04:39Yeah, yeah, yeah, yeah.
04:40And, you know, much of what we invest in is enterprise AI.
04:44My friend Andre Karpathy described this well, which is the simplest way to think about the use of AI is
04:50automation, right?
04:52It's automation of tasks we already do.
04:54Right.
04:55And if we give these models the tools and the harness such that they can do these tasks, they can
05:00take over things for us.
05:01And so I do think that the opportunity is as big as Elon says.
05:08But does Elon believe the opportunity or is this him trying to justify why SpaceX took XAI?
05:13I think it's interesting that, you know, what you put in the S1 is a choice, right?
05:18And he could have easily just said, you know, like, something broader or, you know, focus on other parts of
05:25the SpaceX TAM overall.
05:27But he is making the commitment that we are going to go make, you know, our offerings relevant in enterprise
05:34AI.
05:35And so I think he's very committed to this.
05:37Yes or no?
05:37You think he will pull it off?
05:40So there's a question about whether the value is infrastructure, models, or applications.
05:45I see this in the, you know, Anthropic deal and the Cursor deal as well.
05:49They have the infrastructure.
05:50The infrastructure and the capability to build more is extraordinarily valuable.
05:55And I think the question will be, like, I think he's going to make money on that no matter what.
05:59But the question is, do they also need to own the model and the application layer?
06:03I think the question is, do they need to own the model and the application layer?
Comments