WEBVTT

00:00:00.000 --> 00:00:02.077
And I think we're all pretty AI pilled.

00:00:02.160 --> 00:00:05.197
And if you're AI pilled, that means we
got to build a lot more compute than the

00:00:05.280 --> 00:00:08.597
world thinks. And that these models
are going to be a lot more valuable than

00:00:08.680 --> 00:00:12.317
people think. You combine that with
their core business, I don't know

00:00:12.400 --> 00:00:14.557
another entrepreneur or another business

00:00:14.640 --> 00:00:19.397
that's a better bet on the future,
right, than SpaceX. And so I think for

00:00:19.480 --> 00:00:21.517
most institutional investors, it's a

00:00:21.600 --> 00:00:24.397
must buy, a must own, a set it and

00:00:24.480 --> 00:00:28.917
forget it, right, in order to have
a a real bet on both the space and the AI

00:00:29.000 --> 00:00:31.480
future.

00:00:40.720 --> 00:00:44.997
All right, here we go.
Early morning Silicon Valley, BG2 is

00:00:45.080 --> 00:00:49.317
back. We're chopping it up on all things
tech and markets. To do that, I have

00:00:49.400 --> 00:00:53.637
none other than GB in the house, Gavin
Baker from a Treaties. He's brought his

00:00:53.720 --> 00:00:58.037
main guy, Andrew Fox. And of course,
I had to draft Clark Tang into the mix, my

00:00:58.120 --> 00:01:00.597
partner, um to talk
to to talk about some of the

00:01:00.680 --> 00:01:04.077
big questions of the day. You know,
how should we be thinking about the SpaceX

00:01:04.160 --> 00:01:06.037
IPO? You know, what are the big levers?

00:01:06.120 --> 00:01:08.957
There are big numbers out there
for what's going to happen over the course

00:01:09.040 --> 00:01:12.717
of the next few years. So, let's break
that down a bit, help simplify it for

00:01:12.800 --> 00:01:17.957
folks. I Mythos launched yesterday. I want
to talk a little bit about like who's up,

00:01:18.040 --> 00:01:20.917
who's down in the race
for superintelligence. Where are we? What

00:01:21.000 --> 00:01:22.917
did we learn with the Mythos launch? And

00:01:23.000 --> 00:01:25.317
Clark was in uh in in in Taiwan last

00:01:25.400 --> 00:01:28.557
week um with Jensen at Computex and GTC.

00:01:28.640 --> 00:01:31.957
So, what was our takeaway there? What's
going on with GPUs, memory? Where are

00:01:32.040 --> 00:01:37.237
the bottlenecks? And where do we go
from here? To start everything off, um you

00:01:37.320 --> 00:01:41.037
know, maybe just kick it over to you,
Gavin, talking about the SpaceX IPO. The

00:01:41.120 --> 00:01:45.157
IPO is in 2 days. Uh you're a big
shareholder. Congratulations. We're also

00:01:45.240 --> 00:01:49.757
a shareholder. Uh you know, we also we
expect to be buying in the IPO. The Wall

00:01:49.840 --> 00:01:51.637
Street Journal's reporting

00:01:51.720 --> 00:01:54.037
um you know, the Goldman Sachs are both

00:01:54.120 --> 00:01:57.797
saying 160 billion in revenue in 2028.

00:01:57.880 --> 00:02:01.397
Um we know that the IPO is $135 a share,

00:02:01.480 --> 00:02:05.677
1.77 trillion. Um so

00:02:05.760 --> 00:02:08.357
when we think about kind of what the big

00:02:08.440 --> 00:02:11.157
levers are, there's so many moving parts

00:02:11.240 --> 00:02:15.437
in this IPO. Um nobody's better than
you at just breaking it down, simplifying

00:02:15.520 --> 00:02:17.917
it. What are the key levers that we
ought to be thinking about that you're

00:02:18.000 --> 00:02:20.157
thinking about over the course
of the next few years?

00:02:20.240 --> 00:02:23.557
>> Sure. Um so great to be here. Thank you
for having me. I thought we're going to

00:02:23.640 --> 00:02:27.357
call it B BGG B,
but [laughter] we can stick with BG2.

00:02:27.440 --> 00:02:29.797
I've been your house. >> hey.
All subject to revision.

00:02:29.880 --> 00:02:33.477
>> That's okay. That's okay.
So I think there's two

00:02:33.560 --> 00:02:37.797
big levers or variables that I think
people should focus on. And you know,

00:02:37.880 --> 00:02:41.517
I'm not going to comment on where
I think um those variables go.

00:02:41.600 --> 00:02:44.357
But one is um you guys have this chart.

00:02:44.440 --> 00:02:46.677
Um did you did you post this on X?

00:02:46.760 --> 00:02:52.077
>> I did I did before and then we also
included a new addition with uh XAI's

00:02:52.160 --> 00:02:54.117
new deals as well. >> Yeah. >> Yeah.

00:02:54.200 --> 00:02:57.757
>> So Clark, who I've
known for many years, um

00:02:57.840 --> 00:03:00.877
uh made a did a great analysis here.

00:03:00.960 --> 00:03:03.997
And he shows that XAI's deal um with

00:03:04.080 --> 00:03:06.917
Google for cloud computing

00:03:07.000 --> 00:03:08.677
uh generates more operating profit per

00:03:08.760 --> 00:03:15.107
gigawatt um than Anthropic, than Meta,
than Google, than OpenAI. Uh their deal um

00:03:15.190 --> 00:03:18.037
>> [clears throat] >>
actually with uh Anthropic

00:03:18.120 --> 00:03:21.357
also generates probably more operating

00:03:21.440 --> 00:03:25.157
profit than anyone but um Anthropic.

00:03:25.240 --> 00:03:29.357
And so you know, um the your your
colleague at Altimeter, Freida, also she

00:03:29.440 --> 00:03:32.797
calculated a 55% IRR >> Mhm.
>> on Claude's one.

00:03:32.880 --> 00:03:34.797
>> Mhm. >> You know,
if you can borrow money at 6,

00:03:34.880 --> 00:03:39.277
7, 8% and invest
in something with a 55% ARR,

00:03:39.360 --> 00:03:41.837
I'm not the most sophisticated
thinker, but that math maths.

00:03:41.920 --> 00:03:44.837
>> Right. >> And so I
think the most important

00:03:44.920 --> 00:03:48.597
variable, one of the two most important,
is how quickly they bring on terrestrial

00:03:48.680 --> 00:03:52.837
data centers. >> Mhm.
>> We do know from Jensen that uh Elon

00:03:52.920 --> 00:03:57.677
brings data centers up faster than
anyone 122 days. Speed is literally cost

00:03:57.760 --> 00:04:00.917
because every day you're paying
electricians and plumbers. That's cost.

00:04:01.000 --> 00:04:04.077
And they're now monetizing them at

00:04:04.160 --> 00:04:08.997
arguably the highest rate. And so I
think, you know, everybody should run

00:04:09.080 --> 00:04:11.917
their own math on that,
but that is a massive variable.

00:04:12.000 --> 00:04:15.077
>> Yes. >> Truly massive
variable. The second thing

00:04:15.160 --> 00:04:19.277
is, you know, we have a chart and it's
wildly out of date now. It's kind of

00:04:19.360 --> 00:04:21.637
freaking amazing. This chart is I think

00:04:21.720 --> 00:04:23.797
is this chart from 10 days ago?

00:04:23.880 --> 00:04:28.837
But it like the 10 or 12 days since
this chart since we made this chart

00:04:28.920 --> 00:04:30.757
which shows the Pareto curves for Opus

00:04:30.840 --> 00:04:32.837
4.7 for coding

00:04:32.920 --> 00:04:35.957
for Codex from OpenAI. And now we've had

00:04:36.040 --> 00:04:38.237
Opus 4.8. It was already out of date.

00:04:38.320 --> 00:04:40.397
And now we have Fable >> Totally.

00:04:40.480 --> 00:04:43.277
>> and Mythos, which is freaking wild.
In 10 days >> Yes.

00:04:43.360 --> 00:04:45.517
>> like we would have had
to update the chart twice.

00:04:45.600 --> 00:04:48.437
>> Right. >> But what
the Pareto curve sure shows is

00:04:48.520 --> 00:04:52.237
how much intelligence you can get
for a given amount of cost. And I do think

00:04:52.320 --> 00:04:55.917
being all revenue will accrue
to the Pareto curve. All at least kind of

00:04:56.000 --> 00:04:59.197
frontier model revenue will accrue
to the Pareto curve. And this is Pareto

00:04:59.280 --> 00:05:04.157
curve for code coding.
And what I think is so impressive

00:05:04.240 --> 00:05:07.757
is that you can see
in the chart that Composer 2

00:05:07.840 --> 00:05:11.397
was Pareto dominant or,
you know, at the lowest level of

00:05:11.480 --> 00:05:14.037
intelligence with very little training.

00:05:14.120 --> 00:05:16.637
This just reflects, and I know you know
Cursor well. I think you know Cursor

00:05:16.720 --> 00:05:19.577
a lot better than I do.
A vast amount better than I do.

00:05:19.660 --> 00:05:21.517
[laughter]

00:05:21.600 --> 00:05:24.037
But my understanding is that Cursor and

00:05:24.120 --> 00:05:28.557
Anthropic have more tokens
of proprietary coding data than anyone

00:05:28.640 --> 00:05:31.957
else. And they have more tokens
of proprietary coding data than exist on

00:05:32.040 --> 00:05:33.397
the public internet.

00:05:33.480 --> 00:05:36.437
And so they fed Cursor fed

00:05:36.520 --> 00:05:41.437
um used ChemK 0.25, used their own
private data, did some RL, some

00:05:41.520 --> 00:05:44.637
supervised fine-tuning and they
got a really good model. And

00:05:44.720 --> 00:05:46.837
then they spent 3 weeks in the Colossus

00:05:46.920 --> 00:05:50.157
2 cluster and they got a model that

00:05:50.240 --> 00:05:54.597
12 days ago was Pareto dominant
with Composer 2.5. That's on their own

00:05:54.680 --> 00:05:58.317
benchmark. Um Cursor bench,
so maybe take it with a

00:05:58.400 --> 00:06:00.317
grain of salt. But I think this just

00:06:00.400 --> 00:06:02.997
suggests that the Cursor data is very

00:06:03.080 --> 00:06:07.317
valuable for coding and when it is

00:06:07.400 --> 00:06:10.437
trained, you know, to Chinchilla
optimal or beyond Chinchilla optimal with

00:06:10.520 --> 00:06:14.637
reinforcement learning, you know,
I think it suggests that XAI

00:06:14.720 --> 00:06:17.677
and SpaceX AI has a shot
of being a real player in coding.

00:06:17.760 --> 00:06:21.637
>> I mean, I think one of the
interesting things is, you know, we

00:06:21.720 --> 00:06:23.637
you know, the way he answered
the question, right? We didn't talk about

00:06:23.720 --> 00:06:26.037
launch. Right?
We didn't talk about Starlink or

00:06:26.120 --> 00:06:31.277
communications. Those up until
really 6 months ago were the business.

00:06:31.360 --> 00:06:33.997
>> Yeah. >> Right? And,
you know, and then we merged

00:06:34.080 --> 00:06:37.237
in x.ai and we merged in Cursor and then

00:06:37.320 --> 00:06:41.277
we announced these deals where it
was very clear he was kind of building AWS

00:06:41.360 --> 00:06:45.037
right under our nose, you know, in in in
in terms of this. But what I want to do

00:06:45.120 --> 00:06:49.357
is go to go to Fox. Give us the breakdown.
Three big lines of business,

00:06:49.440 --> 00:06:52.997
right? We've got the the the
communication Starlink launch business,

00:06:53.080 --> 00:06:56.357
we've got the, you know, AI compute
business and then I want to come back to

00:06:56.440 --> 00:07:01.157
x.ai that you were just clicking on. But
if we just go to the core business, what

00:07:01.240 --> 00:07:05.597
do we have to assume goes right in the
core business both with launch and with

00:07:05.680 --> 00:07:08.517
Starlink in order to achieve
the numbers that are out there?

00:07:08.600 --> 00:07:11.717
>> Yeah, sure. So, look, I think
the thing that's foundational to

00:07:11.800 --> 00:07:13.437
everything is the launch business.

00:07:13.520 --> 00:07:15.757
>> Right. >> Right? This is the kind of

00:07:15.840 --> 00:07:20.557
crown jewel of SpaceX. Um it's something
that no one else really has, notably

00:07:20.640 --> 00:07:24.277
reusability. >> Right.
>> And soon rapid reusability. Right? This

00:07:24.360 --> 00:07:26.917
is, I think, what you need to believe in

00:07:27.000 --> 00:07:31.037
to get to the economics in AI that make

00:07:31.120 --> 00:07:34.677
orbital compute something that's
very economically attractive.

00:07:34.760 --> 00:07:36.877
>> Right. >> Outside
of the idea that we are in

00:07:36.960 --> 00:07:39.957
shortage for power, shortage for chips.

00:07:40.040 --> 00:07:44.077
Right? Um so, I think rapid reusability
is the main thing that we're watching

00:07:44.160 --> 00:07:47.077
for and I think most
people should watch for. Um,

00:07:47.160 --> 00:07:50.437
you know, Elon talks about it a lot,
but getting these rockets to fly at a

00:07:50.520 --> 00:07:53.837
cadence that's comparable to an airline,
right? And and Gavin has used this

00:07:53.920 --> 00:07:57.637
analogy before, but um,
the old rocket industry was kind of

00:07:57.720 --> 00:08:01.477
like, imagine boarding a plane, flying
to California, getting off the plane,

00:08:01.560 --> 00:08:02.997
the plane explodes

00:08:03.080 --> 00:08:06.077
after. Um, so I think what SpaceX are

00:08:06.160 --> 00:08:08.637
ultimately trying to achieve is have a

00:08:08.720 --> 00:08:13.637
Starship fly both stages,
not just the booster. Um,

00:08:13.720 --> 00:08:17.237
uh, 30, 40, 50 times

00:08:17.320 --> 00:08:20.197
before you have to retrofit that ship.

00:08:20.280 --> 00:08:23.077
Um, and when you do that,
you're amortizing

00:08:23.160 --> 00:08:27.077
the cost of the vehicle over many
flights, right? And that's what brings

00:08:27.160 --> 00:08:31.557
the cost down significantly. Um,
>> But that's a really hard problem to

00:08:31.640 --> 00:08:35.197
solve. >> Extremely
difficult and look, I think

00:08:35.280 --> 00:08:38.837
the company, you know, have been
loud and clear, they're going to

00:08:38.920 --> 00:08:41.077
attempt to bring back
the second stage >> Right.

00:08:41.160 --> 00:08:42.757
>> of Starship later this year.

00:08:42.840 --> 00:08:45.917
>> Right. >> Um, and then
make it reusable, you know,

00:08:46.000 --> 00:08:49.637
re-fly the second stage next year. Um,

00:08:49.720 --> 00:08:54.597
and from there, ramp up the cadence. But
at the end of the day, driving down the

00:08:54.680 --> 00:08:58.237
cost of launch is what
enables all of these other

00:08:58.320 --> 00:09:03.197
businesses and is what makes them
so attractive relative to incumbents.

00:09:03.280 --> 00:09:06.557
>> So how many how many times are we
Starship just launched Starship 3, you

00:09:06.640 --> 00:09:10.677
know, just launched. How many launches
are you know, do you think kind of the

00:09:10.760 --> 00:09:14.717
consensus out there is assuming, you know,
two or three years from now? Like

00:09:14.800 --> 00:09:17.917
what is the launch cadence? Are we
launching one of these every day or we

00:09:18.000 --> 00:09:21.477
launching one of these every week
or every month? Like where are we in terms

00:09:21.560 --> 00:09:24.557
of expectations? >> Yeah,
so look, I think expectations for

00:09:24.640 --> 00:09:27.477
now, you know, we're going from, you

00:09:27.560 --> 00:09:30.517
know, call it 160 165 launches last year

00:09:30.600 --> 00:09:34.797
up into the high hundreds of launches
in several years and, you know, getting

00:09:34.880 --> 00:09:38.397
into the thousands of launches
probably in the next 3 years thereafter.

00:09:38.480 --> 00:09:41.597
>> Okay. >> Um I think
the company have aspirations.

00:09:41.680 --> 00:09:43.077
>> Thousands of launches,
you're launching,

00:09:43.160 --> 00:09:45.397
you're doing two or three launches a day.

00:09:45.480 --> 00:09:48.397
>> Right. >> Right.
And then talk to us a little bit,

00:09:48.480 --> 00:09:51.837
what is this enable? Obviously,
you know, I'm here in Silicon Valley. I

00:09:51.920 --> 00:09:57.077
can't even I can't even keep a call
on Sand Hill Road, two decades into the

00:09:57.160 --> 00:10:01.157
mobile revolution. I mean, it's the
craziest thing. It's like a third world.

00:10:01.240 --> 00:10:04.197
>> It's It's a major business problem
when you're freaking out here. [laughter]

00:10:04.280 --> 00:10:08.597
>> It's crazy. It's crazy. Right by the
Starwood dead zone. And I'm like, how

00:10:08.680 --> 00:10:11.797
can this possibly be? So almost like
it's a joke. It's the epicenter of

00:10:11.880 --> 00:10:15.997
technology in America and you can't
maintain a call. Okay, so we're all

00:10:16.080 --> 00:10:18.797
going to switch to Starlink mobile when
it comes along because I don't want to

00:10:18.880 --> 00:10:21.197
lose that call on Sand Hill Road. So

00:10:21.280 --> 00:10:23.157
walk me through a little bit just again

00:10:23.240 --> 00:10:28.517
high-level. Um it's a big portion of the
revenue growth expected in the business

00:10:28.600 --> 00:10:31.957
over the course of the next two to three
years. My hunch is a lot of this is

00:10:32.040 --> 00:10:37.037
driven uh by uh by direct to cell
connectivity. Walk me through a little

00:10:37.120 --> 00:10:41.237
bit those economics. >> Yeah,
so look, it's actually interesting. Um

00:10:41.320 --> 00:10:45.637
the broadband business is still very
early stage when you think about um the

00:10:45.720 --> 00:10:49.397
percent of households that have actually
been penetrated to date. You look at the

00:10:49.480 --> 00:10:53.637
percent of global households with
Starlink, it's less than 1%. Uh and

00:10:53.720 --> 00:10:57.357
that's the broadband, you know, you kind
of have a base terminal at your house,

00:10:57.440 --> 00:11:00.237
on your car, on your boat, um and then

00:11:00.320 --> 00:11:04.717
airlines now as well.
Um so I actually think broadband can scale

00:11:04.800 --> 00:11:08.917
to hundreds of millions of terminals,
hundreds of millions of users. And today

00:11:09.000 --> 00:11:09.837
the subscriber base

00:11:09.920 --> 00:11:13.797
>> Hundreds of millions if if they get

00:11:13.880 --> 00:11:17.397
rapid reusability of Starship,
which is really hard.

00:11:17.480 --> 00:11:20.917
Um you know, if if
there's not competition.

00:11:21.000 --> 00:11:23.957
Um hundreds of millions, um it's possible.

00:11:24.040 --> 00:11:27.197
But maybe >> I I always
say around here, it's funny.

00:11:27.280 --> 00:11:31.117
I love seeing PM and kind of analysts
in in this situation. It's exactly what I

00:11:31.200 --> 00:11:34.357
do with Clark. Clark will say something,
I'll say the future is a distribution of

00:11:34.440 --> 00:11:37.757
unknown probabilities. It's either more
likely or less likely, so give me the

00:11:37.840 --> 00:11:40.557
distribution. Are we talking 20% 30%?

00:11:40.640 --> 00:11:43.037
It's hilarious. It's the same >> Well,
no, 100% same thing. And like I've

00:11:43.120 --> 00:11:46.757
watched Elon do many hard things and
this is a really hard thing. So, I think

00:11:46.840 --> 00:11:50.717
it's reasonable to think that they're
going to succeed with rapid reusability,

00:11:50.800 --> 00:11:54.877
but just I just think it's
important to acknowledge that like

00:11:54.960 --> 00:11:56.517
orbital compute

00:11:56.600 --> 00:11:58.397
you know, Starlinks, you know, Starlink

00:11:58.480 --> 00:12:01.957
V3, Starlink direct to cell, we need we

00:12:02.040 --> 00:12:06.757
need first reusability for Starship V3
and then rapid reusability unlocks a lot

00:12:06.840 --> 00:12:09.757
of this. >> Right.
When I see when I see the models

00:12:09.840 --> 00:12:12.357
that the banks are putting out there,
right? And Wall Street Journal,

00:12:12.440 --> 00:12:15.917
everybody's reported on these.
These same things have been widely leaked.

00:12:16.000 --> 00:12:20.157
They they they largely have the revenue
on connectivity, so let's call it

00:12:20.240 --> 00:12:22.397
Starlink direct to cell, etc. Going

00:12:22.480 --> 00:12:24.717
from, you know, let's call it 10 billion

00:12:24.800 --> 00:12:27.957
to 50 billion uh by 2028. And so, I'm

00:12:28.040 --> 00:12:32.197
not asking you guys to react to, you know,
to tell me your specific numbers,

00:12:32.280 --> 00:12:34.557
but when I'm talking to Clark all I'm

00:12:34.640 --> 00:12:37.557
trying to size up is order of magnitude.

00:12:37.640 --> 00:12:41.797
Do we think we can 5x the business over
the course of the next 3 years? Is there

00:12:41.880 --> 00:12:45.237
enough TAM both in terms of broadband
and direct to consumer? And I think the

00:12:45.320 --> 00:12:47.957
answer to that is yes. >> Yeah,
here's what I just say very simply

00:12:48.040 --> 00:12:50.437
is I have um

00:12:50.520 --> 00:12:55.277
I I travel with Starlink.
Um I'm I'm a big video gamer and very

00:12:55.360 --> 00:12:58.557
consistently wherever I am in the world,
Starlink is the best connection.

00:12:58.640 --> 00:13:01.597
>> Yes. >> It's the fastest,
it's lowest latency

00:13:01.680 --> 00:13:05.757
and I do think once they get to rapid
reusability, it's also going to be

00:13:05.840 --> 00:13:07.917
they're going to have the cheapest cost

00:13:08.000 --> 00:13:11.477
per gigabyte or megabyte delivered um

00:13:11.560 --> 00:13:15.677
and better faster, cheaper
has been a winning formula.

00:13:15.760 --> 00:13:21.157
And so, 50 billion, that's, you know,
0.3% penetration of the global telecom

00:13:21.240 --> 00:13:24.877
market. Now, maybe there's some
deflation with Starlink pricing.

00:13:24.960 --> 00:13:26.597
Um but that's the way I'd frame it up.

00:13:26.680 --> 00:13:28.317
>> Yeah. I like betting
on better, faster, cheaper.

00:13:28.400 --> 00:13:32.677
>> Um, Clark, I would say probably
the biggest surprise of the last six six

00:13:32.760 --> 00:13:36.277
weeks is that Elon,
you know, we talked about

00:13:36.360 --> 00:13:38.637
it on all-in podcast, we called it EWS,

00:13:38.720 --> 00:13:44.397
Elon web services, right? That that he
struck these huge deals with Anthropic

00:13:44.480 --> 00:13:47.317
and Google. I don't even
think people were thinking

00:13:47.400 --> 00:13:50.517
about SpaceX in the AI compute game,

00:13:50.600 --> 00:13:54.957
right? We If you looked at the models
as of a few months ago, it was

00:13:55.040 --> 00:13:56.997
connectivity, so Starlink, and then it

00:13:57.080 --> 00:13:59.597
was x.ai, the model.

00:13:59.680 --> 00:14:02.677
But this whole category of taking all of

00:14:02.760 --> 00:14:06.637
this compute, which he's uniquely
good at standing up, right? And then

00:14:06.720 --> 00:14:10.317
reselling it in a way that's highly
profitable was not in a lot of people's

00:14:10.400 --> 00:14:13.277
forecast. Now it's a major component of

00:14:13.360 --> 00:14:17.957
the forecast. You know, you and I did
this podcast with Jensen, where Jensen

00:14:18.040 --> 00:14:22.797
said Elon is an N of 1. >> What
they achieved is is singular. Never

00:14:22.880 --> 00:14:27.797
been done before. Just to put
in perspective, 100,000 GPUs, that's

00:14:27.880 --> 00:14:31.837
you know, easily the fastest supercomputer
on the planet as one cluster.

00:14:31.920 --> 00:14:34.677
Um, a supercomputer uh,

00:14:34.760 --> 00:14:39.117
that you would build would
take normally 3 years to plan.

00:14:39.200 --> 00:14:41.917
>> Right. >> And then they
deliver the equipment, and

00:14:42.000 --> 00:14:44.237
it takes 1 year

00:14:44.320 --> 00:14:49.117
to get it all working. Yes.
We're talking about 19 days.

00:14:49.200 --> 00:14:52.157
>> Wow. >> N of 1 is right.
Elon is an N of 1.

00:14:52.240 --> 00:14:55.397
>> And his ability to secure supply, stand

00:14:55.480 --> 00:14:57.917
up the supply, you know, deploy it in a

00:14:58.000 --> 00:15:00.237
way that's uh, you know, coherent and

00:15:00.320 --> 00:15:02.517
effective for both himself and I guess

00:15:02.600 --> 00:15:05.717
now for others. So, walk us through kind

00:15:05.800 --> 00:15:09.877
of that. It looks to me again like
this is a major component of the revenue

00:15:09.960 --> 00:15:13.557
story. >> Totally.
I mean, so we we were all at

00:15:13.640 --> 00:15:18.197
the macro hard data center, and it
was just very evident the amount of

00:15:18.280 --> 00:15:21.997
engineering that was that had
gone into building these sites.

00:15:22.080 --> 00:15:24.877
Um, you people always
talk about Google and

00:15:24.960 --> 00:15:29.277
their ability to to build a TPU
and sell the TPU to Anthropic to generate

00:15:29.360 --> 00:15:30.997
revenues for AI.

00:15:31.080 --> 00:15:32.997
I think it's a pretty similar dynamic

00:15:33.080 --> 00:15:36.917
here with Elon able to secure power, uh

00:15:37.000 --> 00:15:39.197
build these sites faster than anyone

00:15:39.280 --> 00:15:42.397
else and also be able now to monetize it

00:15:42.480 --> 00:15:47.037
to um to the this massive
AI market that's ahead of us.

00:15:47.120 --> 00:15:51.557
Um if you look at the the relationships

00:15:51.640 --> 00:15:55.237
that he's forged with a lot of his
suppliers, you know, be it Jensen, be

00:15:55.320 --> 00:15:58.397
it, you know, all of these different um

00:15:58.480 --> 00:16:03.357
different sites that actually want
xAI as a tenant. Um his his ability to

00:16:03.440 --> 00:16:07.677
finance these deals at at very
attractive uh financing rates relative

00:16:07.760 --> 00:16:12.357
to a lot of the other players in in the
space, you know, these are advantages

00:16:12.440 --> 00:16:15.997
that compound over time. And when you've
built the credibility to stand up these

00:16:16.080 --> 00:16:20.477
sites and monetize at these levels, um
you know, it's very it's actually a very

00:16:20.560 --> 00:16:22.357
attractive uh

00:16:22.440 --> 00:16:27.517
very attractive proposition for for a lot
of folks involved. Um and actually,

00:16:27.600 --> 00:16:31.077
you know, if you look
at the these deals in particular,

00:16:31.160 --> 00:16:33.557
Gavin, you you pointed out, but,
you know, they're they're actually

00:16:33.640 --> 00:16:38.637
monetizing, you know, perhaps better
than other players in the space by selling

00:16:38.720 --> 00:16:40.197
this infrastruc- >> a lot higher.

00:16:40.280 --> 00:16:45.157
>> Um Google is obviously paying SpaceX
a huge premium for this compute. Fox, you

00:16:45.240 --> 00:16:47.117
said something that I thought was really

00:16:47.200 --> 00:16:49.637
important, which is, you know, it may

00:16:49.720 --> 00:16:52.277
very well be that in order to get, you

00:16:52.360 --> 00:16:56.957
know, first in line on space compute,
which Google certainly wants to do, that

00:16:57.040 --> 00:17:01.357
they're willing to pay a premium
for their terrestrial compute. And so, to

00:17:01.440 --> 00:17:04.597
me, that's how you kind of square
the circle as to why the premium. Any

00:17:04.680 --> 00:17:06.597
thoughts? >> Yeah, look,
I think there's some of that

00:17:06.680 --> 00:17:10.997
embedded there, but um look, at the end
of the day, SpaceX can stand up compute

00:17:11.080 --> 00:17:15.237
quickly. They can stand it up coherently.
And they can stand up a lot of it in one

00:17:15.320 --> 00:17:18.317
place and have it readily available.
So, look, I think that's most of the

00:17:18.400 --> 00:17:22.757
premium, but outside of that certainly
people are going to space over time.

00:17:22.840 --> 00:17:25.197
>> I have to pay a little call
option to get first in line for space.

00:17:25.280 --> 00:17:27.117
>> There you go. Good one.
>> We've all been investing in the neo

00:17:27.200 --> 00:17:30.837
cloud space. So, like there's
a fundamental belief around this table I

00:17:30.920 --> 00:17:34.117
that that we lack the compute needed to

00:17:34.200 --> 00:17:37.957
continue to push the frontier on
intelligence. So, we have to build a lot

00:17:38.040 --> 00:17:42.517
of compute, okay? Now there's a there's
competition going on. On one end you

00:17:42.600 --> 00:17:46.037
have the hyper scalers who are building
out that capability. Then we have AI

00:17:46.120 --> 00:17:48.557
dedicated clouds that are building out

00:17:48.640 --> 00:17:50.197
that capability. And now literally in a

00:17:50.280 --> 00:17:53.197
matter of weeks, right, we have a a you

00:17:53.280 --> 00:17:55.677
know a giant that's emerged in this

00:17:55.760 --> 00:17:58.477
category which is SpaceX.

00:17:58.560 --> 00:18:00.277
The question to you Gavin is can they

00:18:00.360 --> 00:18:03.277
consolidate this market, right? Because

00:18:03.360 --> 00:18:07.757
if I think about a marketplace, Elon has
a unique ability to get the supply. He

00:18:07.840 --> 00:18:11.957
has a unique ability to cut deals on
the other side and nobody can stand it up

00:18:12.040 --> 00:18:16.757
like he can stand it up. So, I think
there might be a real consolidation in

00:18:16.840 --> 00:18:20.757
the AI compute market where you have the
hyper scalers on the one hand and on the

00:18:20.840 --> 00:18:25.717
other hand, you know, he may emerge
as the largest, strongest player in the AI

00:18:25.800 --> 00:18:29.197
compute market. >> Yeah,
so I think they're are they the

00:18:29.280 --> 00:18:32.277
number four number five hyper
scaler today after the Google deal?

00:18:32.360 --> 00:18:36.277
>> Um it will be number four.
>> Kind of wild.

00:18:36.360 --> 00:18:40.597
>> Yeah. >> In 30 days we
went from not being an AI hyper

00:18:40.680 --> 00:18:45.717
scaler to being number four. And we passed
a lot of companies including Oracle.

00:18:45.800 --> 00:18:47.437
>> Coreweave is a huge business, right?

00:18:47.520 --> 00:18:51.437
That we're we're investors in, you know,
and have been investors in, right? But

00:18:51.520 --> 00:18:54.317
there are a lot of other players, the
Nebius's of the world, the Iron's of the

00:18:54.400 --> 00:18:57.677
world. And I would say that they're
probably 50 neo labs being funded in

00:18:57.760 --> 00:19:01.437
Silicon Valley right now as we speak
because of the shortage in compute.

00:19:01.520 --> 00:19:06.557
>> Absolutely. So, that's kind
of crazy in 30 days. That's just

00:19:06.640 --> 00:19:09.637
extraordinary. What I would
say is there I think there

00:19:09.720 --> 00:19:12.877
is a belief that these data
centers are commodities.

00:19:12.960 --> 00:19:15.677
>> Mhm. >> And I do not share that belief.

00:19:15.760 --> 00:19:19.157
Um I don't think anybody around
this table shares that belief.

00:19:19.240 --> 00:19:23.477
And in the same way that Elon was able
to re-engineer a rocket from first

00:19:23.560 --> 00:19:27.437
principles and make it reusable,
he engineered an electric car from first

00:19:27.520 --> 00:19:30.517
principles. You know, everyone else
was trying to, you know, make an electric

00:19:30.600 --> 00:19:34.877
car like an internal combustion engine
car and he thought about it differently.

00:19:34.960 --> 00:19:39.597
And um I think he looked
at data center design

00:19:39.680 --> 00:19:43.677
from first principles and he designed
something fundamentally different. And I

00:19:43.760 --> 00:19:45.637
did actually ask the team. I said, "Hey

00:19:45.720 --> 00:19:51.037
guys, maybe I'd be a little less public
about things that are very obvious to

00:19:51.120 --> 00:19:54.557
you [laughter] >> Right.
>> about how to design a data center, but

00:19:54.640 --> 00:19:59.237
are revelations to other people
because I think what you're doing is

00:19:59.320 --> 00:20:02.157
maybe um more differentiated
than you perhaps

00:20:02.240 --> 00:20:05.117
realize cuz what you're
doing is so logical to you,

00:20:05.200 --> 00:20:08.957
but maybe lot not logical to everyone
else. And that's how he was able to do

00:20:09.040 --> 00:20:12.197
it in 122 days. >> Yeah,
to I mean to that point, Brad,

00:20:12.280 --> 00:20:14.997
yesterday we were meeting one of our
portfolio companies and we were talking

00:20:15.080 --> 00:20:18.397
about behind the meter and we're, you
know, really thinking about it. There's

00:20:18.480 --> 00:20:21.277
only maybe two or three two or three

00:20:21.360 --> 00:20:26.237
players now that can actually reliably
engineer behind the meter data center.

00:20:26.320 --> 00:20:29.037
And you know, there's real engineering
work that goes into all of this. So, if

00:20:29.120 --> 00:20:31.677
you think about this,
if you're a gas combustion if you're

00:20:31.760 --> 00:20:33.757
Vernova and you say we only have a

00:20:33.840 --> 00:20:38.837
certain number of gas combustion engines.
Now, we can sell them to x.ai

00:20:38.920 --> 00:20:42.277
or we can sell them to one of these
startup neo clouds. Who are you going to

00:20:42.360 --> 00:20:44.437
sell them to? >> Well,
and there's another dynamic.

00:20:44.520 --> 00:20:46.317
Everyone starts making more money when

00:20:46.400 --> 00:20:49.797
the GPUs get energized and sold faster.

00:20:49.880 --> 00:20:55.037
So, literally speed is money for all
of the suppliers. Power, land, turbines.

00:20:55.120 --> 00:20:57.917
So, I think it's we'll we'll see.

00:20:58.000 --> 00:21:00.717
>> Right. >> Hey Brad, man.
>> And but but this is just we're just

00:21:00.800 --> 00:21:02.797
talking terrestrial. I do I do want to

00:21:02.880 --> 00:21:04.837
hit on and then you can flip it back on

00:21:04.920 --> 00:21:07.837
me. Talk to me, okay, So, let's let's

00:21:07.920 --> 00:21:10.637
assume, right, that they continue to

00:21:10.720 --> 00:21:13.157
build out the terrestrial landscape.

00:21:13.240 --> 00:21:15.597
They continue to find buyers for that.

00:21:15.680 --> 00:21:19.877
Um, walk us through, you know, what
this unlocks, you know, and how this is

00:21:19.960 --> 00:21:23.357
related to space data centers because I
think, you know, once you start talking

00:21:23.440 --> 00:21:27.477
terrafab capacity and beyond. So, we're
talking 1,000 gigs, right? And this year

00:21:27.560 --> 00:21:31.397
what what we're doing, 25 or 30 gigs
just to put it all in perspective.

00:21:31.480 --> 00:21:33.677
>> 20, yeah. >> Right? >> 20, 25 gigs.

00:21:33.760 --> 00:21:36.757
>> Okay, so so once we start scaling up,

00:21:36.840 --> 00:21:41.237
walk us through, do we have
to have space data centers in

00:21:41.320 --> 00:21:46.357
order to get excited about buying the IPO,
right? And then there's obviously

00:21:46.440 --> 00:21:49.837
this this debate in the world. I I heard
Jeff Bezos say, you know, I think it's

00:21:49.920 --> 00:21:53.397
more like 6 years, but Elon's going
to say three because if if he says six,

00:21:53.480 --> 00:21:57.037
then it will take even longer. So, say
three and we may get it in four or five.

00:21:57.120 --> 00:22:00.357
But are space data centers integral and

00:22:00.440 --> 00:22:04.717
essential to, you know, the IPO? And what
do you think the timeline is, Erin?

00:22:04.800 --> 00:22:07.237
There are you guys. >> So,
I don't think I think if you think

00:22:07.320 --> 00:22:10.717
about those variables around what

00:22:10.800 --> 00:22:15.157
Crusher could mean for XAI.
And we do have an existence proof that

00:22:15.240 --> 00:22:18.477
once you really get
on that Pareto frontier,

00:22:18.560 --> 00:22:21.917
revenue can scale rapidly and it's
called Entropic. And there does seem to

00:22:22.000 --> 00:22:26.437
be an exhaust There seems to be a lot
of demand for coding. And I do think I'm

00:22:26.520 --> 00:22:29.157
John Massad posted
something very interesting.

00:22:29.240 --> 00:22:32.837
>> The The founder of Replit.
>> The founder of Replit. It he called it

00:22:32.920 --> 00:22:35.317
bitter lesson adjacent that coding may

00:22:35.400 --> 00:22:40.357
be the fastest path to AGI and ASI
because if you really go to coding, you

00:22:40.440 --> 00:22:43.837
can write code if a model's good
at coding to do anything. So, I think

00:22:43.920 --> 00:22:46.597
that's a profound point and I think
coding is going to continue to be very

00:22:46.680 --> 00:22:49.797
important. So, I think if
you think about that variable,

00:22:49.880 --> 00:22:51.277
if you think about Starlink direct to

00:22:51.360 --> 00:22:54.157
cell enabled by Starlink V3, and you

00:22:54.240 --> 00:22:57.357
think about how quickly they can or

00:22:57.440 --> 00:23:00.197
cannot bring on terrestrial compute, I I

00:23:00.280 --> 00:23:02.637
think orbital compute is is

00:23:02.720 --> 00:23:06.437
is necessary for the IPO valuation,

00:23:06.520 --> 00:23:11.237
but it's certainly important and it's >>
Well, maybe another way to say it is you

00:23:11.320 --> 00:23:15.157
think you you may think
we're going to get to ASI

00:23:15.240 --> 00:23:17.077
faster than we're going to get to

00:23:17.160 --> 00:23:19.037
orbital compute. That may take us from

00:23:19.120 --> 00:23:24.197
300 IQ to 400 IQ, 500 IQ, and beyond.

00:23:24.280 --> 00:23:28.517
Um and the ability to scale it up
to consume 10% of, you know, global GDP,

00:23:28.600 --> 00:23:30.997
but maybe maybe that's
where we should move next.

00:23:31.080 --> 00:23:34.717
>> No, no, I think on the orbital compute,
I think Foxy would be great our Clark to

00:23:34.800 --> 00:23:39.197
lay out the math from first principles on,
you know, Clark has this great chart

00:23:39.280 --> 00:23:42.557
on, you know, the gigawatts it costs,
you know, the dollars per gigawatt.

00:23:42.640 --> 00:23:44.957
>> Right. >> Walk us
through the economic case.

00:23:45.040 --> 00:23:48.357
>> Yeah. Yeah, so I mean, on this point of

00:23:48.440 --> 00:23:51.637
is orbital key to investing here? I

00:23:51.720 --> 00:23:55.677
don't think it is and I'll first
point I'll make is what are the implied

00:23:55.760 --> 00:23:59.157
monetization rates based
on expectations today for the AI

00:23:59.240 --> 00:24:02.197
business? And you know, I think you
threw out the $160 billion number that's

00:24:02.280 --> 00:24:04.557
been leaked out there
that people are talking about.

00:24:04.640 --> 00:24:08.877
The implied monetization rate on that
number is something like $14 billion per

00:24:08.960 --> 00:24:11.397
gigawatt per year for the AI business.

00:24:11.480 --> 00:24:13.997
They just signed Anthropic at 22 to 23.

00:24:14.080 --> 00:24:15.477
They just signed Google at 50.

00:24:15.560 --> 00:24:17.917
>> Right. >> Right. So,
I I think you can invest

00:24:18.000 --> 00:24:22.797
behind the AI business terrestrially
and still be excited about it. But with

00:24:22.880 --> 00:24:24.957
orbital >> an important point.
Excited about it if

00:24:25.040 --> 00:24:26.357
they can get the land and the power.

00:24:26.440 --> 00:24:30.477
>> Right. But but but I mean I I think
for most investors, right? They get They

00:24:30.560 --> 00:24:35.837
have an easier time getting their head
around how SpaceX wins terrestrially.

00:24:35.920 --> 00:24:39.037
Like can they go get land, power,
and chips? The answer to that is high

00:24:39.120 --> 00:24:43.557
probability yes, okay? And what we're
saying is at the rate they're monetizing

00:24:43.640 --> 00:24:47.237
that, that gets you to the numbers
that are being leaked out there before you

00:24:47.320 --> 00:24:50.277
even have to take the leap of faith
that they're going to extend the lead with

00:24:50.360 --> 00:24:52.637
orbital data centers.
But take us there on that, too.

00:24:52.720 --> 00:24:58.477
>> Sure. Yeah, so so look, with orbital,
I think the key thing is um two-stage

00:24:58.560 --> 00:25:02.757
reusability. >> Yeah. >> And beyond
that, rapid two-stage reusability.

00:25:02.840 --> 00:25:05.517
>> Yeah. >> So, today
with Starship, they've shown

00:25:05.600 --> 00:25:09.437
that they can successfully
reland the booster.

00:25:09.520 --> 00:25:12.317
The second stage, we'll see what
happens later this year. I think they're

00:25:12.400 --> 00:25:16.517
attempting to bring that back
and then make it reusable by next year.

00:25:16.600 --> 00:25:20.357
Um but the thing that's important about
two-stage reusability when it comes to

00:25:20.440 --> 00:25:22.917
the economics for orbital compute,

00:25:23.000 --> 00:25:25.917
right, is the cost per kg comes down

00:25:26.000 --> 00:25:27.517
significantly. You know, we're talking

00:25:27.600 --> 00:25:29.957
about going from $1,500 per kg on

00:25:30.040 --> 00:25:33.477
Falcon, somewhere in that range, to 250

00:25:33.560 --> 00:25:38.517
per kg, something lower. Um and the
more that you can reuse the rocket,

00:25:38.600 --> 00:25:39.757
the more that price comes down.

00:25:39.840 --> 00:25:43.277
>> Right. >> Right, cuz you're just
depreciating the cost of the launch.

00:25:43.360 --> 00:25:46.237
And eventually, you asymptote
to the cost of the fuel.

00:25:46.320 --> 00:25:49.397
>> Right. >> Right. Assuming
you can use a rocket for forever.

00:25:49.480 --> 00:25:51.477
>> Yes. >> Right, which
will take a very long time

00:25:51.560 --> 00:25:54.717
for us to to really achieve that.
But, um and at that point, we're talking

00:25:54.800 --> 00:25:58.037
about something well south of 250 per kg.

00:25:58.120 --> 00:26:03.037
So, then you look at the specs of these
AI satellites. You know, Elon did a great

00:26:03.120 --> 00:26:05.957
>> Yeah, that pod that that pod
was incredible that he laid out the other

00:26:06.040 --> 00:26:07.677
day, the specs on the satellites.

00:26:07.760 --> 00:26:10.117
>> It was really great
because I think they

00:26:10.200 --> 00:26:12.917
are finally showing people, here's how

00:26:13.000 --> 00:26:17.317
you could viably design
one of these satellites.

00:26:17.400 --> 00:26:21.477
And how heavy is the satellite? How many
could you fit into a Starship launch?

00:26:21.560 --> 00:26:27.077
And when you back into the numbers,
you get to something like 5 MW of capacity

00:26:27.160 --> 00:26:29.237
per Starship launch. >> Right.

00:26:29.320 --> 00:26:33.077
>> There's 100 metric tons in one of those
Starships. So, you can back into the

00:26:33.160 --> 00:26:37.397
math of how much will it cost per gigawatt

00:26:37.480 --> 00:26:39.917
to launch these satellites into space.

00:26:40.000 --> 00:26:43.597
>> Right. >> Launch
this compute into space. Um

00:26:43.680 --> 00:26:45.037
and the math that you get to before you

00:26:45.120 --> 00:26:47.517
account for things like

00:26:47.600 --> 00:26:50.717
bad GPUs, bad satellites, right, these

00:26:50.800 --> 00:26:52.517
will all be things that happen.

00:26:52.600 --> 00:26:54.797
But the math you get to is it's about $5

00:26:54.880 --> 00:26:59.877
billion per gigawatt of CapEx
to put these in space.

00:26:59.960 --> 00:27:03.237
>> Right. >> For comparison,
terrestrially,

00:27:03.320 --> 00:27:04.997
talk about the switch gears, the

00:27:05.080 --> 00:27:08.517
generators, the transformers, the shell,

00:27:08.600 --> 00:27:10.477
getting the power, that today is about

00:27:10.560 --> 00:27:13.677
25 20 to 25 billion per gigawatt.

00:27:13.760 --> 00:27:19.997
So we're talking about a 5x reduction
in cost on half of your bill of materials

00:27:20.080 --> 00:27:21.157
>> Right. >> for the data center.

00:27:21.240 --> 00:27:22.437
>> Right. >> Which is a huge number.

00:27:22.520 --> 00:27:26.797
>> Yeah, just just very simply, I mean
just to say that put it cost $60 to put a

00:27:26.880 --> 00:27:30.237
gigawatt on the ground today.
And we'll call it

00:27:30.320 --> 00:27:34.437
35 of that is are the GPUs and the
silicon that's doing the training and

00:27:34.520 --> 00:27:38.677
the inference.
And 25 billion is the land, the shell,

00:27:38.760 --> 00:27:42.797
the power, and the cooling.
I would hypothesize that those elements

00:27:42.880 --> 00:27:44.157
are probably going to be inflationary,

00:27:44.240 --> 00:27:47.437
so that 25 billion may not go down.

00:27:47.520 --> 00:27:52.757
And because space, power, cooling are

00:27:52.840 --> 00:27:57.037
effectively free in space, and when
I say space, I mean land. You know,

00:27:57.120 --> 00:28:00.560
there's no land in space,
but there is space.

00:28:00.760 --> 00:28:04.557
Um you're you're talking about putting a

00:28:04.640 --> 00:28:07.117
gigawatt into space for 30 billion and

00:28:07.200 --> 00:28:11.877
having lower operating costs.
Now the dynamic versus 60 billion that's

00:28:11.960 --> 00:28:15.917
inflationary, and that third and that 30
billion, that five may be deflationary

00:28:16.000 --> 00:28:19.437
over time. But what we need to consider

00:28:19.520 --> 00:28:21.317
is you know, the reliability and the

00:28:21.400 --> 00:28:25.037
maintenance. And so as long as you know,

00:28:25.120 --> 00:28:26.717
everybody can do the math,

00:28:26.800 --> 00:28:30.117
but as long as these satellites in space

00:28:30.200 --> 00:28:33.517
aren't failing at an at an astronomical

00:28:33.600 --> 00:28:36.717
rate, the math maths. As you can see,

00:28:36.800 --> 00:28:40.917
and by the way, we know GPUs melt
and lasers fail. We know this happens in

00:28:41.000 --> 00:28:44.437
data centers, particularly
during big training runs.

00:28:44.520 --> 00:28:47.037
And yeah, I mean GPUs melt.

00:28:47.120 --> 00:28:50.237
Um so as long as the reliability and

00:28:50.320 --> 00:28:54.757
maintenance is not dramatically lower,
the math is there once we have

00:28:54.840 --> 00:28:57.757
reusability and then rapid
reusability for Starship V3.

00:28:57.840 --> 00:29:00.317
>> I when you when we look
at this, okay, so we we

00:29:00.400 --> 00:29:04.837
went through Starlink and we said,
"Okay, like it it just stands to reason

00:29:04.920 --> 00:29:07.677
we're going to have direct to cell
on Starlink." Like the assumptions there

00:29:07.760 --> 00:29:11.757
are, you know, again, seem like you can
get your head around. Then when it comes

00:29:11.840 --> 00:29:15.517
to building terrestrial data centers,
again, not a hard one to think that

00:29:15.600 --> 00:29:18.677
based on these couple deals that Elon's
going to build a much bigger Starlink's

00:29:18.760 --> 00:29:22.437
going to build or SpaceX is going
to build a much bigger business there. And

00:29:22.520 --> 00:29:25.837
then you have this call option on space
that would drop the price even further.

00:29:25.920 --> 00:29:27.877
The one thing we haven't talked about is

00:29:27.960 --> 00:29:32.997
their model, right? And I find this
surprising, right? Six Six months ago,

00:29:33.080 --> 00:29:36.117
x.ai was competing, they were doing
pretty well, but they've done something

00:29:36.200 --> 00:29:38.797
dramatic over the course
of the past couple

00:29:38.880 --> 00:29:40.037
couple months, which is they bought

00:29:40.120 --> 00:29:44.557
Cursor, right? Cursor is 700 800 people

00:29:44.640 --> 00:29:48.837
was already doing incredibly well
from a revenue perspective. Our own

00:29:48.920 --> 00:29:52.557
projections were that they could
exit this year at up to $10 billion

00:29:52.640 --> 00:29:56.117
of revenue, so they were growing very fast

00:29:56.200 --> 00:29:59.837
one of the leading coding agents, but
they also had this incredible team with

00:29:59.920 --> 00:30:03.237
the potential, right, to really build
a frontier level model, but they were

00:30:03.320 --> 00:30:05.557
compute constrained. So all of a sudden,

00:30:05.640 --> 00:30:07.837
they get bought by X. X has massive

00:30:07.920 --> 00:30:10.917
compute that they can now train on.

00:30:11.000 --> 00:30:13.877
And when I think about the revenue in AI

00:30:13.960 --> 00:30:15.797
that like if I look at that line item in

00:30:15.880 --> 00:30:18.557
the models having it go from $10 billion

00:30:18.640 --> 00:30:23.197
to $150 billion, yes, a lot of that will
be the core weave type business that

00:30:23.280 --> 00:30:28.277
they have, but the question is how much
of that is going to be the core x.ai

00:30:28.360 --> 00:30:31.877
business that's really powered by the new
team from Cursor. So any thoughts on

00:30:31.960 --> 00:30:34.997
that, Kevin? >> Right now,
so Composer 2.5 was Pareto

00:30:35.080 --> 00:30:39.637
dominant 12 days ago. It was trained
on the Kimmy K2.5 base model.

00:30:39.720 --> 00:30:43.437
>> Right. >> Now, what's
happening is the Grok 4.3

00:30:43.520 --> 00:30:46.717
1.5 trillion parameter model is training.

00:30:46.800 --> 00:30:48.997
One would hypothesize based on scaling

00:30:49.080 --> 00:30:51.957
laws that that will might be a better

00:30:52.040 --> 00:30:56.957
base model. And then the cursor data
is being injected into the pre-training

00:30:57.040 --> 00:30:59.677
process, not just reinforcement learning.

00:30:59.760 --> 00:31:03.597
And we'll see, and I think that is going
to be a very important data point when

00:31:03.680 --> 00:31:07.597
that comes out. And I just think
everyone should keep in mind that once

00:31:07.680 --> 00:31:12.597
you are at multiple places on that
Pareto curve, if you have compute, you

00:31:12.680 --> 00:31:16.517
can scale really rapidly. >> You know,
that that to me is if I had to

00:31:16.600 --> 00:31:20.437
say what the one piece that's
being lost in the story,

00:31:20.520 --> 00:31:23.277
right? Like it's easy for everybody
to get excited about the deals with

00:31:23.360 --> 00:31:26.597
Anthropic because you can put your hands
around that. You know how much revenue

00:31:26.680 --> 00:31:30.517
it is. I see debate about, you know,
the 90-day termination and how long they

00:31:30.600 --> 00:31:33.917
last and what multiple do you put
on those revenues. But I think the thing

00:31:34.000 --> 00:31:38.477
that's getting lost is I think they've
dramatically advanced their capability

00:31:38.560 --> 00:31:42.157
when it comes to building a frontier
model. People outside Silicon Valley may

00:31:42.240 --> 00:31:45.877
not know, you know, Michael and the
team at Cursor as well. This is an

00:31:45.960 --> 00:31:48.357
extraordinary team that he just

00:31:48.440 --> 00:31:53.117
downloaded, right, into SpaceX. SpaceX
was already building good models. And

00:31:53.200 --> 00:31:56.797
what they have is they have
this way to monetize compute

00:31:56.880 --> 00:32:00.997
that gives you this call option that you
can pull all that compute in-house,

00:32:01.080 --> 00:32:04.397
right, to train a model and then to run
the model. I suspect if there's an

00:32:04.480 --> 00:32:08.477
upside surprise, if we went around the
table, I'd say this is the place that's

00:32:08.560 --> 00:32:11.437
getting the least amount of attention
and could have the biggest upside

00:32:11.520 --> 00:32:14.477
surprise. Any any thoughts, Clark, on

00:32:14.560 --> 00:32:18.997
what you think is being overlooked
or areas that you think are misunderstood

00:32:19.080 --> 00:32:23.477
about the business today?
>> I I would say I would say

00:32:23.560 --> 00:32:28.437
what the last few weeks
have proven is that Elon, um,

00:32:28.520 --> 00:32:31.917
their team can stand up all this compute.
Actually, if you just, you

00:32:32.000 --> 00:32:36.317
know, went back 1 and 1/2 years,
you know, they were behind in the race to

00:32:36.400 --> 00:32:40.797
stand up compute. They were you know
they they didn't have that many H100s.

00:32:40.880 --> 00:32:44.717
They brought in Colossus. Then they
brought in Colossus 2 at a scale much

00:32:44.800 --> 00:32:49.237
larger than anyone else.
And now you know as we gear for Vera Rubin

00:32:49.320 --> 00:32:52.197
you know from you know
a lot of my conversations

00:32:52.280 --> 00:32:54.037
it looks like they've you know secured

00:32:54.120 --> 00:32:57.197
maybe up to 20% of Vera Rubin capacity

00:32:57.280 --> 00:33:00.197
especially in the early days of you know

00:33:00.280 --> 00:33:06.117
when you know these these chips are very
scarce that that they're going to have a

00:33:06.200 --> 00:33:11.037
a a lead on all of this because you know
people think that they can stand up this

00:33:11.120 --> 00:33:14.637
compute better. So I think they'll all
you know what what the last few weeks

00:33:14.720 --> 00:33:18.397
have actually shown
is that Elon you know Elon will

00:33:18.480 --> 00:33:23.237
take you know take a shot at hitting
the frontier but if it you know if for

00:33:23.320 --> 00:33:26.117
whatever reason um they

00:33:26.200 --> 00:33:28.557
they have over procured some capacity

00:33:28.640 --> 00:33:30.917
this is a very scarce asset that they've

00:33:31.000 --> 00:33:35.877
shown that they can monetize at actually
you know best in class margins and

00:33:35.960 --> 00:33:38.757
payback periods.
>> The irony is like you know

00:33:38.840 --> 00:33:41.717
you and I've been doing this long enough
to know I mean that's why Bezos built

00:33:41.800 --> 00:33:46.037
AWS. Right? He had to build
capacity for Black Friday.

00:33:46.120 --> 00:33:47.677
>> Yeah. >> Right?
But then the rest of the year he

00:33:47.760 --> 00:33:50.837
sat on all this capacity they had
to build and he figured out a really

00:33:50.920 --> 00:33:56.077
incredible way to monetize this. And by
the way investors at the time 2009 2010

00:33:56.160 --> 00:33:59.277
when he was building out
the capability around AWS hated it.

00:33:59.360 --> 00:34:01.117
>> Of course. >> Because he
was consuming all that free

00:34:01.200 --> 00:34:05.637
cash flow. My meanwhile he was digging
the biggest gold mine in the history of

00:34:05.720 --> 00:34:07.077
the world. One of the biggest.

00:34:07.160 --> 00:34:10.117
>> One of the biggest. >> Among them among
them at the time was probably the biggest.

00:34:10.200 --> 00:34:14.017
>> Yeah Google search might want to have
a we'll have a we'll have a discussion.

00:34:14.100 --> 00:34:16.197
>> [clears throat] >> By the way
I do think it is important.

00:34:16.280 --> 00:34:18.677
Grok 4.3 I think the cursor if they

00:34:18.760 --> 00:34:20.637
acquire it that may end up being very

00:34:20.720 --> 00:34:23.437
important. But Grok 4.3 was on the

00:34:23.520 --> 00:34:25.957
Pareto frontier and has of 10 or 12 days

00:34:26.040 --> 00:34:28.477
ago and this these things move fast. But

00:34:28.560 --> 00:34:33.677
most intelligent 500 billion parameter
model in the world. And they were on the

00:34:33.760 --> 00:34:36.157
frontier and there are four
companies on the frontier.

00:34:36.240 --> 00:34:39.997
xAI, SpaceX AI, Google one with Gemini

00:34:40.080 --> 00:34:43.837
3.1 Pro, and then the rest of it
was dominated by Anthropic and OpenAI. But

00:34:43.920 --> 00:34:46.717
they were on the Pareto frontier and now
we'll see what they do with Cursor.

00:34:46.800 --> 00:34:49.237
>> Yeah. Um I want to come
back to that in a second.

00:34:49.320 --> 00:34:50.597
>> way, man, I want
to ask you some questions.

00:34:50.680 --> 00:34:53.477
>> go go go. What do you think? So you
think the biggest source of potential

00:34:53.560 --> 00:34:55.917
upside is the model?
>> Yes. >> What do you think?

00:34:56.000 --> 00:34:58.757
>> I think that's the I think that's
the thing that's least talked about.

00:34:58.840 --> 00:35:01.957
>> Least talked about.
>> Right? And so, listen.

00:35:02.040 --> 00:35:06.037
When I look at the bull bear case on
the IPO, right? The bears are looking at

00:35:06.120 --> 00:35:09.237
last year's revenue.
Say it was $18 billion

00:35:09.320 --> 00:35:13.717
and they're looking at the forecast from
the banks of $160 billion, you know, 3

00:35:13.800 --> 00:35:15.957
years from now and they're saying,
"Listen, not many companies in the

00:35:16.040 --> 00:35:18.237
history of the world have basically 8x

00:35:18.320 --> 00:35:20.637
their revenue over 3 to 4 years." Right?

00:35:20.720 --> 00:35:25.877
So that's where, you know, I think and
people get nervous about the valuation.

00:35:25.960 --> 00:35:29.677
When I look at this, again,
when you break it down as an analyst first

00:35:29.760 --> 00:35:33.717
principles, part by part, which is what
I tried to do here, right? When you look

00:35:33.800 --> 00:35:35.957
at Starlink, it looks totally doable.

00:35:36.040 --> 00:35:39.957
When I look at what they're building
in AI compute terrestrially, looks totally

00:35:40.040 --> 00:35:41.997
doable over the course of next 3 years.

00:35:42.080 --> 00:35:45.757
When I look at the model itself after
the acquisition of Cursor, you know,

00:35:45.840 --> 00:35:49.397
combining those things around the compute
they have, that looks to me like

00:35:49.480 --> 00:35:53.637
it could be an upside surprise. So I
would say that I think that uh you know,

00:35:53.720 --> 00:35:57.517
in the IPO, but I think when you look
back 3 years from now, there's a decent

00:35:57.600 --> 00:35:59.517
chance that everybody's like, "Oh my

00:35:59.600 --> 00:36:02.117
god, that was super obvious." Right?

00:36:02.200 --> 00:36:06.397
Even though today all of these things
have risk associated and back to where

00:36:06.480 --> 00:36:10.397
we started. I'm not, you know, none
of us are here to pump the IPO at 1.77

00:36:10.480 --> 00:36:14.077
trillion. It's really to just break
it down as we do inside our shop and to

00:36:14.160 --> 00:36:17.237
say, "What is that distribution
of future probabilities? What's the

00:36:17.320 --> 00:36:19.117
probability that it's higher from here?

00:36:19.200 --> 00:36:22.997
What's the prob" And I think we're
all pretty AI pilled. And if you're AI

00:36:23.080 --> 00:36:26.437
pilled, that means we got to build a lot
more compute than the world thinks and

00:36:26.520 --> 00:36:29.797
that these models are going to be a lot
more valuable than people think. You

00:36:29.880 --> 00:36:33.677
combine that with their core business.
I don't know another entrepreneur or

00:36:33.760 --> 00:36:37.837
another business that's
a better bet on the future,

00:36:37.920 --> 00:36:41.677
right, than SpaceX. And so I think
for most institutional investors, it's a

00:36:41.760 --> 00:36:44.557
must buy, a must own, a set it and

00:36:44.640 --> 00:36:49.637
forget it, right, in order to have a real
bet on both the space and the AI future.

00:36:49.720 --> 00:36:51.037
>> From your lips to God's ears.

00:36:51.120 --> 00:36:54.717
>> I mean, listen, I I again, I think that
I I I think that you're going to have to

00:36:54.800 --> 00:36:58.397
wait, but you know, we had this chart
last week, right, that came out.

00:36:58.480 --> 00:37:00.117
Everybody was sending around Twitter,

00:37:00.200 --> 00:37:02.717
conveniently timed, and you know, it's

00:37:02.800 --> 00:37:05.837
like shows the average max drawdown post

00:37:05.920 --> 00:37:08.477
IPO for like 20 companies from Facebook,

00:37:08.560 --> 00:37:10.997
Twitter, Alibaba, Shopify is, you know,

00:37:11.080 --> 00:37:16.037
over 50%. And so maybe that again will
will will end this section here. You

00:37:16.120 --> 00:37:19.037
know, Gavin, you and I've been doing
this a long time. We know it's going to

00:37:19.120 --> 00:37:21.877
be bouncy around the IPO. Um,

00:37:21.960 --> 00:37:25.357
you know, how do you as a manager try to

00:37:25.440 --> 00:37:27.677
try to manage that? Um, do you try to

00:37:27.760 --> 00:37:32.677
trade around the IPO? Do you set it kind
of and forget it? I would say from an

00:37:32.760 --> 00:37:36.477
Altimeter perspective, what we tend
to do is we take a base position that we

00:37:36.560 --> 00:37:38.677
set and forget, right? And then we may

00:37:38.760 --> 00:37:41.397
size up or size down depending upon how

00:37:41.480 --> 00:37:43.877
the market reacts in, you know, in a

00:37:43.960 --> 00:37:46.917
particular moment. Um, but any thoughts

00:37:47.000 --> 00:37:51.717
on on this chart or you know,
how how people you guys are

00:37:51.800 --> 00:37:54.797
thinking about it in particular.
You obviously own a lot going into it.

00:37:54.880 --> 00:37:57.397
>> First agree with absolutely everything
you said and I actually think about it

00:37:57.480 --> 00:37:58.717
the same way, set it and forget it.

00:37:58.800 --> 00:38:01.957
You've talked about you have ballast,
you move around and you move the ballast

00:38:02.040 --> 00:38:05.077
to one side of the ship when you want
to the ship to lean into the wind to go

00:38:05.160 --> 00:38:08.157
faster and you move it to the other
side when you don't want the ship to tip

00:38:08.240 --> 00:38:09.677
over. I think that's a great analogy.

00:38:09.760 --> 00:38:13.237
Think about all important
companies in the portfolio the

00:38:13.320 --> 00:38:16.277
same way. So 100% agree. I mean, this

00:38:16.360 --> 00:38:20.877
this chart is a bummer. What I would
say is, you know, this data on IPOs, but

00:38:20.960 --> 00:38:22.917
what I would just say is this is a

00:38:23.000 --> 00:38:24.677
really unprecedented situation.

00:38:24.760 --> 00:38:28.037
>> Yes. We've never had an IPO this big.

00:38:28.120 --> 00:38:32.717
We've never had an IPO that's going
to go into an index this quickly.

00:38:32.800 --> 00:38:36.477
We simply do not know how much selling

00:38:36.560 --> 00:38:42.197
there will be from investors.
I would hazard a guess. I mean, I'm I

00:38:42.280 --> 00:38:45.197
don't know. But Elon,
I don't think he needs

00:38:45.280 --> 00:38:48.997
liquidity and I think he
owns What does he own, Foxy?

00:38:49.080 --> 00:38:52.597
>> It's 50% >> 50% of the company.

00:38:52.680 --> 00:38:57.557
>> way, he's locked up for 365 days
or 366 days. So, we know he's not selling,

00:38:57.640 --> 00:38:59.677
right? >> So, I just
think it's an unprecedented

00:38:59.760 --> 00:39:02.357
situation and the right answer >> Yeah.

00:39:02.440 --> 00:39:05.917
>> is I don't know what's going to happen
in the short term. And the right answer

00:39:06.000 --> 00:39:09.197
that I would just, you know, encourage
every investor making their own decision

00:39:09.280 --> 00:39:12.917
is to just think exactly [clears throat]
the way you articulated it. We have

00:39:13.000 --> 00:39:15.837
these different levers. We have these
different variables. Think about each

00:39:15.920 --> 00:39:19.517
one of them from first principles.
Make your own decision.

00:39:19.600 --> 00:39:22.757
Do your own due diligence.
Be thoughtful. But, there are a lot of

00:39:22.840 --> 00:39:24.397
variables here and that it is a little

00:39:24.480 --> 00:39:27.157
funny to me that uh you know, it was 100

00:39:27.240 --> 00:39:31.677
times trailing TTM revenue. Well, after
the deals they signed, I think it's at

00:39:31.760 --> 00:39:33.917
39 times. >> That can change fast.

00:39:34.000 --> 00:39:36.397
>> So, they added $29 billion in a month.

00:39:36.480 --> 00:39:39.317
>> Yes. Now, it's >> [laughter] >> By
the way, have you ever seen that happen?

00:39:39.400 --> 00:39:42.677
>> Never. Never.
And you know, it just goes to show

00:39:42.760 --> 00:39:46.717
first um Elon is not
only a great engineer.

00:39:46.800 --> 00:39:50.117
He and Gwen and the team
are great at business.

00:39:50.200 --> 00:39:52.477
>> And Brad, >> They they
they understand what needs to

00:39:52.560 --> 00:39:56.997
be done to raise the capital to get
to the next phase. They have a long-term

00:39:57.080 --> 00:40:01.277
mission in the business. And so, to me,
again, what we saw in the course of the

00:40:01.360 --> 00:40:05.397
last few weeks with cursor, what we saw
with these deals that they cut, I don't

00:40:05.480 --> 00:40:09.757
know that any of the mag seven could
have moved that quickly to adjust the

00:40:09.840 --> 00:40:14.677
business that they did. It's
exceptionally entrepreneurial at scale,

00:40:14.760 --> 00:40:16.637
which we very rarely see in businesses.

00:40:16.720 --> 00:40:19.397
Two other things I would
just say >> you a hug, Brad?

00:40:19.480 --> 00:40:23.557
>> Two Two other things I I I I would
just say. Number one is people talk a lot

00:40:23.640 --> 00:40:27.317
about the total amount of capital being
raised. If you add up the capital here,

00:40:27.400 --> 00:40:32.517
right, for Anthropic what they may raise,
what OpenAI may raise, what, you

00:40:32.600 --> 00:40:35.997
know, SpaceX may raise,
let's call it $250 billion.

00:40:36.080 --> 00:40:40.077
That's 1% of the Mag 7.
Okay, it's 1% of the Mag 7.

00:40:40.160 --> 00:40:43.437
>> I Yeah. And And we will
as well. You know, like that

00:40:43.520 --> 00:40:47.117
to me is like a bet on the future
that we all believe in. And so, if I said,

00:40:47.200 --> 00:40:50.917
"Where are we out of consensus? What
is our variant perception?" We actually

00:40:51.000 --> 00:40:53.877
think it's going to be bigger, faster,
and we've thought that for a couple

00:40:53.960 --> 00:40:56.757
years. Um so, first, it's only 1% of the

00:40:56.840 --> 00:41:01.357
Mag 7 market cap. And then you
referenced it, the amount of selling. Um

00:41:01.440 --> 00:41:05.037
I've got a chart we'll post here.
This is, you know, the the dribble share

00:41:05.120 --> 00:41:07.957
release for SpaceX shareholders. You

00:41:08.040 --> 00:41:12.517
know, so there's not a lot that can
be released um up until after the first

00:41:12.600 --> 00:41:16.917
earnings. This We saw this in the Cerebras
IPO. Um there's a version of it

00:41:17.000 --> 00:41:21.077
here in this IPO. And so, again, I think
the banks have been thoughtful here,

00:41:21.160 --> 00:41:23.437
knowing that this is a very large IPO.

00:41:23.520 --> 00:41:26.437
And I'm not saying that won't trade down.
Like there's possibility, you

00:41:26.520 --> 00:41:31.077
know, these things trade down.
But again, for me, telescope out, is there

00:41:31.160 --> 00:41:34.317
any company better positioned as a bet
on the future? I think what they've

00:41:34.400 --> 00:41:36.837
shown over the course
of last 5 weeks, they're

00:41:36.920 --> 00:41:39.797
they're they're probably number one.
But let's move on.

00:41:39.880 --> 00:41:42.237
>> No, no, can I just say one thing
about the employees? I think another thing

00:41:42.320 --> 00:41:45.837
that's unprecedented here
is the employees >> Yeah.

00:41:45.920 --> 00:41:50.277
>> and to a large degree the investors
here have had liquidity every 6 months.

00:41:50.360 --> 00:41:52.037
>> Exactly. >> the last 10 years.

00:41:52.120 --> 00:41:54.877
>> Yes. >> So,
if you're a SpaceX employee or

00:41:54.960 --> 00:41:57.397
former employee, and you wanted to sell

00:41:57.480 --> 00:42:02.597
you've had whatever that is,
close to 20 chances. And it is a matter

00:42:02.680 --> 00:42:04.157
of historical record that large

00:42:04.240 --> 00:42:08.477
investors have been able to sell. So

00:42:08.560 --> 00:42:11.677
I would think a lot
of the people >> they've

00:42:11.760 --> 00:42:14.637
>> chosen to own it. Now, there's a new
valuation and we'll see what they do,

00:42:14.720 --> 00:42:17.557
but just this is utterly
unprecedented and we'll see.

00:42:17.640 --> 00:42:20.317
>> Yeah, I know. It's It's It's a great
point. We have in fact called these

00:42:20.400 --> 00:42:25.157
companies quasi-public. Um you and I
both know that SpaceX and I'd put

00:42:25.240 --> 00:42:29.277
Anthropic in in in this category as well,
Databricks in this category. These

00:42:29.360 --> 00:42:32.117
things in many ways have been more
liquid over the course of the past 3

00:42:32.200 --> 00:42:36.077
years than some public biotech companies
we know. Right? And so there's a

00:42:36.160 --> 00:42:38.957
continuum of liquidity here. We We treat

00:42:39.040 --> 00:42:43.197
it as a binary, private versus public,
but it's really about this continuum.

00:42:43.280 --> 00:42:45.197
You know, let's keep going on models.

00:42:45.280 --> 00:42:47.517
You know, um Anthropic launched Fable 5,

00:42:47.600 --> 00:42:49.877
which you referenced um yesterday, which

00:42:49.960 --> 00:42:52.237
is basically Mythos um with some

00:42:52.320 --> 00:42:54.997
classifiers and safeguards um around

00:42:55.080 --> 00:42:59.397
cyber and biology, chemistry,
um and distillation. When those things

00:42:59.480 --> 00:43:01.277
get triggered, it fails back [snorts] to

00:43:01.360 --> 00:43:06.517
Opus 4.8. Um you know, there was a
Copart tweet about this yesterday. He

00:43:06.600 --> 00:43:09.597
said, you know, it sold on all
the benchmarks, but what really makes it

00:43:09.680 --> 00:43:13.477
special is long-running tasks. Okay? You

00:43:13.560 --> 00:43:15.397
retweeted our good friend, you know,

00:43:15.480 --> 00:43:18.997
Noam Brown. Um you know, ChatGPT 5.5

00:43:19.080 --> 00:43:21.677
also exhibited these capabilities.

00:43:21.760 --> 00:43:24.397
Um you know, it it led Noam, right, to

00:43:24.480 --> 00:43:27.117
suggest that it's not very relevant to

00:43:27.200 --> 00:43:29.077
do these snapshot benchmarks anymore.

00:43:29.160 --> 00:43:31.957
Yeah, like the x-axis has to be time or

00:43:32.040 --> 00:43:34.517
tokens or compute because we can solve

00:43:34.600 --> 00:43:36.797
most problems now if we just let these

00:43:36.880 --> 00:43:39.597
frontier models for a very long uh point

00:43:39.680 --> 00:43:42.877
in time. So, Gavin, what is this new

00:43:42.960 --> 00:43:46.357
class of model, right, Fable Fable 5,

00:43:46.440 --> 00:43:49.717
ChatGPT 5.5? What does it mean for the

00:43:49.800 --> 00:43:51.597
race in superintelligence? Who's up?

00:43:51.680 --> 00:43:54.877
Who's down? Who's still on the frontier?

00:43:54.960 --> 00:43:59.317
Um give us your thoughts. >> I mean,
it's hard to say that Anthropic's not up.

00:43:59.400 --> 00:44:03.117
>> Yeah. >> Like after
the revenue numbers they've put up,

00:44:03.200 --> 00:44:05.517
after the Fable 5 release, and Mythos is

00:44:05.600 --> 00:44:07.317
evidently even better.

00:44:07.400 --> 00:44:10.397
But I just think that Gnome Brown post

00:44:10.480 --> 00:44:15.037
from yesterday,
polynomial, is so profound.

00:44:15.120 --> 00:44:19.197
And just the idea that we do not
know how smart these models are.

00:44:19.280 --> 00:44:22.277
And we made >> Say more about that.
Why don't we know how smart they are?

00:44:22.360 --> 00:44:25.957
>> Because nobody has
run Mythos for a year

00:44:26.040 --> 00:44:28.997
continuously. And we may never know how

00:44:29.080 --> 00:44:31.597
smart each generation of models actually

00:44:31.680 --> 00:44:36.197
is or was, but because we don't have
time to appropriately evaluate their

00:44:36.280 --> 00:44:39.557
intelligence before the next model
comes out. I mean, this is a profound

00:44:39.640 --> 00:44:42.757
statement. And just just imagine, okay?

00:44:42.840 --> 00:44:45.277
So, I always say like
when you think about FSD,

00:44:45.360 --> 00:44:48.157
just imagine a human being who never

00:44:48.240 --> 00:44:50.957
gets distracted, never gets tired, never

00:44:51.040 --> 00:44:55.357
talks on the phone in the car, never
drinks and drives, never yells at their

00:44:55.440 --> 00:44:59.557
kids, never has to go to the backseat
to give their baby a bottle.

00:44:59.640 --> 00:45:03.117
And like of course you would think that
over time that is superior to humans who

00:45:03.200 --> 00:45:06.637
are distracted. I don't know
how long How long can you

00:45:06.720 --> 00:45:08.597
think deeply about one topic, Brad?

00:45:08.680 --> 00:45:11.277
>> What do you Give me an hour. Give
me an [laughter] hour. Give me an hour.

00:45:11.360 --> 00:45:15.157
>> A BIT. THAT MAKES me
feel terrible cuz I think

00:45:15.240 --> 00:45:18.717
I can think deeply about one topic
continuously before having a stray

00:45:18.800 --> 00:45:22.397
thought enter my mind for like
maybe 5 minutes. Then I can

00:45:22.480 --> 00:45:26.517
come back to that.
Imagine if Albert Einstein

00:45:26.600 --> 00:45:29.837
had been able instead of,
you know, and maybe that maybe I

00:45:29.920 --> 00:45:34.037
maybe he could think for 3 hours at a
time. Clearly an exceptional intellect.

00:45:34.120 --> 00:45:38.557
But imagine Albert Einstein had just
thought about fundamental physics

00:45:38.640 --> 00:45:41.997
24 hours a day. He doesn't
have to eat, he doesn't have

00:45:42.080 --> 00:45:46.077
to sleep, he doesn't have to relax,
he doesn't drink, >> never gets old,

00:45:46.160 --> 00:45:48.957
>> never gets old, >> never
has diminished intelligence,

00:45:49.040 --> 00:45:53.037
>> and he thought for 1 year.
I mean, we might already, you know,

00:45:53.120 --> 00:45:55.317
>> have solved a lot
of these intractable problems.

00:45:55.400 --> 00:45:57.997
>> So, I just think
that's an extraordinary

00:45:58.080 --> 00:46:01.037
thought. And just my takeaway was

00:46:01.120 --> 00:46:05.317
however bullish I
was on compute before then,

00:46:05.400 --> 00:46:09.037
I'm just a lot more bullish. >> Right.
Right. Right. So, so, so that is

00:46:09.120 --> 00:46:11.397
a, you know, we saw when

00:46:11.480 --> 00:46:14.317
that was probably what really unlocked

00:46:14.400 --> 00:46:17.117
Opus 4.6. It was the first really

00:46:17.200 --> 00:46:22.277
long-running model that could maintain
that context, maintain that memory, um

00:46:22.360 --> 00:46:26.357
solve some of these longer-running
problems, right? For us, the signal was

00:46:26.440 --> 00:46:29.037
in January. We knew we felt like that

00:46:29.120 --> 00:46:33.757
was a big moment, but then when you
started to see the revenue go up, we

00:46:33.840 --> 00:46:37.877
knew that lots of people were voting
independently, that that was a profound

00:46:37.960 --> 00:46:41.677
moment that they became much,
much more useful. So,

00:46:41.760 --> 00:46:45.837
but one of the things that the consensus
going into this year, right? So, the big

00:46:45.920 --> 00:46:51.157
question going into this year was was
the AI revenue going to show up? Were we

00:46:51.240 --> 00:46:55.317
going to get to these thresholds of
intelligence that caused enterprises and

00:46:55.400 --> 00:46:57.757
consumers to use them more? And I think

00:46:57.840 --> 00:46:59.477
the consensus at the time, at least on

00:46:59.560 --> 00:47:02.477
this podcast, um the the the debate with

00:47:02.560 --> 00:47:05.237
with my with with with Bill was the

00:47:05.320 --> 00:47:09.797
open-source models, cheap tokens,
were catching up on the frontier, that

00:47:09.880 --> 00:47:11.637
perhaps these models were beginning to

00:47:11.720 --> 00:47:13.997
asymptote, um that people wouldn't

00:47:14.080 --> 00:47:16.877
really pay for premium tokens,

00:47:16.960 --> 00:47:19.317
and it seems to me that the evidence on

00:47:19.400 --> 00:47:21.637
the field, 6 months into the year, is

00:47:21.720 --> 00:47:26.997
just the opposite, right? That frontier
tokens are capturing the vast majority

00:47:27.080 --> 00:47:30.957
of all the revenues,
and that in fact, if you believe in the

00:47:31.040 --> 00:47:34.677
long-running capabilities and more
compute allows you to do that, they may

00:47:34.760 --> 00:47:37.117
actually be extending their lead, right?

00:47:37.200 --> 00:47:40.597
On some of these models that were built
on distillation. So, I just open it up

00:47:40.680 --> 00:47:42.397
to anyone around the table, what are

00:47:42.480 --> 00:47:45.477
your thoughts on whether or not, you

00:47:45.560 --> 00:47:48.477
know, have we challenged this thesis

00:47:48.560 --> 00:47:52.597
that cheap open-source tokens are going
to always, you know, close the gap on

00:47:52.680 --> 00:47:55.517
these frontier models,
or are they extending their leads?

00:47:55.600 --> 00:48:00.997
>> I I think this debate, like this same
debate has existed since the beginning

00:48:01.080 --> 00:48:05.157
of since we started training these
models to begin with, which was hey, we're

00:48:05.240 --> 00:48:09.597
always kind of three, six months behind
the frontier. But empirically, like you

00:48:09.680 --> 00:48:13.397
can just see all of the revenue has
actually just accrued at the frontier.

00:48:13.480 --> 00:48:14.997
And that I think that's because every

00:48:15.080 --> 00:48:21.477
time we release the frontier,
a whole new like slew of use cases

00:48:21.560 --> 00:48:24.517
>> Right. >> that that
previously we could have never

00:48:24.600 --> 00:48:28.637
tackled before, like coding.
Um but also just, you know, you know,

00:48:28.720 --> 00:48:33.037
we've we've just been locked at our desk
for the last last day just, you know,

00:48:33.120 --> 00:48:36.797
hammering Claude because, you know, it's
just fascinating the things that now we

00:48:36.880 --> 00:48:40.837
can do with fable five that we could
just couldn't do with opus 48 just a day

00:48:40.920 --> 00:48:42.517
before. >> So what are
some of those things, man?

00:48:42.600 --> 00:48:45.117
I'm curious. >> So So I think
it's really really good at

00:48:45.200 --> 00:48:47.757
multi-agent orchestration. So they they

00:48:47.840 --> 00:48:50.357
Anthropic released a um a blog post

00:48:50.440 --> 00:48:55.237
about like different uh agent um six
different agent like orchestration

00:48:55.320 --> 00:48:58.797
patterns that, you know, they've they've
talked about. But really like once you

00:48:58.880 --> 00:49:01.837
start being able to manage all these

00:49:01.920 --> 00:49:06.797
agents, the harness and the model itself
is being arled with one another, they're

00:49:06.880 --> 00:49:09.357
actually being, you know,
fused closer and closer

00:49:09.440 --> 00:49:13.637
together, but the model can understand
the, you know, the extent of your work.

00:49:13.720 --> 00:49:15.917
So, you know, one of the things, for

00:49:16.000 --> 00:49:21.077
instance, is um I just threw
in like seven of our models

00:49:21.160 --> 00:49:22.717
and just said, "Okay, like I want to

00:49:22.800 --> 00:49:26.397
create a master view of like my beliefs

00:49:26.480 --> 00:49:31.397
given all of these assumptions of all
these companies, TSMC capacity, like and

00:49:31.480 --> 00:49:35.757
then and then produce me a report on all
this stuff." And you know, the the model

00:49:35.840 --> 00:49:39.117
is able to reason through all
of our assumptions. Like actually, if you

00:49:39.200 --> 00:49:42.397
believe this >> Right.
What are the contradictions exactly?

00:49:42.480 --> 00:49:46.477
>> Yeah, it's it was fascinating. And and
and you know, before we'd never do that,

00:49:46.560 --> 00:49:50.277
but but now, you know, I think
we're just step one into multi-agent

00:49:50.360 --> 00:49:54.197
orchestration. We're going to do
this even further and that's one example.

00:49:54.280 --> 00:49:56.157
I've also dumped all my all my notes

00:49:56.240 --> 00:49:58.877
into it and it's reason across all my

00:49:58.960 --> 00:50:03.197
notes from the last 3 years and said,
you know, here are some of your ideas

00:50:03.280 --> 00:50:07.197
that were consistent. Here are like, you
know, the sources that were actually the

00:50:07.280 --> 00:50:11.717
highest signal to what actually played
out, you know, and then it is actually

00:50:11.800 --> 00:50:16.117
just super fascinating what you could
do and we've just blown through our blown

00:50:16.200 --> 00:50:18.477
through our limits. >> mean it's
it's it's unlocking all this.

00:50:18.560 --> 00:50:21.757
I mean like they gave examples yesterday
and the release Anthropic did, you know,

00:50:21.840 --> 00:50:24.077
50 million line Ruby code base at Stripe

00:50:24.160 --> 00:50:29.197
that was, you know, refactored in a day
versus many weeks with many people. You

00:50:29.280 --> 00:50:32.677
think about where this is impacting
biology and life sciences just across

00:50:32.760 --> 00:50:37.157
the spectrum. Um and to me
it really gets back to this

00:50:37.240 --> 00:50:39.877
fundamental point. Number one,
if you believe this to be true about

00:50:39.960 --> 00:50:44.117
long-running agents, then we're going
to produce and consume more tokens in the

00:50:44.200 --> 00:50:46.677
future as far as the eye can see. So the

00:50:46.760 --> 00:50:51.237
world this gets me back to, you know,
terrafab and space orbital and all this

00:50:51.320 --> 00:50:55.357
because we we we may in fact unlock real

00:50:55.440 --> 00:50:58.997
thresholds of intelligence, but we're
going to have to let these horses run

00:50:59.080 --> 00:51:00.877
for a long time in order to get there.

00:51:00.960 --> 00:51:03.757
Yeah, I would just say two
things I two things can be true.

00:51:03.840 --> 00:51:07.677
>> Mhm. >> The majority
of economic value may

00:51:07.760 --> 00:51:11.277
continue to accrue to the frontier and
man has it ever accrued to the frontier

00:51:11.360 --> 00:51:16.117
thus far and for sure the first 6 months
of this year, but the majority of tokens

00:51:16.200 --> 00:51:17.797
consumed in the world may be open source.

00:51:17.880 --> 00:51:18.797
>> And they are >> today.

00:51:18.880 --> 00:51:21.397
>> Yes. I and I think that this current

00:51:21.480 --> 00:51:24.997
state is likely to persist. Harvey had a

00:51:25.080 --> 00:51:28.357
great blog post that they put out on X

00:51:28.440 --> 00:51:33.477
and they used and it's just amazing how
everything gets out out of date like in

00:51:33.560 --> 00:51:37.077
5 days, you know. But they
used their own proprietary

00:51:37.160 --> 00:51:40.277
legal data to do reinforcement learning

00:51:40.360 --> 00:51:44.677
and supervised fine-tuning with
Fireworks on an open source model

00:51:44.760 --> 00:51:47.677
and then And used a router and a router
being something that picks which model

00:51:47.760 --> 00:51:51.837
you send which query to, and which model
you use to check which model. And they

00:51:51.920 --> 00:51:57.157
got better outcomes than Opus 4
either 4.7 or 4.8 at a lower cost.

00:51:57.240 --> 00:52:01.357
>> Yes. >> And I think that is
the future. And the reality is

00:52:01.440 --> 00:52:04.557
they were still consuming a lot of Opus,
but a majority of the tokens they were

00:52:04.640 --> 00:52:08.237
processing probably were
in their own open-source models.

00:52:08.320 --> 00:52:11.117
>> We heard the same thing.
We did a We did um

00:52:11.200 --> 00:52:15.917
We did an enterprise survey that we'll
post of 300 companies how which ones

00:52:16.000 --> 00:52:19.957
were optimizing, so these are folks who
are kind of looking at model routing and

00:52:20.040 --> 00:52:22.637
saying we're going to send certain
tokens over here, which ones are

00:52:22.720 --> 00:52:26.037
thinking about optimizing, which ones
aren't optimizing yet, and then what is

00:52:26.120 --> 00:52:30.317
their expected use of frontier model
tokens, right? And they're all expecting

00:52:30.400 --> 00:52:33.757
to consume a lot more even though
they're already in the process of

00:52:33.840 --> 00:52:37.517
optimizing. Think of it in the in in the
context of JP Morgan. If they're doing

00:52:37.600 --> 00:52:42.197
some back of the house stuff, right,
on customer service or whatever, they may

00:52:42.280 --> 00:52:46.117
very well use an open-source model.
Now, I think they're loath to use Chinese

00:52:46.200 --> 00:52:50.397
open-source models, so they're waiting
on kind of US open-source models to, you

00:52:50.480 --> 00:52:54.317
know, be able to really deliver the bang
that they need, but my hunch is for

00:52:54.400 --> 00:52:57.757
these enterprises, a lot of that back
of the house stuff will get rooted there.

00:52:57.840 --> 00:53:00.797
That will probably be a majority
of the tokens, but I think the really

00:53:00.880 --> 00:53:04.277
high-value stuff, you know, coding
as an example, they don't want to write

00:53:04.360 --> 00:53:07.797
second-tier code. I think the vast
majority of that will continue to be on

00:53:07.880 --> 00:53:11.117
the frontier. >> Um >> You don't
need Albert Einstein to book

00:53:11.200 --> 00:53:14.997
you a trip. You don't need
Albert Einstein to do KYC.

00:53:15.080 --> 00:53:18.717
>> But but but this is the debate we had
at literally at this table two years ago.

00:53:18.800 --> 00:53:23.597
However, if you just look at the revenue
curves, right? What bill What what folks

00:53:23.680 --> 00:53:25.557
concluded when they said that, they

00:53:25.640 --> 00:53:30.717
said, "Therefore, the frontier models
will not accrue most of the revenue."

00:53:30.800 --> 00:53:32.717
And what we're seeing
right now, it's 90% of the

00:53:32.800 --> 00:53:34.237
>> That has been
decisively wrong. Probably

00:53:34.320 --> 00:53:38.237
more than 90%, and it may continue to be
decisively wrong. Frontier might be 90%

00:53:38.320 --> 00:53:42.557
of the economic value.
Open-source >> might be 80% of tokens.

00:53:42.640 --> 00:53:45.437
Something that I think is very
important on open source

00:53:45.520 --> 00:53:47.917
is that you know, I think there's this

00:53:48.000 --> 00:53:50.157
belief that it's bearish for AI.

00:53:50.240 --> 00:53:53.917
It's actually it may be very bearish
for the frontier models. There's that bear

00:53:54.000 --> 00:53:58.597
case you talked about. It's actually
really bullish for compute and hardware

00:53:58.680 --> 00:54:02.677
because if the frontier models are
capturing less of the margin, then

00:54:02.760 --> 00:54:04.437
you're going to spend more on compute.

00:54:04.520 --> 00:54:08.517
So, the better open source does,
the better it is for compute providers.

00:54:08.600 --> 00:54:12.277
>> And I yeah, I I will
say it there is a very

00:54:12.360 --> 00:54:16.477
I would say between um spending time
in the heart of like the West, Silicon

00:54:16.560 --> 00:54:20.277
Valley, and also spending time
in Asia, there is like a very

00:54:20.360 --> 00:54:24.317
big um like a deep-seated
belief in one versus

00:54:24.400 --> 00:54:27.477
the other, which is like if you spend
a lot of time here, it's like all closed

00:54:27.560 --> 00:54:32.397
source, cloud, every all traffic
is going to go, you know, by way of this

00:54:32.480 --> 00:54:36.197
direction. And then you
spend time in Asia, you know,

00:54:36.280 --> 00:54:40.237
the the overwhelming belief is that
we're going to find the right model to

00:54:40.320 --> 00:54:43.117
the right workload,
and we're not going to overspend.

00:54:43.200 --> 00:54:46.157
>> Right. >> And I think,
you know, I would say I would say

00:54:46.240 --> 00:54:51.877
the next year is probably going to be
the most indicative of which way this

00:54:51.960 --> 00:54:55.237
falls um because

00:54:55.320 --> 00:54:59.517
I think I think the reason why uh
closed source models have captured so

00:54:59.600 --> 00:55:04.157
much of the value is because um
the models actually get the intention and

00:55:04.240 --> 00:55:07.237
actually carry through the work. And
this is the first year where we actually

00:55:07.320 --> 00:55:10.557
had agents that actually carried out

00:55:10.640 --> 00:55:15.437
user intention from just answering
a chatbot request to actually producing

00:55:15.520 --> 00:55:18.917
useful work. >> Right.
>> Um now the the the level of this

00:55:19.000 --> 00:55:22.917
intelligent has scaled so rapidly,
and we continue to push against like the

00:55:23.000 --> 00:55:27.077
most economically valuable tasks,
which are coding and finance and all these

00:55:27.160 --> 00:55:29.357
like knowledge work tasks. But like for

00:55:29.440 --> 00:55:32.477
the long tail of tasks, if open source

00:55:32.560 --> 00:55:37.637
continues to maintain a 6-month lag,
we might actually see a lot more open

00:55:37.720 --> 00:55:42.397
source used for you know,
our everyday tasks that we might actually

00:55:42.480 --> 00:55:43.877
>> basically Jensen's argument, right?

00:55:43.960 --> 00:55:46.597
Jensen's argument is you're
going to have model routing

00:55:46.680 --> 00:55:50.837
and we're just in a moment in time where
the frontier models gain the advantage

00:55:50.920 --> 00:55:54.397
can do long-running tasks that open
source models couldn't do it very well

00:55:54.480 --> 00:55:57.677
and so they're accruing all of the value,
but as soon as the open source

00:55:57.760 --> 00:56:01.677
models can do the long-running tasks
as well, which is not far away that they

00:56:01.760 --> 00:56:04.477
too will grab a bunch
a bunch of this revenue.

00:56:04.560 --> 00:56:06.717
>> Are you about to burst into reflection?

00:56:06.800 --> 00:56:08.957
>> I'm not. >> Okay. No,
no, no, no are we, but I'm

00:56:09.040 --> 00:56:12.237
very impressed by Misha and and the team

00:56:12.320 --> 00:56:14.237
and what they're doing. I very much want

00:56:14.320 --> 00:56:17.717
a frontier open source US lab to win. We

00:56:17.800 --> 00:56:21.757
know that, you know, I heard you say
recently and I believe it to be true

00:56:21.840 --> 00:56:25.557
Nvidia any day that they really wanted to,
right? They already have some great

00:56:25.640 --> 00:56:29.197
open source models. They could
absolutely build a frontier open source

00:56:29.280 --> 00:56:33.317
model whenever they chose to do it and so
it's not a question in my mind as to

00:56:33.400 --> 00:56:36.917
whether or not the US is going to have
a frontier open source model. It's just a

00:56:37.000 --> 00:56:42.317
question about timing and then like
at that point in time is that you know,

00:56:42.400 --> 00:56:46.157
let's say let's assume they get
these long-running capabilities.

00:56:46.240 --> 00:56:50.597
Have the frontier labs now achieved
something yet again that allows them to

00:56:50.680 --> 00:56:53.437
keep keep the the stranglehold
on the revenues?

00:56:53.520 --> 00:56:56.037
>> Yeah, and I just think it's if you're

00:56:56.120 --> 00:57:01.437
Wow, that's a cute ASIC you've built
there. That is so cute. How would you

00:57:01.520 --> 00:57:04.517
like open source to join the frontier?

00:57:04.600 --> 00:57:06.077
>> Right. >> How would you
like that? How do you like

00:57:06.160 --> 00:57:10.437
them apples? So, I mean I'm not sure
that's the explicit calculation, but I

00:57:10.520 --> 00:57:13.757
do think Jensen >> Say more. Just double
click on that for everybody at home.

00:57:13.840 --> 00:57:15.757
>> Yeah. >> If you were
if they were to put an open

00:57:15.840 --> 00:57:19.597
source model out there, how does
that impact the ASIC landscape?

00:57:19.680 --> 00:57:23.677
>> Well, you might not have the revenue

00:57:23.760 --> 00:57:26.797
to fund [laughter] to fund
that the revenue of the margins

00:57:26.880 --> 00:57:31.117
to fund that ASIC. And I do
think Nvidia is highly likely

00:57:31.200 --> 00:57:35.797
to be the world's dominant provider
of open source AI. And I do think Jensen

00:57:35.880 --> 00:57:40.437
will bring open source, you know,
right now it's whatever, 6

00:57:40.520 --> 00:57:42.237
months behind the frontier. >> Yeah.

00:57:42.320 --> 00:57:47.237
>> We might see it creep
closer and closer and closer.

00:57:47.320 --> 00:57:50.557
And I do think Jensen has a big business
decision. I see this, you know, chart

00:57:50.640 --> 00:57:53.837
here, so let's, you know,
chop it up about Nvidia, as you say.

00:57:53.920 --> 00:57:59.197
But if all of his customers are
going to compete with him, >> Yes.

00:57:59.280 --> 00:58:02.557
>> then why not compete
with his customers? And

00:58:02.640 --> 00:58:04.317
we have all these neo clouds. >> Right.

00:58:04.400 --> 00:58:08.077
>> So that's a cloud computing business
that can compete with all these cloud

00:58:08.160 --> 00:58:11.277
computing businesses.
He has his own models that are really,

00:58:11.360 --> 00:58:14.957
really good. Nematron 3 or 3.1
was actually really, really cool from a

00:58:15.040 --> 00:58:18.477
computer efficiency perspective.
And he's always careful to release small

00:58:18.560 --> 00:58:22.237
models so as to not tread
on Anthropic and OpenAI, >> Right.

00:58:22.320 --> 00:58:26.597
>> Google's toes. But I do think
that is a choice he is making.

00:58:26.680 --> 00:58:31.117
And just, you know, at if if
the economics change, >> Right.

00:58:31.200 --> 00:58:33.077
>> I think Nvidia can
join the frontier and

00:58:33.160 --> 00:58:37.077
become one of the world's largest cloud
computing companies much faster than

00:58:37.160 --> 00:58:41.397
people think. >> Interesting. Interesting.
Clark, walk us through this this chart.

00:58:41.480 --> 00:58:43.637
>> Yeah, so so I think
one of the takeaways

00:58:43.720 --> 00:58:47.077
from spending time in Taiwan was there

00:58:47.160 --> 00:58:48.877
there is certainly a lot of excitement

00:58:48.960 --> 00:58:51.957
around the next wave of ASICs.

00:58:52.040 --> 00:58:57.317
Um, but I think I think it's like
a very clear moment now where Nvidia

00:58:57.400 --> 00:59:01.197
it used to be an argument of Nvidia
versus ASICs one or the other and, you

00:59:01.280 --> 00:59:03.277
know, total domination one or the other.

00:59:03.360 --> 00:59:07.997
Now I think it increasingly every year
every every one assumed that Nvidia was

00:59:08.080 --> 00:59:12.877
going to lose share dramatically on a
revenue scale, on a gigawatt scale, on a

00:59:12.960 --> 00:59:16.517
unit scale. And actually, if you
actually look at the last few years, you

00:59:16.600 --> 00:59:21.797
know, they've actually maintained
their share very, very handsomely.

00:59:21.880 --> 00:59:26.397
Um, actually, um, if you accounted
for the fact that Anthropic

00:59:26.480 --> 00:59:28.317
was not really using Nvidia. They

00:59:28.400 --> 00:59:31.557
probably actually gain share against

00:59:31.640 --> 00:59:35.437
if not for in 25 26. So, I think I think

00:59:35.520 --> 00:59:39.517
what was very interesting
though was a new class of

00:59:39.600 --> 00:59:43.277
accelerators or ASICs. MediaTek with their

00:59:43.360 --> 00:59:48.557
with their new V8T versus,
you know, Broadcom's V8I for

00:59:48.640 --> 00:59:50.237
TPUs

00:59:50.320 --> 00:59:52.957
actually was a big topic of discussion.

00:59:53.040 --> 00:59:56.477
And, you know, I I think for ASICs

00:59:56.560 --> 00:59:58.357
the argument now is that more and more

00:59:58.440 --> 01:00:01.317
will look custom to the actual workload

01:00:01.400 --> 01:00:03.597
and that is like one vector that people

01:00:03.680 --> 01:00:06.477
are moving in versus Nvidia now is

01:00:06.560 --> 01:00:09.557
has kind of shown itself as the the

01:00:09.640 --> 01:00:12.677
predominant provider of compute to

01:00:12.760 --> 01:00:17.357
a lot of the world and for,
you know, internal internal workloads,

01:00:17.440 --> 01:00:21.117
perhaps they will go more and more
custom and more and more down the stack.

01:00:21.200 --> 01:00:25.997
And I I remember just, you know, 1 year
ago when it was kind of a Broadcom or

01:00:26.080 --> 01:00:27.837
Nvidia battle. It seems there's a lot

01:00:27.920 --> 01:00:30.557
more nuance now to, you know, what type

01:00:30.640 --> 01:00:33.197
of accelerators will fit which workloads

01:00:33.280 --> 01:00:37.837
and fit which customers and fit
which business models. Um

01:00:37.920 --> 01:00:43.717
and yeah, I thought I thought
that was a a new topic.

01:00:43.800 --> 01:00:45.757
>> It's actually >> New
realization though, I think we all

01:00:45.840 --> 01:00:47.957
kind of shared this view for a long time.

01:00:48.040 --> 01:00:50.957
>> Yeah, I was just shocked. I mean, I'm
I'm out here. I did a board meeting with

01:00:51.040 --> 01:00:54.397
one of our companies and just,
you know, their biggest one

01:00:54.480 --> 01:00:59.157
thing they emphasized is we
thought the world would be have be

01:00:59.240 --> 01:01:01.397
consuming less Nvidia than it is and if

01:01:01.480 --> 01:01:06.237
anything, Nvidia is accelerating and they
just continue to out execute their

01:01:06.320 --> 01:01:10.277
competitors. And I think a lot
of people are indexing to this

01:01:10.360 --> 01:01:14.237
OpenAI gigawatt and you
know, Nvidia has 10.

01:01:14.320 --> 01:01:17.277
Broadcom has 10. Um

01:01:17.360 --> 01:01:21.277
who has six? AMD AMD has
six and they have warrants.

01:01:21.360 --> 01:01:25.797
And then Cerebras has
our shared portfolio company

01:01:25.880 --> 01:01:29.877
has a gigawatt.
And I just that is what's on paper.

01:01:29.960 --> 01:01:33.317
>> Right. >> What actually
gets deployed, let's see.

01:01:33.400 --> 01:01:37.477
I will be very surprised if you know
that 10 out of 27, what's that math?

01:01:37.560 --> 01:01:40.717
Let's see who's best at math.
What percentage market share is that?

01:01:40.800 --> 01:01:44.077
>> 30% yeah. >> Yeah.
I'll be very surprised if that is

01:01:44.160 --> 01:01:48.637
where they land. I think that
is an extremely unlikely outcome.

01:01:48.720 --> 01:01:53.037
And especially as long as we're in a
watt constrained world, if you can get

01:01:53.120 --> 01:01:54.877
more tokens per watt, which is literally

01:01:54.960 --> 01:01:57.397
revenue with Nvidia

01:01:57.480 --> 01:02:03.157
than a lot of alternatives just if you
build your factory with another chip

01:02:03.240 --> 01:02:05.637
you may save some money, but you're
going to have less revenue and the

01:02:05.720 --> 01:02:09.037
margins may be lower and that's a point
that Jensen keeps hammering and I think

01:02:09.120 --> 01:02:12.997
is a really important.
And by the way, credit where credit is due

01:02:13.080 --> 01:02:16.997
the most important the most one of the
most surprising things to me in this

01:02:17.080 --> 01:02:20.997
ASIC landscape >> I'd say
Meta and Microsoft have been

01:02:21.080 --> 01:02:22.557
probably disappointing. >> Yes.

01:02:22.640 --> 01:02:23.757
>> You know who made a good ASIC?

01:02:23.840 --> 01:02:24.997
>> Yes. >> Well, I know you know.

01:02:25.080 --> 01:02:26.677
>> Yes. >> Jalapeno >> Yeah, exactly.

01:02:26.760 --> 01:02:28.917
>> from Open AI. They made a great chip.

01:02:29.000 --> 01:02:31.677
>> Yes. >> Now,
unfortunately needs to run at a

01:02:31.760 --> 01:02:34.997
much lower temperature than the Nvidia
GPUs, which means you need to spend more

01:02:35.080 --> 01:02:37.677
money on cooling and that consumes
more power. They made a great chip.

01:02:37.760 --> 01:02:39.517
>> Well, we can I mean
I think the question

01:02:39.600 --> 01:02:42.197
there and the question for everybody is
going to be is that the highest and best

01:02:42.280 --> 01:02:45.117
use of your time? Right? Like I you

01:02:45.200 --> 01:02:49.317
know, I tend to think that the frontier
companies like there's this belief that

01:02:49.400 --> 01:02:52.757
they got to be vertical vertically
integrated. But if you believe like I do

01:02:52.840 --> 01:02:56.037
that the race to super intelligence
particularly as we get these recursive

01:02:56.120 --> 01:02:58.517
loops working may be over in the next

01:02:58.600 --> 01:03:03.677
two to three years, then I think focus
focus focus focus. You exist to build

01:03:03.760 --> 01:03:07.117
the best intelligence in the world and
to deliver the best intelligence in the

01:03:07.200 --> 01:03:10.717
world and you that means you have
to have all the revenue. Because if you

01:03:10.800 --> 01:03:13.717
want to build out the compute that's
going to be required to continue to push

01:03:13.800 --> 01:03:16.357
the frontier, you have to have
the revenue in order to support it. So I

01:03:16.440 --> 01:03:21.317
think you know, subject to the focus
question, I think they certainly did.

01:03:21.400 --> 01:03:24.837
This all brings me back to kind
of a reality check, though.

01:03:24.920 --> 01:03:28.357
Um you know, we just got done talking
about test time compute, inference time

01:03:28.440 --> 01:03:32.077
compute, long-running agents. This is
really the thing that's unlocked the

01:03:32.160 --> 01:03:34.557
revenue this year. Um it all pushes us

01:03:34.640 --> 01:03:39.197
in the direction of more CapEx.
Google just raised $80 billion,

01:03:39.280 --> 01:03:41.437
right? We've now taken the Mag 5 or Mag

01:03:41.520 --> 01:03:44.037
7 free cash flow, you know, down

01:03:44.120 --> 01:03:47.077
dramatically, 80% um from just a few

01:03:47.160 --> 01:03:49.917
years ago. Um and Morgan Stanley, you've

01:03:50.000 --> 01:03:51.557
got this chart in front of you, up to

01:03:51.640 --> 01:03:56.237
their 2027 CapEx forecast from 950

01:03:56.320 --> 01:03:59.757
billion to 1.1 trillion. I mean, we were
talking about this with Jensen. That was

01:03:59.840 --> 01:04:03.797
his forecast 2 years ago. You know,
obviously, this doesn't even include

01:04:03.880 --> 01:04:09.277
SpaceX, CoreWeave, etc. So, I think
the number on 2027 is likely closer to 1.5

01:04:09.360 --> 01:04:12.677
trillion. And if we
compare this to the total

01:04:12.760 --> 01:04:16.357
incremental inference revenue,
so the thing that the market gets worried

01:04:16.440 --> 01:04:18.437
about, you know, back to my Sam Altman

01:04:18.520 --> 01:04:20.797
podcast, you know, in October of last

01:04:20.880 --> 01:04:23.597
year, can we really afford to spend 1.5

01:04:23.680 --> 01:04:28.037
trillion of CapEx a year if we're
only generating X amount in inference

01:04:28.120 --> 01:04:29.837
revenue? The thing I think that lit the

01:04:29.920 --> 01:04:32.637
fuse this year was Anthropic showed up

01:04:32.720 --> 01:04:35.557
in a major way with revenue, right? And

01:04:35.640 --> 01:04:38.037
so, we have, you know, the AI lab

01:04:38.120 --> 01:04:40.637
revenue everybody combined at around

01:04:40.720 --> 01:04:44.797
$300 billion next year, right? So, can't

01:04:44.880 --> 01:04:48.397
you know, and go roll that out to 2027

01:04:48.480 --> 01:04:50.757
uh or that is 2027, 300 billion. So,

01:04:50.840 --> 01:04:53.117
we're spending 1.5 trillion of CapEx on

01:04:53.200 --> 01:04:55.677
300 billion of inference revenue. Does

01:04:55.760 --> 01:04:58.037
that math math for you? And what would

01:04:58.120 --> 01:05:02.837
cause you, you know, to to get more
nervous again about our ability to

01:05:02.920 --> 01:05:04.157
continue to make these investments?

01:05:04.240 --> 01:05:07.837
Because the second we get nervous about
it, the entire semi complex is going to

01:05:07.920 --> 01:05:10.637
come down a lot. Well, what do you
think the gross margins are on that 300

01:05:10.720 --> 01:05:13.277
billion? Yeah, let's call it 50%.

01:05:13.360 --> 01:05:16.317
>> I I would guess they're probably
a little bit higher than that. I might say

01:05:16.400 --> 01:05:20.157
60 or 70. But, I mean,
that math starts to math,

01:05:20.240 --> 01:05:24.077
and what I would just say is I
think that 300 billion is low, man.

01:05:24.160 --> 01:05:25.797
>> Yeah. Yeah. >> I just think it's low.

01:05:25.880 --> 01:05:29.557
>> From your mouth to God's >> Yeah,
yeah, exactly. I think I think we

01:05:29.640 --> 01:05:34.437
end this year well over 200 billion
in inference revenue, well over.

01:05:34.520 --> 01:05:38.357
And so, I think the math really maths,
and I do think we have to give uh Jensen

01:05:38.440 --> 01:05:42.357
>> Yeah, our friend.
>> some credit because he said some things

01:05:42.440 --> 01:05:44.197
that seemed outlandish. >> Right.

01:05:44.280 --> 01:05:49.037
>> And he was conservative. He was low.
He said a trillion 2 years ago.

01:05:49.120 --> 01:05:50.677
And I mean, he was really low.

01:05:50.760 --> 01:05:52.717
>> Right. >> And so, like,
let's give the guy some

01:05:52.800 --> 01:05:55.317
credit and think about
what he is saying right now.

01:05:55.400 --> 01:05:58.077
>> For sure, for sure. And and listen,

01:05:58.160 --> 01:06:02.637
I would say consistently,
Elon's been taking the over.

01:06:02.720 --> 01:06:04.917
Sundar's been taking the over.

01:06:05.000 --> 01:06:10.277
Sam, Dario, you know, Dario did
the podcast with Dwarkesh when he was

01:06:10.360 --> 01:06:13.437
talking about country geniuses in the
data center. He said that will be here

01:06:13.520 --> 01:06:16.477
by 2028. He said revenues will go into

01:06:16.560 --> 01:06:19.397
the low hundreds of billions by 2028.

01:06:19.480 --> 01:06:23.397
So, let's call that, you know, 3 400
billion of revenue by 2028. And he said

01:06:23.480 --> 01:06:27.717
that a while ago now, so he may even be
revising up his number. And he said it's

01:06:27.800 --> 01:06:31.317
hard for me to see that there won't be
trillions of dollars in revenue before

01:06:31.400 --> 01:06:34.317
2030. And if you're on that revenue

01:06:34.400 --> 01:06:36.517
trajectory, if we're on a trajectory to

01:06:36.600 --> 01:06:38.757
200 by the end of this year, let's call

01:06:38.840 --> 01:06:42.277
it 4 or 500 by next year, and a path to

01:06:42.360 --> 01:06:46.357
trillion plus by 2029,
then the math maths.

01:06:46.440 --> 01:06:50.837
>> And we got to keep in mind
that half of the spending is there to

01:06:50.920 --> 01:06:53.237
you know, for training, maybe a little
less than half. What is it, Foxy?

01:06:53.320 --> 01:06:55.917
>> It's probably that depends on the lab,
but it's I would say it's increasingly

01:06:56.000 --> 01:06:59.637
less than half. >> Yes. >> Okay.
So, we'll call it 35% is spending

01:06:59.720 --> 01:07:02.717
that's not revenue generating, but it's
going to kind of make the next model.

01:07:02.800 --> 01:07:04.677
So, I think the math maths. >> Right.

01:07:04.760 --> 01:07:08.117
>> And there's still this prisoner's
dilemma where if you opted out, that may

01:07:08.200 --> 01:07:11.837
be an existential decision.
>> And I think like coming into this year,

01:07:11.920 --> 01:07:15.477
going back to this kind of what
narratives were violated, you know,

01:07:15.560 --> 01:07:20.237
I think into this year everyone expected
token pricing, uh the price of compute,

01:07:20.320 --> 01:07:24.277
it's all deflationary. And it will be
kind of a smooth line deflationary over

01:07:24.360 --> 01:07:26.917
time. But, I think this year what we've

01:07:27.000 --> 01:07:31.437
seen is the opposite. And you know,
it's all comes back to supply-demand. The

01:07:31.520 --> 01:07:35.517
demand side of the equation seems to be
far outstripping the supply. Right? And

01:07:35.600 --> 01:07:38.637
I think you look at the
deals signed by SpaceX

01:07:38.720 --> 01:07:44.037
and others, the monetization
rates per watt are increasing.

01:07:44.120 --> 01:07:46.237
Um and

01:07:46.320 --> 01:07:48.677
look, that is on a a pretty nascent

01:07:48.760 --> 01:07:54.117
small base of users, right? Like Alex
at Well Rock, he has this great um

01:07:54.200 --> 01:07:55.557
way to frame it.

01:07:55.640 --> 01:08:01.437
Less than 0.2% of people on Earth are
actually using AI in an agentic way.

01:08:01.520 --> 01:08:04.077
>> Right. >> Right?
Like I'm not a technical person,

01:08:04.160 --> 01:08:10.477
but I'm consuming 500 CPU cores
in a VM instance, five GPUs 24/7.

01:08:10.560 --> 01:08:12.477
>> Yeah. >> I mean,
if you draw that out to any

01:08:12.560 --> 01:08:17.157
meaningful percentage of the population,
I mean, we're going to be in, you know,

01:08:17.240 --> 01:08:20.517
this kind of shortage environment
maybe for some time. So,

01:08:20.600 --> 01:08:23.717
I think that is all positive
for this ROI question.

01:08:23.800 --> 01:08:26.817
>> Man, foxy, 100 to one CPU to GPU ratio.

01:08:26.900 --> 01:08:30.037
>> [laughter] >> Kind of agentic workflow.

01:08:30.120 --> 01:08:33.157
>> He said of course.
>> [laughter] >> Five.

01:08:33.240 --> 01:08:35.717
>> Five, yes. >> Yeah.
>> I'm being smart with my phone.

01:08:35.800 --> 01:08:37.757
>> Good, good, good. Excellent.

01:08:37.840 --> 01:08:40.597
>> I I will say also that ratio of 300 to

01:08:40.680 --> 01:08:44.237
1. You know, call it 1.2, 1.5.

01:08:44.320 --> 01:08:47.677
Um there there is also a rate that now

01:08:47.760 --> 01:08:50.237
physically we can only expand

01:08:50.320 --> 01:08:55.077
how much we can produce and how much
we can actually increase that spend by,

01:08:55.160 --> 01:08:59.837
whereas we're seeing the opposite right
now on the on the the willingness to pay

01:08:59.920 --> 01:09:03.317
for these tokens. And actually like when
the willingness to pay for these when

01:09:03.400 --> 01:09:04.997
the monetization per gigawatt is

01:09:05.080 --> 01:09:07.477
actually increasing from, you know, call

01:09:07.560 --> 01:09:11.477
it like 20 20 billion um in the in the

01:09:11.560 --> 01:09:16.837
best best of cases for at the beginning
of the year to now like 30 to even

01:09:16.920 --> 01:09:20.317
pushing 40 >> per
gigawatt >> per gigawatt.

01:09:20.400 --> 01:09:24.877
Um all of that is is a very heavy fixed
cost base, but all of that is like pure

01:09:24.960 --> 01:09:26.517
margin flow through now. And you're

01:09:26.600 --> 01:09:29.437
actually, you know, as we scale like the

01:09:29.520 --> 01:09:33.877
willingness to pay for for all of this
and and now the all of this stipulated

01:09:33.960 --> 01:09:36.957
by like, you know, everything we're
talking about of like how much is open

01:09:37.040 --> 01:09:40.717
source versus not and all of these
different flows, but really like as

01:09:40.800 --> 01:09:43.077
we're climbing this curve, you know, the

01:09:43.160 --> 01:09:48.157
the the revenue is might actually
outstrip our fixed cost base by by

01:09:48.240 --> 01:09:52.437
significant amount. And I think that's
why all the labs are pushing, you know,

01:09:52.520 --> 01:09:57.037
the the the gas to the pedals because
they all they all see like within if we

01:09:57.120 --> 01:09:59.957
continue this curve within like 3 years,
you know, we're just going to be so

01:10:00.040 --> 01:10:04.197
short on all the computer >> It's
a great I'm sorry, but I mean I

01:10:04.280 --> 01:10:07.837
like it's a great point. Like if you
thought you were getting a when you made

01:10:07.920 --> 01:10:11.917
these decisions >> Yes.
>> in November of 2025,

01:10:12.000 --> 01:10:13.477
you thought you were
getting a certain return.

01:10:13.560 --> 01:10:17.197
>> Yeah. >> You may be getting
triple that return today.

01:10:17.280 --> 01:10:20.677
>> At Tropic, no way no way did they think
they were going to be anywhere close to

01:10:20.760 --> 01:10:23.557
break even. >> Yeah. >> Right?
And and and and in this part of

01:10:23.640 --> 01:10:28.397
the curve, and the reason like I I I
called it accidental profitability that,

01:10:28.480 --> 01:10:30.917
you know, people have been talking about
that because they want to spend a lot

01:10:31.000 --> 01:10:33.797
more money on computer. They just had
a hard time doing it. Now maybe with

01:10:33.880 --> 01:10:37.117
SpaceX, you know, they could take some
of those dollars and and and go spend

01:10:37.200 --> 01:10:39.317
them other places. But that to me is,

01:10:39.400 --> 01:10:42.717
you know, a a fundamental change. Um the

01:10:42.800 --> 01:10:46.397
first argument against the frontier
labs was they'll never generate revenue.

01:10:46.480 --> 01:10:48.677
Okay? And then we that got blown up.

01:10:48.760 --> 01:10:51.237
Then it was like even if they generate
revenue, it'll be really shitty gross

01:10:51.320 --> 01:10:54.677
margins, and they'll never be able
to get make money. And then kind of that

01:10:54.760 --> 01:10:58.477
that's blown up. And and you know,
I think now, you know, people are falling

01:10:58.560 --> 01:11:02.277
back and they're saying, "Well, they're
overcharging. This is token maxi." My

01:11:02.360 --> 01:11:05.877
good friend, you know, Chamath has said
there's no ROI on any of this spend.

01:11:05.960 --> 01:11:10.917
It's all this token maxi. My best
evidence for why we all know, of course,

01:11:11.000 --> 01:11:14.557
when somebody puts on this much spend
like at Altimeter, we're not optimally

01:11:14.640 --> 01:11:19.197
spending every single dollar. But,
the question is, why are millions of

01:11:19.280 --> 01:11:23.637
independent businesses, small, medium,
and large, why are millions of consumers

01:11:23.720 --> 01:11:25.717
all choosing to do the same thing?

01:11:25.800 --> 01:11:30.397
They're not dumb. These are, you know,
rational economic actors that are all

01:11:30.480 --> 01:11:34.237
simultaneously saying, "I want to do
this because it makes my life better. It

01:11:34.320 --> 01:11:38.437
makes my business better, etc." To me,
that is the best evidence as to why I

01:11:38.520 --> 01:11:40.157
think this revenue can continue.

01:11:40.240 --> 01:11:44.317
>> Yeah. And Clark, I think like the point
you made is dead on cuz I mean, you want

01:11:44.400 --> 01:11:47.717
to own asset-heavy businesses
in inflationary environments, and token

01:11:47.800 --> 01:11:49.557
pricing is going up, and supply and

01:11:49.640 --> 01:11:52.037
demand is tightening, so totally agree.

01:11:52.120 --> 01:11:57.067
>> Um you know, as we begin to uh
find our way to the exit ramp and and

01:11:57.150 --> 01:12:00.757
[laughter] and wrap here, one of the
things I you know, you and I've been

01:12:00.840 --> 01:12:04.517
doing this for a long time, Gavin,
a couple decades. Um you may even sketch

01:12:04.600 --> 01:12:08.357
longer than me, even though I'm
a little bit older than you. Um

01:12:08.440 --> 01:12:12.877
you know, we have uh I always like to do
a market check, because I find a lot of

01:12:12.960 --> 01:12:16.557
time that analysts come on these things,
and they talk their, you know, talk

01:12:16.640 --> 01:12:19.117
their book, and you know, there are
a lot of people who listen to these

01:12:19.200 --> 01:12:20.997
things, retail investors and others.

01:12:21.080 --> 01:12:23.437
It's just kind of like, what do
we really think? And so, I always

01:12:23.520 --> 01:12:27.357
characterize as kind of small, medium,
and large. Like, what am I doing? Do I

01:12:27.440 --> 01:12:30.677
have small exposure on? Do I have medium
exposure on? Do I have large exposure

01:12:30.760 --> 01:12:34.357
on? You know, and if you look at what's
happened in the markets, semis ripped

01:12:34.440 --> 01:12:37.637
this year. I mean, like uh you've been
doing this a long time. I don't I've

01:12:37.720 --> 01:12:41.197
never seen it before, right? I've never
seen, you know, the doubles and the

01:12:41.280 --> 01:12:43.557
triples across the board like we saw.

01:12:43.640 --> 01:12:47.757
But, there's been huge dispersion,
right, in the market. Internet's down

01:12:47.840 --> 01:12:52.037
16%, uh software's
down 8% on the year. You

01:12:52.120 --> 01:12:55.837
know, spy and and Nasdaq are up,
but really up because of their components

01:12:55.920 --> 01:12:59.957
that are related to AI and compute.
And so, the market itself has kind of

01:13:00.040 --> 01:13:03.957
struggled. Meanwhile, if you were in
the stuff that we were invested in, we've

01:13:04.040 --> 01:13:07.477
all done pretty well. I think you know,
I've said it a couple times. I think if

01:13:07.560 --> 01:13:11.677
the Anthropic revenue had not shown up
this year, because that was the overhang

01:13:11.760 --> 01:13:15.397
on the market, I think the whole market
could be down this year. Right? Um but

01:13:15.480 --> 01:13:17.317
that showed up. You know, we just had

01:13:17.400 --> 01:13:19.757
these huge months in in in April and

01:13:19.840 --> 01:13:23.517
May. Um for us, you know, because prices

01:13:23.600 --> 01:13:27.557
came up so much, because I have some
worry about, you know, geopolitics, the

01:13:27.640 --> 01:13:32.077
macro backdrop with, you know, with with
what's going on with inflation in the

01:13:32.160 --> 01:13:35.637
short run, and just like, you know,
needing a little consolidation in this

01:13:35.720 --> 01:13:39.077
market to answer some of these questions,
because now expectations are

01:13:39.160 --> 01:13:42.597
higher. You know, we dialed back from what
I would call large for Altimeter to

01:13:42.680 --> 01:13:47.637
something kind of like medium small. Um
again, it's never all or nothing for us.

01:13:47.720 --> 01:13:50.997
It's like, what is the
risk-reward at a given price?

01:13:51.080 --> 01:13:54.517
Um and so, we think this is a,
you know, maybe going to be a period of

01:13:54.600 --> 01:13:59.157
consolidation on way to much higher highs.
Um curious just how you run the

01:13:59.240 --> 01:14:01.677
book, how you think about
it like a portfolio manager.

01:14:01.760 --> 01:14:05.237
>> Very similarly, man. I always think
stocks, the markets, I imagine them as

01:14:05.320 --> 01:14:08.957
runners. Okay? And like in '22,

01:14:09.040 --> 01:14:12.637
that runner had gone downhill.
It had a lot of energy, man.

01:14:12.720 --> 01:14:17.037
Yeah, it was painful. It wasn't fun.
Um but coming out of that, there was a lot

01:14:17.120 --> 01:14:19.557
of kind of pent-up upside in the market.

01:14:19.640 --> 01:14:23.197
And you know, the market, particularly
last 2 months, it has run up a very

01:14:23.280 --> 01:14:26.797
steep hill. And a lot
of companies, semiconductor

01:14:26.880 --> 01:14:30.677
companies in particular, you know,
ironically, you know, Nvidia and

01:14:30.760 --> 01:14:32.757
Broadcom, they they have been laggards.

01:14:32.840 --> 01:14:36.517
>> Totally. >> And so,
but a lot of these, like I do

01:14:36.600 --> 01:14:40.277
see a lot on X about finding the next
bottleneck. I think that was the last

01:14:40.360 --> 01:14:44.557
game. That game is over.
You've had a lot of stocks that forget

01:14:44.640 --> 01:14:47.077
climbing a mountain or a hill. They've

01:14:47.160 --> 01:14:49.557
gone straight up a cliff, okay?

01:14:49.640 --> 01:14:52.877
>> Yes. They're tired. They need to rest.

01:14:52.960 --> 01:14:57.477
And we'll see, do they just rest at the
top of that cliff they climbed? Do they

01:14:57.560 --> 01:15:00.837
hang out on the in their
harness for a while?

01:15:00.920 --> 01:15:03.877
We've seen some.
Or do they need to go downhill for a

01:15:03.960 --> 01:15:07.637
bit? We'll see, but I'm
thinking very similarly to you.

01:15:07.720 --> 01:15:11.557
But it is and I think there's, you know,
the market is seasonal. I think there's

01:15:11.640 --> 01:15:15.717
real real concerns around
inflation and rates.

01:15:15.800 --> 01:15:20.197
>> What was CPI this morning? >> uh 4.2.
I think we added core came in at

01:15:20.280 --> 01:15:23.357
like 0.2 versus 0.3,
so a little bit better.

01:15:23.440 --> 01:15:26.557
Um but you know, clearly
we're we're above four again.

01:15:26.640 --> 01:15:28.557
And um and and there's short-term

01:15:28.640 --> 01:15:31.917
pressure on, you know, core PCE, etc.

01:15:32.000 --> 01:15:35.237
Um and we have some unknown unknowns,
but the market, I mean, if I had told

01:15:35.320 --> 01:15:38.837
you the fact pattern for this year,
that we're going to be in a war with Iran,

01:15:38.920 --> 01:15:42.757
that, you know, oil was going to be
at 100 bucks, that CPI was going to be

01:15:42.840 --> 01:15:46.117
creeping back up, that internet was going
to be down 15%. Software is going

01:15:46.200 --> 01:15:49.797
to be down 8%. You would have said,
"I want nothing to do with that market,

01:15:49.880 --> 01:15:53.677
right?" And here we are. The market's
done pretty good in the stuff that we

01:15:53.760 --> 01:15:58.237
traffic in because the world
underestimated AI revenues and

01:15:58.320 --> 01:16:00.997
underestimated the amount of compute
that was going to be needed.

01:16:01.080 --> 01:16:04.277
>> It's odd to say you know, we're heading
into a seasonally weak period with all

01:16:04.360 --> 01:16:09.037
of these fears. AI has actually been
seasonal for the last three summers.

01:16:09.120 --> 01:16:12.237
Token consumption is kind of plateaued,
slowed down, and that's cuz you know,

01:16:12.320 --> 01:16:15.957
college kids are big AI consumers
and they don't use as much AI, you know,

01:16:16.040 --> 01:16:19.477
hopefully they're all using
it to learn and not cheat.

01:16:19.560 --> 01:16:23.317
But that may happen. It may not
happen because of generative AI.

01:16:23.400 --> 01:16:27.637
>> is building swarms of agents, building
a SpaceX model. He's going to the SpaceX

01:16:27.720 --> 01:16:29.957
IPO with me at the exchange on Friday,

01:16:30.040 --> 01:16:32.637
but I he had to build an AI model using

01:16:32.720 --> 01:16:34.677
AI agents. He had to build a model, a

01:16:34.760 --> 01:16:39.997
DCF before we go to the exchange. He
is mesmerized. He is absolutely and it's

01:16:40.080 --> 01:16:41.517
extraordinary what he's doing.

01:16:41.600 --> 01:16:44.237
>> So he's one kid who's not easy to less
computer >> [laughter] >> or something.

01:16:44.320 --> 01:16:46.037
>> He's burning it. He's burning it.

01:16:46.120 --> 01:16:48.557
>> Yeah, but you know,
if token consumption

01:16:48.640 --> 01:16:50.957
plateaus, if open source takes some

01:16:51.040 --> 01:16:55.917
share, there's a Silicon data index that
has showed, which is an index of kind of

01:16:56.000 --> 01:16:59.277
consumption and pricing. I think there
may have been a little bit of a shift

01:16:59.360 --> 01:17:02.677
over the last 2 weeks to open source
tokens that are cheaper. Like people

01:17:02.760 --> 01:17:06.917
looking at that data as bearish or not
understanding it. But nonetheless, like

01:17:07.000 --> 01:17:09.917
I just think there's reasons,
you know, to look around, be careful, be

01:17:10.000 --> 01:17:12.877
thoughtful. I always assume a bullet
is coming for me. Head on [laughter] a

01:17:12.960 --> 01:17:16.597
swivel. It's the bullet you don't see
that gets you. So, I'm trying to spin as

01:17:16.680 --> 01:17:20.237
fast as I can. But yeah,
it's the market may need to

01:17:20.320 --> 01:17:24.957
take a breather. But man, when I
think about what Noam Brown said

01:17:25.040 --> 01:17:28.997
and when I see the capabilities of Fable,

01:17:29.080 --> 01:17:32.997
it's just hard for me to get too bearish.

01:17:33.080 --> 01:17:37.797
>> I mean, like to me um and we
got two, I think, of the most

01:17:37.880 --> 01:17:41.997
extraordinary guys of, you know,
the next generation, you know, sitting in

01:17:42.080 --> 01:17:45.357
the room. We have at Altimeter, we have
deep admiration for the work that you

01:17:45.440 --> 01:17:47.557
guys do. I always appreciate when you

01:17:47.640 --> 01:17:52.837
send me a note about the work that we do
and we publish. Um but for the guys who

01:17:52.920 --> 01:17:55.877
are newer to the business, they might
think this is the way that it kind of

01:17:55.960 --> 01:17:58.957
always was, right? And like this line,

01:17:59.040 --> 01:18:01.397
the steepening of the line of creative

01:18:01.480 --> 01:18:04.197
destruction, the steepening of the line

01:18:04.280 --> 01:18:07.717
of, you know, scale advantages. Um I

01:18:07.800 --> 01:18:11.757
always believed it was to it was going
to be true. I never thought it would be

01:18:11.840 --> 01:18:14.477
true at this rate. I went back last

01:18:14.560 --> 01:18:16.957
night. In the last 7 years, we've added

01:18:17.040 --> 01:18:19.957
1 trillion of revenue to the Mag 7

01:18:20.040 --> 01:18:22.637
in the last 7 years, okay? To get to a

01:18:22.720 --> 01:18:25.477
trillion, to get to the first trillion

01:18:25.560 --> 01:18:28.157
of, you know, took over 20 years. In the

01:18:28.240 --> 01:18:30.397
last 7, we had another tr- trillion and

01:18:30.480 --> 01:18:33.357
that added 17 trillion in market cap.

01:18:33.440 --> 01:18:35.197
That trillion dollars, okay?

01:18:35.280 --> 01:18:37.637
I The forecast now that we're going to

01:18:37.720 --> 01:18:40.437
add another trillion of revenue in just

01:18:40.520 --> 01:18:45.597
three companies SpaceX Anthropic
and open AI over the next four to five

01:18:45.680 --> 01:18:49.397
years. Okay, like not
seven companies three

01:18:49.480 --> 01:18:54.837
companies and in half the time right
and so I would say that you know, we are

01:18:54.920 --> 01:18:58.157
going to have bumps in the road.
I know that it's going to be like this but

01:18:58.240 --> 01:19:00.197
we're going to higher highs because the

01:19:00.280 --> 01:19:02.757
size of the prize. This is going to

01:19:02.840 --> 01:19:06.877
transform five ten 15% of global GDP.

01:19:06.960 --> 01:19:11.797
There is no doubt in my mind and 10%
of global GDP is 10 trillion dollars. It's

01:19:11.880 --> 01:19:15.077
an exciting future to be a part of it's
fun to do it with you guys. I think

01:19:15.160 --> 01:19:18.357
we're going to have to do our work to do
the things to make sure America wins and

01:19:18.440 --> 01:19:22.597
that we evolve the social contract keep
everybody you know lift the floor take

01:19:22.680 --> 01:19:27.357
everybody with us on this ride
but it's a it's a it's a really exciting

01:19:27.440 --> 01:19:30.077
time to be doing what we're doing
it's fun to be doing it with you guys.

01:19:30.160 --> 01:19:31.477
>> Yeah, I just want
to say Brad thanks for

01:19:31.560 --> 01:19:34.437
having us and thank you for what you've
done with the Trump accounts. I actually

01:19:34.520 --> 01:19:38.877
think it's super important for America
for the world to give people an equity

01:19:38.960 --> 01:19:42.277
stake at a very young age.
They they will see it compound over their

01:19:42.360 --> 01:19:46.237
lifetimes. This is a great thing you've
done for the world. So thank you. I'd

01:19:46.320 --> 01:19:50.317
echo all your comments like deep
admiration for you your team gratitude

01:19:50.400 --> 01:19:55.077
for the collegiality and friendship
between our firms. I know Clark and Foxy

01:19:55.160 --> 01:19:57.317
they hang out like all the time.

01:19:57.400 --> 01:20:00.677
>> That's a people think that you know
and there are people in our business who

01:20:00.760 --> 01:20:04.877
don't want to share anything.
Our view is like we open source it

01:20:04.960 --> 01:20:08.277
but there are very few people who we
actually call and ask their opinion

01:20:08.360 --> 01:20:12.757
because there are very few people who do
the thousands of hours of work that we do

01:20:12.840 --> 01:20:16.317
you know that are adding to that and you
do it and we appreciate that and you do

01:20:16.400 --> 01:20:20.317
as well Gavin we appreciate that. So
with that love fest let's call it a wrap.

01:20:20.400 --> 01:20:23.960
Thanks for being here. >> Thank you.

01:20:47.280 --> 01:20:49.320
>> Mhm.