00:00:00
I'm here at Nvidia's annual GTC
00:00:02
conference and I'm going to interview
00:00:04
four amazing AI CEOs. Stick with us.
00:00:15
>> Our episode is sponsored by the New York
00:00:17
Stock Exchange. Are you looking to
00:00:19
change the world and raise capital? Do
00:00:21
it at the NYSE. The NYSE is a modern
00:00:25
marketplace and a massive platform built
00:00:28
for scale and long-term impact. So if
00:00:31
you're building for the future, the NYSE
00:00:33
is where it happens.
00:00:37
>> One of the great companies of the AI era
00:00:40
is of course Cororeweave. They're
00:00:41
building massive infrastructure for
00:00:43
these hyperscalers. And in some ways,
00:00:45
Michael Intrader, welcome to the
00:00:47
program. You're the original
00:00:49
hyperscaler. you guys got in very early
00:00:52
and secured your I don't know which GPUs
00:00:56
you wound up getting but you were very
00:00:58
early to this trend. How did you get to
00:01:01
it so early and how did you build out
00:01:04
this you know first I guess at the time
00:01:06
neocloud? Yeah. So we we didn't really
00:01:09
start it as a Neocloud and I I uh I was
00:01:12
uh running an algorithmic hedge fund uh
00:01:15
focused on natural gas and uh when when
00:01:19
you build an algorithmic hedge fund um
00:01:21
once the algorithms are built you're
00:01:23
really just monitoring it and testing
00:01:25
different uh thesises and doing all
00:01:27
that. But there's also a lot of downtime
00:01:29
and we got super interested in crypto.
00:01:32
Um, and you know, we're pretty nerdy. We
00:01:35
kind of dig under the hood and we
00:01:36
started to get interested in the
00:01:37
security layer. Uh, we looked at Bitcoin
00:01:40
and the mining for Bitcoin and we didn't
00:01:42
like it. We just thought that like
00:01:43
there's some brilliant engineer that
00:01:46
built the ASIC and they're probably
00:01:47
going to be better at running it than we
00:01:49
are. So, we really began to focus on the
00:01:51
GPUs mostly because the GPUs were you
00:01:55
can mine Ethereum with them. uh but you
00:01:58
could also do all these other things and
00:01:59
really so right from the start we looked
00:02:02
at the compute as an option to be able
00:02:06
to deploy our computing power to
00:02:10
different use cases and so you know
00:02:12
began the company in 2017 uh you know um
00:02:15
spent the first kind of three years
00:02:19
mining crypto went through a couple of
00:02:21
crypto winters um because we had come
00:02:24
from a hedge fund were, you know, we we
00:02:27
have real chops in risk management and
00:02:29
how we think about uh capital and risk
00:02:32
exposure and allocation and all of that.
00:02:34
And so we were really careful around
00:02:35
that right from the start. So we
00:02:37
weathered crypto winter really well um
00:02:39
and began to scale the company and
00:02:41
immediately started to look for other
00:02:43
use cases that you could use this
00:02:45
compute for because crypto was pretty
00:02:47
volatile.
00:02:48
>> Yeah. And crypto was a question mark at
00:02:50
that time.
00:02:50
>> Absolutely.
00:02:51
>> Yeah. I mean Bitcoin was speculative and
00:02:53
there were many other specular projects.
00:02:55
the only other people using this type of
00:02:57
hardware quants
00:02:59
>> medical researchers.
00:03:01
>> So a good way to think about it is like
00:03:03
the progression of products that we kind
00:03:05
of started to work on. You know first
00:03:07
was uh um crypto but we immediately
00:03:10
moved from crypto to CGI rendering and
00:03:12
we built projects that would allow uh um
00:03:16
folks that were trying to animate and
00:03:18
render images um you know kind of what
00:03:22
makes the movies cool, right? and and uh
00:03:24
we started to work on that and then we
00:03:26
moved to batch computing and started to
00:03:27
look at medical research and different
00:03:29
ways of using the compute to be able to
00:03:31
drive science. Um, and we just kind of
00:03:33
kept moving up the stack in terms of
00:03:36
complexity uh uh on how GPUs could be
00:03:39
used. And ultimately uh in like call it
00:03:43
like 2020 2021 we started to really try
00:03:47
to figure out how you can go ahead and
00:03:49
use GPUs for neural networks and that
00:03:52
was not something that uh we knew how to
00:03:54
do. Um, and so we actually went out and
00:03:56
bought a bunch of A100s and donated them
00:03:59
to a uh a group that was working on uh a
00:04:02
Luther AI. They were working on an open-
00:04:04
source project with the thought that um
00:04:07
these guys are taking the GPU compute
00:04:11
because we're donating it. They can't
00:04:13
really get pissed at us if we're not
00:04:14
very good at it initially. And uh that
00:04:16
worked out really well because
00:04:17
>> they can't complain about the SLA.
00:04:18
>> They they kept telling us like we need
00:04:20
more of this, you got to work on this.
00:04:22
And that began to really uh uh give us
00:04:26
an understanding of what was necessary
00:04:28
to run scale parallelized computing. And
00:04:32
uh you know that that uh um we went
00:04:34
through it. I I I kind of feel like
00:04:35
buying those initial GPUs was the
00:04:37
tuition we paid to learn how to run this
00:04:40
business. And then one of the
00:04:41
interesting things is all of those guys
00:04:43
went back to their day jobs because they
00:04:45
were all volunteers working on this.
00:04:47
They were like-minded scientists. And
00:04:49
when they got to their day jobs, they
00:04:51
were all like, I want that
00:04:52
infrastructure. It's built the right
00:04:54
way. That's the way that researchers are
00:04:56
going to want to use it. And that
00:04:57
launched our our business. It was an
00:04:59
amazing story. And
00:05:00
>> so you went from crypto to these
00:05:03
researchers into academia and deep
00:05:05
research. What's the next card to turn
00:05:07
over in the poker game?
00:05:09
>> Yeah. So, so um what became very clear
00:05:11
to us very very early on was that the
00:05:15
scaling laws were going to drive and
00:05:18
remember this is really back in the you
00:05:20
know 2020 2021 before uh uh chatgpt
00:05:24
moment occurred and we began to
00:05:27
understand that like computing
00:05:29
decommoditizes at scale right like when
00:05:31
when you know anybody can run a GPU but
00:05:34
can you run a cluster that's large
00:05:35
enough to train a model that can change
00:05:37
the world and that's a different
00:05:38
question. And so we really began to
00:05:41
think about like how do you go about
00:05:44
scaling up your delivery of this
00:05:47
computing to clients, larger and larger
00:05:49
clients. And that was the next card to
00:05:51
turn is to think about it from a okay,
00:05:54
you know, there's a component of this
00:05:55
that is going to lean into uh our
00:05:58
ability to access the capital to be able
00:06:01
to deliver our solution to the broadest
00:06:04
possible audience to the most
00:06:05
sophisticated consumers of this compute.
00:06:07
And and that was really the next card is
00:06:10
thinking about it as a business rather
00:06:12
than as a engineering project to be able
00:06:14
to deliver the the uh uh the
00:06:17
infrastructure and the software and
00:06:19
really everything between you know when
00:06:21
you when you're thinking about what we
00:06:22
do, we kind of live above the Nvidia
00:06:25
GPUs but below the models. Yeah. and
00:06:28
everything in there, all the software,
00:06:30
the integration of software and
00:06:32
operations and uh observability and all
00:06:35
the things that you need to be able to
00:06:36
build uh a cloud that's purpose-built
00:06:40
for this one specific use case, right?
00:06:42
So, we don't we don't do everything. We
00:06:44
really focus on one use case which
00:06:46
allows
00:06:46
>> you want to do web servers different you
00:06:48
got AWS,
00:06:48
>> you know what they do a great job. It's
00:06:50
like it's a it's a great solution. It
00:06:52
was a brilliant solution to solve a
00:06:53
problem. We just looked at it and said
00:06:55
there's a new problem and let's go about
00:06:57
let's go about looking at this problem
00:06:59
and try and come up with a solution to
00:07:01
deliver compute that solves that
00:07:03
problem.
00:07:03
>> And when did the language model start
00:07:05
dialing and calling you for you know
00:07:08
capacity?
00:07:09
>> Yeah. So uh our our our first uh um well
00:07:13
our our first language model was really
00:07:15
a Luther. Um but uh our our first like
00:07:18
large commercial uh was inflection. Um
00:07:22
and so you know we work with Mustafa and
00:07:25
and and and Inflection and then we we
00:07:27
really diversified from there uh into
00:07:31
the hyperscalers into you know uh open
00:07:35
AAI across the the the the model uh the
00:07:38
foundation models across um you know and
00:07:41
and just kept scaling and scaling with
00:07:43
the belief that you know once again the
00:07:46
the the decommoditization
00:07:48
of compute the ability to to deliver a
00:07:52
solution and the solution is building
00:07:54
supercomputers that can change the world
00:07:57
and that's really what we began to focus
00:07:59
on. That was the lead into training and
00:08:01
now the world has gone through, you
00:08:03
know, this this moment where we've moved
00:08:05
from research into the productization of
00:08:09
this. It's it's it's beginning to work
00:08:11
its way in from the the uh the fringe of
00:08:14
organizations into the core of what they
00:08:16
do. And you can see that every day in
00:08:19
the uh in the amount of inference
00:08:21
compute that is being driven through you
00:08:24
know our uh infrastructure layer which
00:08:27
is just massive which is just like one
00:08:28
of the shows you people are consuming it
00:08:31
not just building models but they're
00:08:33
deploying them and and utilizing them.
00:08:35
>> I always think of inference as the
00:08:37
monetization
00:08:39
>> of the investment in artificial
00:08:41
intelligence. So when when when we when
00:08:44
we see our compute being used uh uh to
00:08:48
stand up the massive scale of inference
00:08:50
that's hitting our compute every day and
00:08:52
like you know inference is when people
00:08:54
ask the model a question it comes back
00:08:57
with an answer that's an inference or
00:08:58
when you ask the model a question and
00:09:00
then to go do something that's inference
00:09:02
right and that's actually where you're
00:09:04
you're you have the opportunity to
00:09:06
really drive value outside of the model
00:09:10
itself but into the real world and
00:09:13
that's really exciting for us. That's
00:09:15
what we like to watch. That's what I
00:09:17
like to watch in terms of gauging the
00:09:18
health.
00:09:19
>> What chips are those?
00:09:20
>> Um so so really uh you know we are we
00:09:26
are the tip of the spear in bringing um
00:09:28
the new architecture uh out of Nvidia uh
00:09:32
into uh um into commercial production at
00:09:35
scale. Yeah. And uh so when when you
00:09:38
know we were the first ones to bring the
00:09:39
H100s at scale, we were the first ones
00:09:41
to bring the H200s at scale, first ones
00:09:43
with the GB uh 200s, and now you've got
00:09:46
the GB300s. And one of the things that's
00:09:48
that's that's amazing and really
00:09:50
fascinating for us is is you know people
00:09:53
are using the bleeding edge GPUs to
00:09:55
train models as the new architectures
00:09:58
come out and then they take those GPUs
00:10:01
and they move them into different
00:10:03
experiments and then over time they move
00:10:06
them into inference and they continue to
00:10:09
use them in inference for a very very
00:10:10
long time.
00:10:11
>> What is the shelf life of a 100 right
00:10:13
now? That's been a big debate is I think
00:10:16
for your company for Microsoft and I
00:10:19
guess Michael Bur you know who you must
00:10:21
have known when you were a quant you
00:10:23
know saying oh my god the whole industry
00:10:25
is the sky's falling and then we all
00:10:27
know in the industry that people don't
00:10:29
just throw this hardware away that they
00:10:30
find uses for it the street finds its
00:10:32
own use for technology so what's the
00:10:34
reality of the lifespan of these things
00:10:36
>> so so my my take on the the uh uh the
00:10:40
GPU uh depreciation bait is that it's
00:10:43
nonsense Right. It's a debate that is
00:10:45
being uh brought to the forefront by uh
00:10:48
some traders that have a short position
00:10:50
in the stock and they're trying to uh
00:10:52
talk down. Look, here's what we know,
00:10:54
right? Um
00:10:57
when when we buy infrastructure, we're a
00:10:59
success based company, right? We're a
00:11:00
small company on a relative basis
00:11:02
compared to the enormous companies that
00:11:04
we're competing with. And so they come
00:11:07
our clients come into us and they buy
00:11:09
compute for five years, for six years.
00:11:11
Our average contract is 5 years. So any
00:11:15
commentary by anyone either inside or
00:11:18
outside of the industry that this stuff
00:11:19
becomes obsolete in 16 months or
00:11:21
whatever nonsense they're spewing, it's
00:11:23
it doesn't it doesn't in any way match
00:11:26
up with the facts on the ground. The
00:11:28
facts on the ground is they're buying it
00:11:29
for 5 years. Right? If and my approach
00:11:32
to this has always been if people are
00:11:34
willing to pay me for it,
00:11:36
>> it still has value.
00:11:38
>> Correct.
00:11:38
>> Pretty simple way of of approaching it.
00:11:40
We use a six-year depreciation. Um, we
00:11:43
believe that the GPUs will last in
00:11:45
excess of six years, but we felt like
00:11:47
that was a fair and reasonable approach
00:11:49
to a technology cycle that's moving at
00:11:51
this velocity. Um, the A100s, the ampers
00:11:55
this year, the price has appreciated
00:11:57
through the year.
00:11:58
>> And why is that? I I think it's because
00:12:00
one of the things that happens is as
00:12:03
more installed capacity becomes
00:12:06
available, you have new companies that
00:12:08
come into existence that have new use
00:12:09
cases that have different size models
00:12:11
that are trying to uh build new
00:12:14
commercial ventures that maybe have been
00:12:16
blocked out of the H100s and never had
00:12:19
an opportunity to run on that. I mean to
00:12:20
make a very simple example for the
00:12:22
audience like when you trade in your
00:12:24
iPhone after 3 or 4 years you're like
00:12:27
who's going to use an iPhone 12 and it's
00:12:29
like have you been to South America or
00:12:32
Africa where you go to the store and you
00:12:35
buy an iPhone 12 or you buy the Pixel 7
00:12:38
and it costs $50 that's still got great
00:12:41
life left in it.
00:12:42
>> Absolutely.
00:12:42
>> Yeah. you know,
00:12:43
>> and so look, you know, we we find these
00:12:46
amazing use cases, new companies that
00:12:48
have come into existence or existing
00:12:51
companies that have integrated new
00:12:53
models into their workflow that are able
00:12:55
to use the Ampierce and so they keep
00:12:59
buying any GPUs that we have available.
00:13:02
And once again, you know, the the
00:13:03
concept that a GPU
00:13:05
>> is no longer relevant or commercially
00:13:07
viable after 16 more 18 months or two
00:13:10
years.
00:13:10
>> Yeah, that's it just it just doesn't
00:13:12
make sense.
00:13:12
>> It's obviously far. I think sometimes
00:13:14
people get caught up in Moore's law or
00:13:16
in just how fast our industry is growing
00:13:19
and that there's so much at stake that
00:13:22
big companies are demanding the most
00:13:24
recent products. That doesn't mean that
00:13:27
the lifespan has gotten shorter. It
00:13:29
means the opportunity and the surface
00:13:31
area of the opportunity has gotten much
00:13:33
larger.
00:13:33
>> Yeah. Uh one of the things is is like
00:13:36
you know the the uh the the industry has
00:13:40
gotten so much attention for the
00:13:43
unprecedented scale of capital that is
00:13:46
coming to bear on this. And
00:13:49
because of that, there tends to be a
00:13:51
incredible focus on
00:13:54
the companies that are building on these
00:13:57
most advanced chipsets. And the truth of
00:14:00
the matter is is you know even within
00:14:03
those companies they have a long tale of
00:14:05
useful life
00:14:06
>> to provide inference horsepower to work
00:14:09
on other experiments to do less bleeding
00:14:12
edge activity but still needs to be done
00:14:14
>> and yeah I mean rendering comes to mind
00:14:16
as well or yeah we're making images on
00:14:19
nano banana like there there will be a
00:14:21
use for it. There is a moment in time
00:14:23
where maybe the compute to power ratio
00:14:25
doesn't make sense. My my expectation is
00:14:29
is obsolescence will be defined by the
00:14:34
moment in time where the power
00:14:37
in the data center for me will be able
00:14:41
to be repurposed for a higher margin
00:14:44
than the existing infrastructure
00:14:46
provides. And you know, like I said, I I
00:14:49
fully expect this infrastructure to last
00:14:51
in excess of 6 years, but the the the
00:14:54
standard in the in in in the space has
00:14:56
really been used with one exception,
00:14:59
which is Amazon, which is Yeah, it's 6
00:15:01
years. That's that seems like the right
00:15:02
schedule. I'm not making it up. That's
00:15:04
what everybody's using.
00:15:05
>> Yeah. And the energy cost is the
00:15:09
opportunity because hey, it's just a we
00:15:11
need that space. there's a better uh
00:15:14
reward here and that might get resold
00:15:16
that hardware to somebody else who wants
00:15:18
it a hobbyist or something. It's
00:15:20
available
00:15:21
>> and or it could be sent someplace else
00:15:22
where they have more capacity when they
00:15:25
can repurpose it there. But I I I um I
00:15:27
kind of feel like, you know, we'll we'll
00:15:29
deal with that part of the business when
00:15:31
we get there. What I know right now is
00:15:34
it is extraordinarily profitable. It's
00:15:36
very creative to my company to continue
00:15:39
to keep the infrastructure that's been
00:15:41
up and running, that's been on these
00:15:42
long-term contracts, and as it rolls
00:15:44
off, as it's been in use for 5 years,
00:15:47
you know, as it becomes available, I am
00:15:49
still able to sell it at a higher price
00:15:50
than it was at a year ago. There's
00:15:53
competition now. When you were buying
00:15:54
these from Jansen back in the day, yeah,
00:15:57
you could buy them and have them
00:15:58
shipped, I would assume, within 30 days
00:16:01
or less. nowadays what's the weight like
00:16:04
even for you a loyal old customer and is
00:16:07
there a bit of a battle is there
00:16:08
politics to who gets the servers like I
00:16:12
you see some like very big names talking
00:16:14
about they got to get an allocation is
00:16:16
it still a little bit crazy what's it
00:16:18
like to be in that category having to
00:16:20
buy something everybody wants
00:16:22
>> look uh you know I I uh I I think of it
00:16:25
as an affirmation of the business that
00:16:26
we're in right like the fact that we are
00:16:29
attracting competitors the the means
00:16:31
that the business is healthy and there's
00:16:33
a lot of people trying to deliver this
00:16:34
service because the need for this
00:16:38
infrastructure the need to integrate the
00:16:40
infrastructure you know into the
00:16:42
software layers to deliver it to
00:16:44
artificial intelligence uh either at the
00:16:46
model level or the inference level or
00:16:48
the application level or whatever you
00:16:50
know level of the five layer cake that
00:16:52
Jensen's you know focused on
00:16:55
the the fact that there are more people
00:16:57
coming into this it doesn't discourage
00:16:58
me um as far as getting access to the GP
00:17:01
CPUs, we show up like everybody else
00:17:03
with a um you know, we'd like to buy
00:17:06
here's a PO and we're ready to pay. Um
00:17:08
the one what's the wait time like? And
00:17:11
is it just really competitive or not?
00:17:15
Because I talked to Jensen about he said
00:17:17
I said, "How do you manage all these
00:17:18
like big egos and names and companies
00:17:20
trying to buy stuff?" And he said,
00:17:22
"Well, they order it and we give it to
00:17:24
them in the order in which they order
00:17:26
it."
00:17:26
>> Is it really like that?
00:17:27
>> It really is. Right. like you know he he
00:17:30
doesn't want to be in the position of
00:17:33
playing favorites or all like that that
00:17:35
just seems like a bad place to be with
00:17:37
your clients
00:17:37
>> or auctioning them off. Can you imagine?
00:17:40
>> Yeah, that would that that
00:17:41
>> that would be crazy.
00:17:42
>> Yeah.
00:17:43
I don't I'm not sure that would be good
00:17:45
for the long-term business. No. Yeah.
00:17:46
So, so our our our approach is
00:17:49
>> you might get some sovereigns coming in
00:17:50
and saying I'll pay double. They do that
00:17:52
with Ferraris too sometimes.
00:17:54
>> I guess these are the Ferraris of
00:17:56
computing, right?
00:17:56
>> In a way they are. Yeah. Bugattis. Our
00:17:59
>> our our approach is to work with
00:18:02
clients across the entire space to find
00:18:05
opportunities that are really
00:18:07
interesting companies that can fit into
00:18:09
our contraction contracting requirements
00:18:12
where we're going to be able to go out
00:18:14
and structure the debt that we require
00:18:16
in order to go out and and uh build
00:18:18
infrastructure at this scale. And um
00:18:21
>> how does all that debt work? I that is
00:18:23
something that you guys specialize in.
00:18:26
um corporate debt uh I'm in the venture
00:18:28
business people are like why should I be
00:18:30
in venture when corporate debt pays so
00:18:31
well corporate paper's so huge I'm
00:18:34
curious how this fits in and like what
00:18:37
uh interest rate people are paying on
00:18:40
you know a billion dollars in
00:18:43
infrastructure what do they pay on that
00:18:44
>> yeah so so coreweave has really been the
00:18:49
innovator around a lot of the financing
00:18:52
engines that have come to bear on this
00:18:54
we did the first GPU based uh loans. Um,
00:18:58
and like I I think it's important or I'm
00:19:01
going to try to explain this in a way
00:19:02
people can understand. So what we do is
00:19:05
we go out and we find a client. Let's
00:19:07
use Microsoft. You brought them up
00:19:09
before, right? And Microsoft comes to us
00:19:11
and says, "We'd like to buy some compute
00:19:12
for you." And we say, "Okay, great.
00:19:14
We're going to sign a contract." Once I
00:19:15
have a contract in hand,
00:19:17
>> then what I do is I create something.
00:19:19
It's not a particularly creative name.
00:19:21
It's called the box. Yeah.
00:19:22
>> Right. And what I do with the box is I
00:19:25
take my contract with Microsoft and I
00:19:27
put it in the box. I go to Jensen and I
00:19:29
buy the GPUs, I put it in the box. I
00:19:31
take my data center contract, I put it
00:19:34
in the box. And now the box governs cash
00:19:36
flow.
00:19:37
>> And it has a waterfall of cash flow that
00:19:39
comes into it and goes out of it. And so
00:19:41
the way it works is then I build the
00:19:44
compute and then I deliver the compute
00:19:46
to Microsoft and they pay the box. They
00:19:48
don't pay me,
00:19:49
>> right? It goes into the box and the
00:19:51
first thing it does is it pays the data
00:19:53
center. It pays the power bill. It pays
00:19:56
the interest and the principal and then
00:19:59
whatever's left flows back to us, right?
00:20:02
And so it is an incredibly well
00:20:05
ststructured, time-tested,
00:20:07
pressure-etested vehicle to be able to
00:20:09
borrow money against client paper and
00:20:13
all of the other collateral around the
00:20:16
deal. which is why Corewave, which is a
00:20:18
company that many people haven't ever
00:20:20
heard of, was able to go out and raise
00:20:22
$35 billion in 18 months to build
00:20:25
infrastructure at scale. But what's
00:20:27
important to understand is the economics
00:20:30
in this box are such that within 2.5
00:20:33
years of a 5-year deal, we have paid for
00:20:37
everything.
00:20:38
>> The principal's been paid off. The well
00:20:40
the principal's been paid off, the
00:20:42
interest has been paid off. The return
00:20:43
into the box is such that we are able to
00:20:47
generate returns to our company at the
00:20:50
box level which gives the most
00:20:53
sophisticated lenders in the world
00:20:56
whether it's banks or private equity
00:20:59
funds or um you know whoever. confidence
00:21:03
that they're going to
00:21:06
be able to achieve the one rule of
00:21:08
lending, which is give me my money back.
00:21:10
>> Yes. Works better when that happens.
00:21:12
>> So, they look at this box and they're
00:21:13
like, "Wow, we're really confident we're
00:21:15
going to get our money back."
00:21:16
>> And maybe they want 10 boxes.
00:21:17
>> That's correct.
00:21:18
>> And if any one box um goes upside down,
00:21:22
you can deal with it and it's not as
00:21:24
acute.
00:21:25
>> That's correct. And they don't
00:21:26
cross-pollinate. They don't cause uh
00:21:29
contagion across the boxes. are all
00:21:31
independent and discreet. One, and
00:21:33
number two is as you do this and as you
00:21:36
show the lenders how this financing tool
00:21:40
and how this financing mechanism works,
00:21:43
what they do is they continue to lend
00:21:45
you money at progressively lower rates.
00:21:48
And so when you think about our cost of
00:21:51
capital over the last two years, we have
00:21:55
dropped our cost of capital by 600 basis
00:21:57
points.
00:21:58
>> Wow. It is enormous, right? And so
00:22:00
you're seeing a company that is driving
00:22:02
its cost of capital down towards where
00:22:05
the hyperscalers borrow, which will
00:22:08
enable us to be able to be competitive
00:22:10
with them over time. And we have been
00:22:14
extremely
00:22:15
uh militant and diligent about feeding,
00:22:19
watering, and caring for those boxes so
00:22:21
that we continue to have access to the
00:22:23
capital markets in a way that allows us
00:22:25
to build and drive our business.
00:22:26
>> Means you has to say no. You have to say
00:22:28
no to maybe some people who want to be
00:22:29
in the box.
00:22:30
>> Yeah. So, we we look at some deals and
00:22:33
we're just like, you know, they want to
00:22:35
buy GPUs for a year and I look at it and
00:22:37
say I I that's not a deal that I can do
00:22:39
because it's too short for me to am
00:22:41
advertise the expenses or and so I won't
00:22:44
do that. Right. Like once
00:22:46
>> and they can go to another provider who
00:22:47
maybe wants to take that risk on who has
00:22:49
extra capacity.
00:22:50
>> Absolutely. But our business is really
00:22:52
built about around the risk management
00:22:54
of being able to get to scale. Because
00:22:56
in my mind
00:22:59
during this period of disequilibrium
00:23:01
during this period where there are not
00:23:03
enough GPUs in the world to uh provide
00:23:06
the compute for all of the different use
00:23:09
cases in artificial intelligence the
00:23:11
part that's important for me and for my
00:23:13
company is to get enormously large so we
00:23:16
can drive down our cost of capital so
00:23:18
that we have information flow coming in
00:23:20
from all different parts of the market.
00:23:22
large language models, high-speed
00:23:24
trading, uh, uh, search, all of these
00:23:27
things. And they're feeding they're
00:23:28
feeding information back into us that is
00:23:31
letting us know what the next product we
00:23:33
need to build is or where, you know,
00:23:35
they need help uh, scaling or what type
00:23:38
of compute they need and all of that
00:23:40
information flow is incredibly valuable
00:23:43
to us.
00:23:44
>> What What can you tell us about demand?
00:23:45
There's been reports of, hey, maybe the
00:23:48
Oracle Starbase thing with OpenAI's been
00:23:52
downsized or maybe not and then you know
00:23:56
uh other folks Microsoft is going big
00:23:58
and Google's going big Meta's going big
00:24:01
and those people obviously have massive
00:24:03
cash flow Apple seems to be MIA they
00:24:05
don't seem to want to play you you
00:24:07
you've you've uh you've named a lot of
00:24:09
really big companies with really big
00:24:10
balance sheets that have the capacity to
00:24:12
drive a lot of demand look I I have been
00:24:15
truly steadfast in this
00:24:18
>> for years now for for for four
00:24:21
The depth of the demand for the service
00:24:23
we provide has been relentless and
00:24:27
overwhelms the global capacity of the
00:24:30
world to deliver enough compute to
00:24:33
enable all of the demand for artificial
00:24:37
intelligence to be sated and that has
00:24:40
been we have been relentless about that.
00:24:42
>> Sounds like Nick's tickets during the
00:24:44
Patrick Euing era like they got up to
00:24:46
50,000 people on the wait list. So if
00:24:49
magically the weight list went away, if
00:24:52
the if the constraint went away and we
00:24:53
just had a large amount of GPUs
00:24:56
available, lot of energy available, a
00:24:59
lot of data center available, how much
00:25:03
capacity would just all of a sudden come
00:25:05
out of the system.
00:25:05
>> So so or would be deployed I should say.
00:25:08
>> So remember how we build our our
00:25:11
business through this box
00:25:12
>> and it's a fiveyear box. So if we had an
00:25:16
air pocket, if if demand were suddenly
00:25:19
to disappear because of a technology
00:25:21
breakthrough, because of a uh a war,
00:25:24
anything, right? Like like the why from
00:25:26
a riskmanagement perspective does not
00:25:28
matter. You have to prepare your company
00:25:31
for the what happens if it happens.
00:25:33
Yeah. And so by entering into these
00:25:36
long-term contracts into entering into
00:25:38
contracts with counterparties that have
00:25:40
large balance sheets, you are or we are
00:25:44
protecting ourselves and our lenders.
00:25:47
Yeah.
00:25:47
>> So that we are confident and they are
00:25:50
confident because you can see how
00:25:51
confident they are by the rate that
00:25:53
they're charging us continuing to
00:25:54
decline that they're ultimately going to
00:25:56
get their money back. And that is the
00:25:58
one rule of lending. And so um you know
00:26:02
I if
00:26:02
>> but just in terms of the capacity if you
00:26:04
were unconstrained and Nvidia Jensen
00:26:06
says hey order as many as you want what
00:26:08
would happen
00:26:09
>> so um the the it's also important to
00:26:13
understand the constraints aren't just
00:26:14
GPUs right electricity it's it's power
00:26:17
shells it's memory it's storage it's
00:26:20
it's networking it's optics all of the
00:26:23
things and there's there's various
00:26:24
throttles that will limit the
00:26:27
>> memory is a throttle right now right
00:26:28
>> oh yeah it Oh yeah, it is.
00:26:30
>> Why? How did memory become the throttle?
00:26:32
>> If um
00:26:34
memory and uh it has historically been a
00:26:39
cyclical business, right? We have seen
00:26:41
these waves of demand driving up the
00:26:44
cost for memory and then it collapses
00:26:46
and then it drives it up. It's a very
00:26:48
boom and bust business. is cyclical in
00:26:50
its nature because the fabs are so
00:26:53
capital intensive that people invest in
00:26:55
the fabs, build a ton of capacity and
00:26:58
then overbuild if there's any type of
00:27:02
turndown. And that we've seen that cycle
00:27:04
again and again. What's happening right
00:27:06
now is the confluence of two things,
00:27:09
right? one is is
00:27:11
with all the demand for artificial
00:27:13
intelligence and the corresponding
00:27:15
demand for compute and the ancillary
00:27:18
services around the GPU, the demand is
00:27:21
through the roof. That's number one.
00:27:23
Number two is is that
00:27:25
>> there was probably an investment cycle
00:27:27
that needed to happen back in 2023
00:27:30
that would have brought on the necessary
00:27:32
fab capacity to be able to serve.
00:27:35
>> Impossible to predict what should happen
00:27:37
just with energy. It's impossible to
00:27:38
predict what just happened. And now
00:27:40
people are chasing energy. The data
00:27:42
centers are going where the energy is.
00:27:44
It's not based on real estate. It's
00:27:46
based on it's and
00:27:47
>> where's there's some wind.
00:27:48
>> And anytime you you have a uh very cap
00:27:52
not every any time, but many times when
00:27:54
you have a uh a capital inensive
00:27:56
business like you know building fabs,
00:27:58
you will get this boom and bust cycle
00:28:00
just like in energy they overbuild.
00:28:02
Yeah. And you know
00:28:03
>> fiber.
00:28:04
>> Yeah. I mean there's there's there's a
00:28:06
lot of examples of that our approach
00:28:09
>> in some ways when you look at that it's
00:28:11
a beautiful aspect of capitalism that
00:28:13
we're able to have a boom bus cycle that
00:28:17
we're able to weather it right if you
00:28:18
think just that capitalism from first
00:28:20
principles something like that happens
00:28:22
and we have too much fiber it creates an
00:28:24
opportunity for Google to buy it all up
00:28:26
or the next person
00:28:27
>> listen the the the um um you know it it
00:28:30
does it does a lot of things having a
00:28:32
boom cycle it clears out the underbrush.
00:28:35
will be able to survive and take
00:28:37
advantage of that and it sews the seeds
00:28:39
of future business.
00:28:45
You put the fiber into the ground which
00:28:47
became the backbone of how you know we
00:28:51
watch movies every day and how we you
00:28:53
know uh communicate and how we hop on a
00:28:55
zoom and you know co and all of these
00:28:58
things were based on that infrastructure
00:29:01
that was available to be consumed. Yeah,
00:29:04
people don't recognize this fact if you
00:29:06
the the premise of YouTube from the
00:29:08
founders who I knew, Chad Hurley and his
00:29:12
other partner. They basically had the
00:29:14
realization at this curve storage is
00:29:16
coming down so quickly we could offer
00:29:18
free unlimited uploads and bandwidth is
00:29:21
coming down. So I guess we don't have to
00:29:23
charge people for sharing a video
00:29:25
online. Before that, if your video went
00:29:28
viral, people are going to have their
00:29:30
minds blown. But your server would turn
00:29:33
off and it would say this person, you
00:29:35
know, needs to pay their bill. Yes.
00:29:36
Because they were getting charged for
00:29:38
carriage by the megabit going out.
00:29:41
>> Yes. I mean it look and and you know
00:29:43
these these the business models change
00:29:45
and evolve and you know like you said
00:29:48
Moore's law and and and certainly Jensen
00:29:50
will talk about the fact that like what
00:29:52
what is going on within the the the
00:29:55
accelerated comput dwarfs
00:29:58
>> Moore's law right and all of that is
00:30:00
going to lead to
00:30:03
>> more opportunity to build more companies
00:30:05
that are going to do things like you two
00:30:07
did which has really changed the world.
00:30:09
>> Yeah. I mean the the concept that I I
00:30:12
don't know if it was like a million
00:30:14
hours being uploaded every hour or
00:30:16
minute but at some point Susan what
00:30:18
Jackie rest in peace said told me just
00:30:21
like how much was being uploaded every
00:30:25
minute and it made no logical sense and
00:30:27
she realized
00:30:28
>> well there's three billion people two or
00:30:31
three billion people in the service and
00:30:32
1% upload or 0.1 10 bit bips upload and
00:30:36
it's like okay one in a thousand people
00:30:37
upload it's a big it's a big denom
00:30:39
denominator like
00:30:40
>> I I was uh sitting on a a panel uh with
00:30:43
Sarah Frier, CFO of uh Open AAI and u um
00:30:50
she every once in a while uh um she she
00:30:53
really puts out like interesting uh
00:30:55
information and so she was talking about
00:30:58
the cost of a million tokens when ChatG3
00:31:01
came out and it was $32 and change and
00:31:05
now a million tokens cost nine cents.
00:31:08
>> Yeah.
00:31:08
>> Right. And so you you just see like like
00:31:11
the incredible power of how the capital
00:31:15
markets, how capitalism is
00:31:19
uh uh fueling engineering and fueling uh
00:31:22
uh competition.
00:31:23
>> It's become recursive now too. I mean
00:31:25
these models if you say to the model,
00:31:27
hey make yourself more efficient, spend
00:31:29
less money and lower the cost of tokens.
00:31:30
It'll be like okay captain.
00:31:32
>> Yeah.
00:31:32
>> I don't know if you saw Cararpathy's
00:31:34
recursive
00:31:36
>> thing last weekend but it's like now
00:31:38
civilians who've never worked in a
00:31:40
language model or done computer science
00:31:41
are like, I'm going to try to do
00:31:42
something recursive this weekend. You
00:31:44
know, it's one of the things that I that
00:31:47
uh uh I talked to, you know, the other
00:31:51
founders about, you know, and it's like
00:31:55
when you think about some of the things
00:31:57
that AI does, right, it's lowering the
00:32:00
barrier to operations. So if you have a
00:32:03
good idea or a great idea, you can open
00:32:06
up your model and you can tell your
00:32:10
model, you can vibe code it, you can do
00:32:11
all kinds of different things and create
00:32:14
things that never existed before. That's
00:32:16
amazing, right? like that's bringing
00:32:18
down this incredible barrier that kept
00:32:20
human creativity contained and now all
00:32:23
of a sudden this whole new vector of uh
00:32:26
uh you know medical research or
00:32:28
different approaches to you know
00:32:31
baseball cards or whatever you want if
00:32:33
you've got a great idea if you've got a
00:32:34
new creative idea that's the valuable
00:32:37
kernel right now that allows you to to
00:32:40
build new things and to create new
00:32:41
things and I just think that's
00:32:42
incredibly exciting like you're bringing
00:32:45
the minds of 8 billion in people a tool
00:32:48
that allows them to overcome what was
00:32:50
insurmountable for
00:32:52
forever
00:32:53
>> for humanity.
00:32:54
>> Yeah, it's a bright new future. Michael,
00:32:56
appreciate you sharing the uh uh
00:32:58
information with us and the vision. I am
00:33:01
really delighted to have Arvin Shri Nas
00:33:04
on the program.
00:33:05
>> Thank you for having me here Jason.
00:33:07
>> It's so great. I want to go through
00:33:09
three stages in which I fell in love
00:33:11
with your product. The first phase was I
00:33:15
could go in pick my language model if I
00:33:18
wanted to choose open AI, if I wanted to
00:33:19
use claude, whatever it was. That was
00:33:22
like a real unlock for me. And on the
00:33:25
sidebar sidebar, I noticed you had done
00:33:29
essentially like what Yahoo did in the
00:33:32
early days, finance, sports, and when I
00:33:36
pulled my nickname up, it gave me a live
00:33:39
version of that. When I pulled my stocks
00:33:40
up, it summarized the news in real time.
00:33:42
time and I was like, "Wow, this
00:33:44
execution's great." And I I kind of made
00:33:46
you my front door, two different models,
00:33:48
and it made it easier for me to check
00:33:50
it. Then you came out with the Comet
00:33:52
browser and I was like, "Holy cow, I can
00:33:55
give this a series of instructions. Go
00:33:56
to my LinkedIn, find everybody from this
00:33:59
company, put them into a Google sheet
00:34:01
and boom, you were the first out of the
00:34:02
gate with that." And then just the last
00:34:04
couple of weeks I had been claw pilled
00:34:07
in using openclaw but you came out with
00:34:09
computer and I started using computer
00:34:12
and boy it's good uh it's a really
00:34:15
strong start uh allowing me to do
00:34:18
repetitive tasks very similar in some
00:34:21
ways to co-work from claude uh or
00:34:24
basically an engineer or developer using
00:34:26
it. So
00:34:28
are are these the evolution of the
00:34:31
company and I should think about it that
00:34:32
way. But how do you look at perplexity
00:34:34
now? You have a very loyal fan base.
00:34:37
You're making a lot of money. I don't
00:34:38
know if you disclose it but I think it's
00:34:40
hundreds of millions to billions. You
00:34:42
can tell us but what is perplexity in
00:34:44
the face of wow Claude's having a great
00:34:47
run, OpenAI still doing strong. Grock
00:34:49
doing very well. Gemini coming on
00:34:51
strong. There's like six or seven of you
00:34:53
and uh you just happen to be one of my
00:34:55
top twos right now.
00:34:57
>> Thank you. So tell me first of all,
00:34:58
first of all, thank you. Thank you so
00:34:59
much. Perplexity has always been built
00:35:01
for people who are always looking for
00:35:04
the extra edge, the curious people. So
00:35:06
it's very natural that you are uh one of
00:35:09
our power users. Uh one common theme for
00:35:13
us uh for the last three and a half
00:35:16
years is accuracy. Plexity wants to be
00:35:19
the company that's building the most
00:35:21
accurate AI. So when you want to give
00:35:23
somebody answers, accuracy is very
00:35:25
essential for building trust because
00:35:27
only then the user is going to ask the
00:35:29
next set of questions. It turns out it
00:35:31
was a great idea to give AI access to
00:35:34
the internet to be accurate. So that's
00:35:36
the perplexity ask product. It turns out
00:35:38
it's a great idea for AI to have full
00:35:40
access to a browser so that it can be
00:35:43
accurate when you task it to go do
00:35:45
something that you would do yourself on
00:35:46
a browser. Aentic browsing comet. Now
00:35:50
the last phase is it turns out it's a
00:35:52
great idea for AI to be given a full
00:35:55
access to a computer so that it can do
00:35:58
whatever you do on a computer on its own
00:36:01
essentially becoming the computer
00:36:03
itself. an orchestra of everything AI
00:36:07
can do today. every single capability
00:36:09
each individual AI model has be it GPT
00:36:12
or cloud or Gemini or anything else an
00:36:16
orchestra of all those capabilities that
00:36:18
what that's what perplexity computer is
00:36:20
and all these sub agents that are
00:36:23
running inside computer are the
00:36:24
musicians the models are essentially the
00:36:27
instruments and they're like hundreds of
00:36:30
models out there each having their own
00:36:32
specialization some are good at coding
00:36:34
some are good at writing some are good
00:36:35
at multimodal visual synthesis is image
00:36:38
generation, video generation, audio, but
00:36:40
what matters is the end output, the
00:36:42
music you play. That's the work AI gets
00:36:45
done for you. And that's what perplexity
00:36:47
computers. The AI itself is the
00:36:50
computer. Now,
00:36:51
>> still lives inside of a browser. Have
00:36:53
you considered giving it desktop root
00:36:56
access? That feels like the next place
00:36:57
this is going, but that comes with a lot
00:37:00
of security issues, a lot of trust
00:37:02
issues. As you mentioned, trust is
00:37:03
paramount. getting the right answer is
00:37:05
what builds it, but also not getting
00:37:07
hacked and not having it delete your
00:37:09
files. So, how do you think about root
00:37:12
access to my Windows machine? Obviously,
00:37:14
iOS, they won't let you, but with an
00:37:16
Android phone, it would let you.
00:37:18
>> Yes.
00:37:18
>> So, do you have that in the works?
00:37:20
>> Yes. So, we announced something called
00:37:21
personal computer. Perplexity personal
00:37:24
computer that's essentially going to
00:37:26
take all the trust and reliability and
00:37:28
the server side execution of perplexity
00:37:30
computer but synchronize it with your
00:37:33
local computer so that you can use it
00:37:36
from your phone and we're going to do
00:37:37
this with the Mac Mini where you
00:37:40
synchronize your computer with the Mac
00:37:41
Mini so that becomes your local server
00:37:44
all the agent orchestration that has to
00:37:46
do with your local private data will run
00:37:48
on that local orchestration loop that
00:37:50
runtime with the Mac Mini. Not on your
00:37:53
servers, not on anthropics.
00:37:55
>> Exactly.
00:37:55
>> Yeah.
00:37:56
>> It could still ping Frontier models if
00:37:58
it needs to with your permission,
00:38:01
>> but it will be orchestrating everything
00:38:03
on your local hardware.
00:38:05
>> Yeah.
00:38:05
>> And if it needs to run on the server
00:38:07
side hardware, if you don't want very
00:38:09
complicated, longunning stask to be
00:38:12
running on your local hardware. Yeah.
00:38:14
>> You can delegate it to run on your
00:38:16
server side computer, which is again
00:38:18
only accessible to you and you alone. So
00:38:21
that way we're going to bring this
00:38:23
perfect hybrid of trustworthy
00:38:25
uh hybrid between local and server side
00:38:28
and you
00:38:29
>> and you'll make it easy to do. It just
00:38:30
be abstracted. You install one
00:38:32
executable, boom, it's done.
00:38:33
>> It's it's like open claw for dummies.
00:38:35
Nobody needs to learn how to use it.
00:38:37
Nobody needs to manage API keys. Nobody
00:38:39
needs to manage separate billing across
00:38:40
like 100 different services. Figure out
00:38:43
what you can give access to and not
00:38:45
access to. We take care of that.
00:38:47
>> So it's a Steve Jobs way of doing it.
00:38:49
you know, end to end integration
00:38:51
>> and and how do you think about local
00:38:53
models? I have started running Kimmy 2.5
00:38:56
on a Mac Studio.
00:38:58
>> It's not as good as Claude or Gemini or
00:39:00
Grock, but you can probably do about 80%
00:39:03
there for free.
00:39:04
>> Yeah.
00:39:05
>> Essentially.
00:39:06
>> Yeah.
00:39:06
>> Uh and so that's quite compelling
00:39:08
considering some of my other bills,
00:39:09
Claude and and stuff were getting
00:39:11
expensive.
00:39:12
>> So, do you have one of those? You
00:39:15
started testing on your local Mac
00:39:17
Studio. I assume you have a Mac Studio
00:39:18
and you're doing this yourself. Yeah.
00:39:20
>> Or now, I don't know if you saw uh Dell
00:39:22
and Nvidia announced a giant
00:39:24
workstation. Um is it 3,800?
00:39:28
>> Something like that.
00:39:28
>> Something like that with 750 gigs of
00:39:30
RAM. So,
00:39:32
>> what do you think about the desktop
00:39:34
going back to workstation/server?
00:39:37
>> Yeah.
00:39:37
>> Status.
00:39:38
>> I think it's very promising. Um my my
00:39:41
prediction is it'll initially start off
00:39:43
as a sub agent. So whatever you need to
00:39:46
go uh like your tax returns, your
00:39:49
personal photos, your emails, your your
00:39:52
calendar, all that stuff, those local
00:39:55
apps, your personal notes, very personal
00:39:58
notes. You can make sure that the models
00:40:00
that access those tokens will be running
00:40:04
on your local hardware if you want to,
00:40:06
if you're that privacy conscious.
00:40:09
uh and more complicated stuff that
00:40:12
accesses your data that's already on the
00:40:14
server side. Example, your Google
00:40:15
calendar, yeah, your Gmail. This is
00:40:18
personal data still, but an AI runtime
00:40:22
can access that through your connector,
00:40:24
your Google calendar connector, your
00:40:26
Google Workspace connector, and that
00:40:28
could run on the server side because
00:40:29
anyway, the data is on the servers. It's
00:40:31
not even lying on your device.
00:40:32
>> So, that sort of hybrid orchestration is
00:40:35
where we're headed to. I don't think
00:40:37
it's a dichotomy between fully local
00:40:39
versus fully server. Uh it's all about
00:40:42
choice. And anyway, when you're on your
00:40:44
phone, uh you want to you don't care
00:40:46
actually which server that workloads
00:40:48
running from because it's not going to
00:40:49
be able to run on your phone anyway. The
00:40:52
chips need to exist on a Mac Studio or a
00:40:54
Mac Mini and or on the server
00:40:56
>> or this new Dell that's coming out. And
00:40:58
I I really think the idea of spending
00:41:01
$10,000 on a powerful desktop will
00:41:04
appeal to people if it lowers their $500
00:41:08
a month
00:41:09
>> claude bill. This is an incredible
00:41:11
savings. Plus, you get the benefit
00:41:13
>> of privacy and not educating the
00:41:15
language models on your personal data.
00:41:18
>> Yes. And it's going to be it's going to
00:41:19
be like you're buying a refrigerator,
00:41:21
your your your internet modem. Like the
00:41:23
cost for these will eventually go down.
00:41:26
>> Yeah. But it's not going to feel like
00:41:28
you're wasting your money. Uh every
00:41:31
every home has a lot of other sensors.
00:41:34
>> Yeah.
00:41:34
>> That runs your home that'll also be part
00:41:37
of this orchestration loop.
00:41:39
>> Yeah.
00:41:40
>> So, so that's where it gets exciting
00:41:41
because now you can just dictate
00:41:43
something to your phone and that can
00:41:45
control your entire home.
00:41:48
>> So that's the dream that everybody has
00:41:49
and all that orchestration loop can run
00:41:52
on your local hardware, no problem. And
00:41:54
I'm curious what you think of the
00:41:56
operating system. What's eventually
00:41:59
going to be the operating system of this
00:42:02
workstation?
00:42:03
>> AI is the operating system. Like earlier
00:42:05
in the traditional operating system, you
00:42:07
execute programmatically.
00:42:09
Now you start with objectives, not
00:42:11
specific instructions.
00:42:13
>> Right?
00:42:13
>> You come up with a highlevel objective.
00:42:15
go build this website for me that you
00:42:18
know takes all the transcripts of all in
00:42:20
podcast and tracks the stock price just
00:42:22
before the podcast and after. Yeah.
00:42:24
>> And charted for the max 7.
00:42:26
>> Yeah.
00:42:26
>> And and charted over time you can so
00:42:28
that's the objective but individually
00:42:31
it's running a file system a code
00:42:33
sandbox access to the internet. It's
00:42:35
having like its own HTML tools and like
00:42:38
so I think that's basically where you
00:42:40
know models systems and files and
00:42:43
connectors are all coming together. You
00:42:44
would think of that as an OS
00:42:46
>> except you're operating at an
00:42:48
abstraction about that where you're
00:42:50
thinking in terms of objectives.
00:42:52
>> Yeah. And does it need to eventually
00:42:55
become its own operating system in your
00:42:57
mind?
00:42:58
>> It could be like people could think
00:43:00
about like yeah I have a my perplexity
00:43:02
computer running all the time whether it
00:43:05
essentially it runs on Linux machines
00:43:07
right now. Every server side computer is
00:43:09
a Linux machine. Yeah.
00:43:11
>> So, I think Mark Anderson tweeted this
00:43:13
right after our release that turns out
00:43:15
Linux computers was the right idea.
00:43:17
Desktop desktop Linux computers are
00:43:19
finally going to work.
00:43:20
>> Yeah. I mean, they're stable. They're
00:43:22
customizable and you're not at the mercy
00:43:25
of Apple's desire to contain the
00:43:28
experience or Microsoft surface area as
00:43:32
for hackers.
00:43:33
>> Exactly.
00:43:34
>> You build something rock solid and it
00:43:36
does feel like Linux might actually
00:43:38
become the correct
00:43:39
>> the eventual winner. It may not need to
00:43:41
have a front end.
00:43:42
>> That's the thing. You could you could
00:43:44
access the Linux machine on your phone.
00:43:47
>> You could be running iOS or Android. It
00:43:49
doesn't matter.
00:43:50
>> The actual valuable runtime is running
00:43:53
on Linux on the server.
00:43:55
>> You've done great as a consumer company.
00:43:58
Lot of love there. Now I'm starting to
00:44:00
see uh corporations with computer
00:44:04
starting engaging it. In fact, you'll be
00:44:06
happy to know this. Last week, I took
00:44:08
two people in my back office and I said,
00:44:10
"Stop working on OpenClaw. Your job is
00:44:13
to do the back office automation at our
00:44:16
venture firm only using Perplexity." And
00:44:19
they were like, "Perplexity computer."
00:44:21
And they were like, "Oh, okay. Um, it
00:44:24
doesn't talk well in Slack. It doesn't
00:44:26
have an agent in Slack." I was like, "It
00:44:28
will. I'm going to see AR and I'll talk
00:44:31
to him about that." So, we need a really
00:44:33
strong Slack connector.
00:44:34
>> It's already out.
00:44:34
>> It is. Okay, great. computer exists as a
00:44:37
Slackbot right now.
00:44:38
>> Okay,
00:44:38
>> that you can add to your Slack workspace
00:44:40
on enterprise plan
00:44:41
>> and our entire company works like that.
00:44:44
People are talking more to computer on
00:44:45
Slack to other than to other people.
00:44:47
>> In our first volley, we were sending
00:44:49
reports in, but it wasn't interactive.
00:44:51
That's perfect.
00:44:54
So now you've got your company going in
00:44:56
two different directions. This
00:44:57
incredible consumer run you have. How
00:44:59
many people are using the product every
00:45:01
month?
00:45:01
>> Several tens of millions. So tens of
00:45:03
millions of people that's very much
00:45:05
similar to the trajectory of the Google
00:45:07
and Yahoo consumer business. Now you've
00:45:09
got corporate. How are you doing on the
00:45:11
corporate side? Thousands of companies.
00:45:13
>> The fastest growing business for us. Ah
00:45:15
>> it's growing faster than the consumer in
00:45:17
revenue and things like computer unlock
00:45:20
entirely new possibilities. For example,
00:45:22
we've saved more than $und00 million for
00:45:25
our uh enterprise max customers who are
00:45:27
on the highest tier of enterprise.
00:45:29
>> Explain what that is. What does it cost?
00:45:31
200 a month per person.
00:45:32
>> So there are two tiers. One is the
00:45:34
enterprise pro which is $40 a month and
00:45:36
there's the enterprise max which is $400
00:45:39
a month. And that that and and and and
00:45:42
on a computer after you run out of your
00:45:44
credits you would pay for the tokens.
00:45:46
You pay for the usage.
00:45:48
>> Are you making money on the $400 a
00:45:50
month, $5,000 a year one or at this
00:45:52
point in time are people going so crazy?
00:45:54
Our uh one thing that Perplexity has is
00:45:57
every revenue we make, unlike certain
00:45:59
other rapper companies, every revenue
00:46:01
Perplexity makes has positive gross
00:46:02
margins.
00:46:03
>> Got it.
00:46:04
>> Because uh we're not just selling
00:46:06
tokens,
00:46:06
>> right?
00:46:07
>> Most of our revenue is recurring because
00:46:09
people are paying a subscription fee
00:46:11
>> and because we route through multiple
00:46:13
different models, we're very efficient
00:46:15
in terms of how we spend on the tokens.
00:46:18
because we have all this advantage with
00:46:19
rag and orchestration and search. We
00:46:22
don't actually need to blow up the
00:46:23
context window of the models.
00:46:25
>> Yeah.
00:46:25
>> As a result of that, we have positive
00:46:28
gross margins on all the revenue. Every
00:46:30
single penny we make, we make profits on
00:46:32
that. But the overall the company is
00:46:34
still yet to be profitable, but we're
00:46:35
working towards that.
00:46:36
>> You've had the opportunity to exit. A
00:46:39
lot of rumors, Apple, other people were
00:46:41
like, "Hey, this is a great team." How
00:46:42
many people on the team now?
00:46:44
>> About 400.
00:46:45
>> Yeah. You you've got a very coveted
00:46:46
team. You obviously understand consumer.
00:46:48
You obviously understand business. It's
00:46:50
a product driven organization. Reports
00:46:52
are you declined,
00:46:55
but the world's getting hyper
00:46:56
competitive here. How do you keep up as
00:46:58
a 400 person organization when you got
00:47:01
Sam Alman over here raising a hundred
00:47:04
billion dollars, you know, and then you
00:47:06
have Elon putting data centers in space
00:47:08
and merging with SpaceX and Twitter. You
00:47:11
have Google with unlimited resources.
00:47:13
Amazon getting in the game and obviously
00:47:16
Gemini uh very strong product and Google
00:47:20
really good at consumer. I think we'd
00:47:22
all agree Facebook and Meta haven't
00:47:25
figured it out yet except maybe for
00:47:27
serving us better ads, but they they
00:47:28
haven't figured out the consumer case
00:47:30
yet, but they'll copy it. They always
00:47:31
do.
00:47:32
How do you look at the playing field?
00:47:34
Because the degree of difficulty, this
00:47:37
isn't playing checkers or this is like
00:47:40
playing against the 10 best chess
00:47:43
players in the world. That's what you
00:47:44
have to do every day.
00:47:46
>> So, how do you think about it? Long-term
00:47:47
and independent company. Do you think
00:47:49
you'll need to join forces at some
00:47:51
point?
00:47:51
>> Well, and why didn't you take the deal?
00:47:53
This deals were incredible that you got
00:47:54
offered.
00:47:55
>> So, one advantage we have that all these
00:47:59
companies you mentioned don't have is
00:48:01
the multimodel orchestration. We're like
00:48:04
Switzerland. We don't have to have one
00:48:06
horse in the race. If GPT wins, Gemini
00:48:09
wins, Claude wins, Llama wins, it
00:48:12
doesn't matter to us. Uh or even open
00:48:15
source models can win, no problem.
00:48:16
>> And you have them on the service. You
00:48:18
have DeepSeek and Kimmy.
00:48:20
>> We have Kimmy, we have Neotron, and we
00:48:22
have uh a lot of usage of Quen, Alibaba
00:48:25
Quen.
00:48:26
>> Yeah.
00:48:26
>> Silently under the hood. So for us like
00:48:29
that advantage of being able to take the
00:48:31
best in each model and give the user the
00:48:35
orchestra of everything they can do. I
00:48:37
don't think any of the companies you
00:48:38
mentioned can do that
00:48:39
>> right nor would they
00:48:40
>> nor would they it makes no sense for
00:48:42
them. It would be an admission that all
00:48:44
the data centers and capex they've built
00:48:46
out mean still couldn't produce them the
00:48:49
best model. And uh Daario uh CEO of
00:48:52
Anthropic said recently in an interview
00:48:54
uh that models are specializing. Towards
00:48:58
the beginning of last year people
00:48:59
thought models are going to commoditize
00:49:01
but towards the end of last year people
00:49:04
models started specializing. Even within
00:49:06
coding
00:49:08
u cloud code and codeex have very
00:49:11
different capabilities. Our iOS
00:49:13
engineers love using codeex. Our backend
00:49:16
engineers love using cloud code.
00:49:17
>> Yeah. So even within a specialization
00:49:19
like coding, models have their own
00:49:22
unique specialtities and there are many
00:49:24
other use cases outside coding where
00:49:26
different models are good at different
00:49:27
things. Which means the orchestra
00:49:29
conductor that has no one model to the
00:49:32
horse in the race can win by providing a
00:49:35
very unique value and service to the
00:49:37
customer that each of these amazing
00:49:40
names that you mentioned cannot. And so
00:49:42
you're buying tokens wholesale from them
00:49:45
and then you'll charge customers to do
00:49:48
it or do you think it's all
00:49:50
>> we're going to take care of all that
00:49:51
orchestration?
00:49:52
>> Yeah.
00:49:52
>> So you don't have to manage tokens
00:49:54
across different models
00:49:55
>> cuz I authenticate I a couple of my
00:49:58
different accounts my pro accounts into
00:49:59
perplexity. But does it I I I don't have
00:50:02
enough knowledge to know if you're
00:50:04
abstracting that and people can just
00:50:05
search across them and it's part of
00:50:07
their perplexity subscription. No, we're
00:50:09
not bundling subscriptions from into
00:50:11
other AIS.
00:50:12
>> We just ping the models directly.
00:50:14
>> Got it.
00:50:14
>> Uh what you get in us is the perplexity
00:50:17
or orchestration.
00:50:18
>> Got it.
00:50:18
>> The harness,
00:50:19
>> right?
00:50:20
>> So the when when when when models are
00:50:22
kind of specializing the there's a
00:50:25
bigger value in the one who knows how to
00:50:27
build a great harness,
00:50:28
>> right?
00:50:28
>> That can take the best in each model.
00:50:30
>> Does it auto route today or do you still
00:50:32
have the drop down somebody's got to
00:50:34
pick
00:50:34
>> it? It definitely auto routes the best
00:50:36
model for each prompt,
00:50:37
>> but we also give users the flexibility
00:50:39
to pick whatever model they want.
00:50:41
>> What do you think of I've seen a bunch
00:50:43
of startups hack this together, but
00:50:45
doing the same query across multiple
00:50:47
>> We built a thing called model council.
00:50:49
>> Model council. Yeah.
00:50:50
>> Yeah. So that's one of the one of the
00:50:52
modes and perplexity where I saw Jensen
00:50:54
say in one of the interviews that he he
00:50:57
puts the same prompt in five different
00:50:58
AIs and sees what each of them says.
00:51:00
>> Yes.
00:51:01
>> Like everybody does that. Yeah. But then
00:51:03
you still have to apply your biological
00:51:05
computers
00:51:10
about your trust or your
00:51:12
>> five different doctors.
00:51:13
>> Five different doctors trying to figure
00:51:14
it out.
00:51:15
>> Exactly. So it's dumb.
00:51:16
>> So the model council is a feature we
00:51:18
built where it will not just give you
00:51:20
the answers of each model, but it will
00:51:22
tell you exactly where they agree, where
00:51:23
they disagree, and where the nuances
00:51:24
are.
00:51:25
>> And that's in the interface. Model
00:51:26
council, I didn't know it was there.
00:51:27
>> It's there.
00:51:28
>> I mean, you you released product at a
00:51:30
pretty great cadence, huh? How where did
00:51:32
you learn that and what's your
00:51:34
philosophy of shipping product?
00:51:36
>> Our philosophy is like speed is our
00:51:38
mode. Like you know again one of the
00:51:40
things that big companies cannot do is
00:51:41
move at the speed we do serve customers
00:51:43
at the speed and qual it's it's very
00:51:45
hard to maintain quality speed and trust
00:51:47
at the same time.
00:51:48
>> Yeah.
00:51:49
>> Like Apple takes a long time to ship
00:51:51
anything
00:51:51
>> because they're very worried about
00:51:53
people not trusting them.
00:51:54
>> Yeah.
00:51:54
>> Uh and so some companies are
00:51:56
bureaucratic and they just take forever
00:51:58
to ship something. They don't maintain
00:52:00
what they ship. They may make a big deal
00:52:01
about an event but nobody even knows how
00:52:04
to go and use that feature.
00:52:05
>> Yeah. They get abandoned.
00:52:06
>> Exactly. So, Perplexity has those
00:52:08
advantages for being very small. And
00:52:10
towards the end of last year, we found
00:52:12
that like AI coding tools have made it
00:52:14
much faster for us to ship things
00:52:16
>> which is honestly one of the reasons why
00:52:18
we built computer because now even
00:52:20
non-engineers are shipping code here by
00:52:22
just pinging a slack bot and asking it
00:52:24
to fix bugs.
00:52:25
>> So, this the the iteration has just been
00:52:27
like exponential. The the moment I had
00:52:30
where I became clawilled was when I was
00:52:34
working with it and I was like, "Hey, I
00:52:36
want to build my network. I know these
00:52:38
20 people in Japan. I had dinner with
00:52:39
them during my recent trip. I want to
00:52:41
know who they know. So, check out
00:52:43
LinkedIn and other things and who
00:52:45
they're associated with and make me like
00:52:46
a mind map of it. And then the next trip
00:52:49
I want to meet with the next circle of,
00:52:52
you know, those connections." So, I
00:52:54
started asking like, "Okay, I got the
00:52:55
results." I was like, "Great."
00:52:58
Um, and they said, "Where do you want me
00:52:59
to put them?" And uh, I was like, "Well,
00:53:01
where can you put them?" And it said,
00:53:02
"Well, I can put it in a Google sheet. I
00:53:04
can put it in notion table. I can put it
00:53:05
here. I can give you a PDF. I can give
00:53:06
you a CSV file. Or I could write you a
00:53:09
CRM." And I was like, "Yeah, sure. Make
00:53:12
me a CRM system." And it may a CRM
00:53:15
system.
00:53:16
>> And I think that becomes, and I think
00:53:18
maybe one out of a thousand people
00:53:20
working with AI have had that
00:53:21
experience. Maybe it's one in 10,000.
00:53:24
Where your agent says, I'll make you
00:53:26
bespoke software.
00:53:27
>> Yeah.
00:53:28
>> Have you had that yet? And and do you
00:53:30
see that as a part of computer that when
00:53:33
a person needs a spreadsheet, you don't
00:53:35
launch Excel or Google Sheets, you just
00:53:38
pop up a spreadsheet?
00:53:40
>> Yeah. Well, we have a board meeting
00:53:42
tomorrow.
00:53:42
>> Okay, I'll come.
00:53:43
>> And and so
00:53:44
>> I'll pitch it to the board.
00:53:45
>> Sure.
00:53:46
>> Uh our computer computer made the memo.
00:53:50
>> Oh, wow.
00:53:50
>> Yeah. And um we had a partner meeting to
00:53:53
pitch a partnership idea and uh earlier
00:53:56
we would have a design team do the whole
00:53:58
deck.
00:53:58
>> Yeah.
00:53:59
>> Computer just oneshotted it. Uh I had a
00:54:02
press briefing with a bunch of
00:54:03
journalists. My comm's person would
00:54:05
>> Sorry about that. Brutal.
00:54:07
>> And then my comms person would usually u
00:54:09
give me a memo what to say.
00:54:12
>> Computer one-shoted him.
00:54:13
>> So
00:54:14
>> it's crazy. And it's the context is so
00:54:17
good because the memor is getting
00:54:18
better. Yeah. Yeah.
00:54:19
>> So it's like I know that journalist from
00:54:22
the last time.
00:54:24
>> I know the board meeting. I have all the
00:54:26
previous decks.
00:54:28
>> When did that happen?
00:54:30
>> I think it it happened with Opus 45
00:54:34
>> Opus 45. That was a inflection point
00:54:36
when models were started being amazingly
00:54:40
good at orchestration and reasoning and
00:54:42
tool calls and cloud code brought in
00:54:45
this new idea in AI that everything can
00:54:47
happen inside a sandbox, a console, a
00:54:51
terminal with access to tools where
00:54:54
tools are just command line tools.
00:54:56
>> Yeah,
00:54:56
>> they don't even need to have graphical
00:54:58
user interface. So when you did that and
00:55:00
when you organize around files and sub
00:55:03
aents and skills and CLIs, the model
00:55:06
started be becoming very good at
00:55:09
handling the context. So the context
00:55:11
window no longer became a problem. It
00:55:14
just put whatever necessary into the
00:55:15
context whenever it wanted to and dumped
00:55:17
dumped them away when it wanted to.
00:55:19
>> Yeah.
00:55:19
>> And that made it like suddenly so good
00:55:21
at doing very long orchestration tasks.
00:55:26
>> Yeah. It's it's pretty crazy. I have
00:55:28
every episode of this week in startups,
00:55:30
all the transcripts and then all of all
00:55:32
in
00:55:32
>> that was one of the tasks I did by the
00:55:33
way I can send it to you. I asked it I
00:55:36
want you to download every all-in
00:55:37
podcast.
00:55:38
>> Yeah.
00:55:38
>> U since the beginning and I want you to
00:55:42
take a mention of all the public
00:55:44
companies they mentioned during the
00:55:46
episode.
00:55:46
>> Yes.
00:55:46
>> I want you to have a histogram of the
00:55:48
counts and I also want you to chart it
00:55:51
across time and then I want you to
00:55:53
analyze the impact on the stock price
00:55:55
>> and the sentiment of what we said.
00:55:57
Exactly. And it did like it clearly
00:55:59
said,
00:55:59
>> "Are we moving stocks
00:56:00
>> around Google's stock going up?"
00:56:02
>> Yes.
00:56:03
>> Prior to that, you guys were talking a
00:56:04
lot about Google.
00:56:05
>> Yes.
00:56:06
>> And it clearly
00:56:06
>> And I said I made a bet publicly on the
00:56:08
thing. I said, "I am buying a bunch of
00:56:10
Google because I believe even though
00:56:12
they're behind,
00:56:13
>> it's because they're too precious." You
00:56:15
were kind of mentioning a company that
00:56:16
might be too precious at times and
00:56:18
doesn't release.
00:56:18
>> I was like, "That's that company. They
00:56:20
need to release more." Yeah.
00:56:21
>> And uh I told Sergey, I was like,
00:56:23
>> like
00:56:25
>> give us the good stuff. started giving
00:56:27
us the good stuff.
00:56:28
>> It literally gives you the timestamps of
00:56:29
every single and then I can go click on
00:56:31
it and actually hear
00:56:33
>> exactly
00:56:34
>> that moment.
00:56:34
>> Yeah.
00:56:35
>> Sweet.
00:56:35
>> Yeah. So that's when that's when I was
00:56:37
like damn like
00:56:38
>> this I would have had somebody do this
00:56:40
as a weekl long project.
00:56:42
>> It would have been 10 hours a week of
00:56:44
researcher. I I'm experiencing the same
00:56:46
thing when I do research notes. I've
00:56:49
created my own uh like mega prompt.
00:56:54
>> Yeah. and it will go and like tell me
00:56:57
where you worked before and who's in
00:56:59
your circle, who your competitors are,
00:57:01
who your friends are, blah blah blah,
00:57:02
and then go find I try to find old
00:57:05
podcast is one of my secrets. If you're
00:57:06
an interviewer watching, I try to find
00:57:09
what was the person talking about 5
00:57:10
years ago, 10 years ago, and then over
00:57:13
10 years ago. And I've gone into
00:57:14
interviews now with Michael Dell and
00:57:16
talked about things he was talking about
00:57:18
in the '9s. Yeah.
00:57:19
>> And it finds me some ancient stuff. Like
00:57:21
you would pay a research or a producer,
00:57:24
>> you know, $70,000 a year, $80,000 a year
00:57:27
to do this and they would have done a
00:57:29
third of the job in 10 times longer.
00:57:33
>> It's really gotten weird just in the
00:57:35
last 6 months. What do you think the
00:57:36
next 6 months looks like?
00:57:38
>> I think the the dream that what we are
00:57:40
going to try to do is help businesses
00:57:42
run as autonomously as possible. You
00:57:45
know, everybody talks about this AI is
00:57:47
going to create this one person $1
00:57:48
billion company. Some people say it's
00:57:51
already happened because people pay
00:57:53
researchers like 1 billion, but it's not
00:57:55
truly moving the GDP by 1 billion. It's
00:57:58
not truly creating new value. So the
00:58:01
best way to do that is to actually help
00:58:02
a small business people who would
00:58:04
otherwise drive Ubers for like yes
00:58:06
>> extra passive income to like buy like a
00:58:10
Mac mini set up perplexity personal
00:58:12
computer and run their business on that
00:58:14
or like run it on the server it doesn't
00:58:16
matter uh and actually make real money.
00:58:19
>> Yeah.
00:58:19
>> Hundreds of thousands or even millions a
00:58:20
year
00:58:21
>> and uh grow it.
00:58:24
>> Have computer go and run your ad
00:58:25
campaigns on Instagram or Google. I mean
00:58:28
>> integrate with SEM and SEO tools, find
00:58:30
new users and uh integrate with Stripe,
00:58:34
charge them, ship new features, have
00:58:36
your own like intercom integration for
00:58:39
customer support and like have this all
00:58:41
working well. You can be sipping wine in
00:58:42
Napa. That's the dream that you know it
00:58:45
feels awesome to say. Everybody thinks
00:58:47
AI is already there. It's not there yet.
00:58:49
Someone has to do that hard work.
00:58:51
>> Yeah,
00:58:51
>> that's what we want to do. Yeah, it it's
00:58:53
a great vision because
00:58:56
when I watched startups 20 years ago,
00:58:58
there were so many check boxes they had
00:59:00
to do. I have to find an office space. I
00:59:01
got to put up a bunch of servers. I I
00:59:04
got to hire hire an HR firm. I I got to
00:59:07
hire a PR person. All this stuff. And
00:59:09
now I talk to young founders. They got a
00:59:11
three-person team. They've come out of
00:59:13
A16Z, my program, Launch Accelerator,
00:59:16
whatever it is, Y Combinator. And I'm
00:59:18
like, "Okay, you raised a half million,
00:59:19
you raised a million. Who are you
00:59:21
hiring?" And they're like, "Um, I don't
00:59:23
know if we need to hire anybody." I'm
00:59:24
like, "If you could hire somebody, would
00:59:26
you hire?" They're like, "Well, I do my
00:59:28
own HR. I have this partner." And
00:59:30
they're I'm like, "How are you doing
00:59:32
hiring anyway?" And they're like, "Well,
00:59:34
I put out an ad and then uh it sorts and
00:59:36
ranks the candidates and then it emails
00:59:40
the top 10, asks them a bunch of
00:59:41
questions, and then I meet with the last
00:59:42
two." And I'm like, "That's what a
00:59:44
recruiter did."
00:59:45
>> Like, the entire recruiting job has been
00:59:47
abstracted. And like a a tool like
00:59:49
computer is going to make that even
00:59:51
faster.
00:59:52
>> Much work to do. Uh lot of connectors, a
00:59:55
lot of specific workflows. People don't
00:59:58
want to like learn how to write like,
01:00:00
you know, essay long prompts. You know,
01:00:01
it needs to be so quick and fast and
01:00:03
autonomous. You just set it up and done.
01:00:06
>> And you have an idea, you can turn it
01:00:07
into a business and start making money.
01:00:09
>> Yeah. It's it's an incredible future. Uh
01:00:12
and it feels like it's right here. Do
01:00:13
you how do you think about job
01:00:15
displacement? is you're actually making
01:00:17
the tool that enables people
01:00:20
>> to be a solo entrepreneur and get to a
01:00:22
million in revenue, but it's also the
01:00:23
same tool that doesn't require them to
01:00:25
hire. And we've had this debate a
01:00:26
million times on the podcast.
01:00:28
>> Do you
01:00:30
I'm wondering if like me, you have
01:00:32
moments where you're like, "Oh my god,
01:00:33
this is really terrifying." Yeah.
01:00:35
>> A lot of people are going to lose their
01:00:36
jobs really fast.
01:00:37
>> Yeah.
01:00:38
>> And then, oh my god, you can learn any
01:00:40
skill you want and all the things that
01:00:42
were hard are now easy.
01:00:44
>> Yeah. I I go back and forth. I'm 70 80%
01:00:48
super positive about this, but I do
01:00:49
worry about like 20% of the time I'm a
01:00:51
little worried. Yeah. Where do you sit?
01:00:53
>> I mean, America has always been about
01:00:54
like entrepreneur entrepreneurship,
01:00:57
right? Like we we've been about like
01:00:58
trying to build new things, discover new
01:01:00
things, go explore.
01:01:02
>> Uh I think this whole like Henry Ford
01:01:04
came and built factories and brought in
01:01:06
jobs and things like that and like put
01:01:09
people into a box. But u I think the
01:01:12
reality is people most people don't
01:01:15
enjoy their jobs. They're doing it for
01:01:16
they hate them.
01:01:17
>> Exactly.
01:01:18
>> So there is suddenly a new possibility a
01:01:20
new opportunity to go use these tools,
01:01:22
learn them and start your own mini
01:01:24
business. And if it pays for your needs
01:01:27
for year or multiple years and lets you
01:01:30
have a high quality life and good work
01:01:32
life balance and true feeling of agency
01:01:34
and ownership and passion to like get
01:01:36
your ideas out there. I think that is
01:01:39
even if there is temporary job
01:01:41
displacement to deal with that sort of
01:01:43
glorious future is what we should look
01:01:45
forward to.
01:01:45
>> I I I think you're exactly right. If
01:01:48
there will be some displacement, but
01:01:49
then there's also going to be so many
01:01:51
opportunities open up and it requires
01:01:53
the individual to not be passive.
01:01:55
>> Exactly.
01:01:55
>> They have to be rugged individualists.
01:01:58
They have to be resilient. Yeah.
01:01:59
>> And they have to be resourceful. And I
01:02:01
think once you start playing with these
01:02:02
tools, that's what happens.
01:02:04
>> Exactly. you you all of a sudden feel
01:02:06
like
01:02:06
>> it brings out the best in you if you
01:02:07
truly are in a good space.
01:02:09
>> Yeah.
01:02:09
>> Yeah.
01:02:10
>> I today uh Comet for iOS is out.
01:02:14
>> Yeah.
01:02:14
>> I'm a Comet super fan. I required
01:02:17
everybody. You were nice enough when I I
01:02:18
emailed you. I was like, "Can you send
01:02:19
me some licenses?" You sent You don't
01:02:21
may not remember. You sent me a bunch of
01:02:22
licenses. I said, "Everybody put this on
01:02:24
because it was $300 a month when you
01:02:26
first came out with the common browser.
01:02:28
Now it's free, I think, for all users.
01:02:30
>> Highly recommend it. Highly recommend
01:02:32
getting a pro account. It's only 20
01:02:34
bucks a month to get into perplexity,
01:02:35
which is a joke. So, you can get on
01:02:37
board for nothing, less than a dollar a
01:02:39
day.
01:02:40
>> But what does iOS allow me to do? And
01:02:43
and how does it connect to computer?
01:02:45
Because that's another thing I'm having.
01:02:47
>> Yeah.
01:02:48
>> Cloud code. Uh computer, there's not a
01:02:51
good enough integration with this mobile
01:02:53
device yet.
01:02:54
>> Yeah. So, computer is already on the
01:02:56
perplexity app. So, you can just toggle
01:02:58
the computer and start using it.
01:03:00
uh comet's uniqueness and perplexity for
01:03:03
the company uh and and and the strategy
01:03:05
is the fact that you can control the
01:03:08
browser. So the browser also becomes a
01:03:10
tool for computer
01:03:12
>> just like your Google workspace and all
01:03:14
these other things. uh until the whole
01:03:17
world is organized around CLI and tools.
01:03:20
>> Yeah,
01:03:20
>> there's still a lot of tasks we have to
01:03:22
do manually on the web on the browser.
01:03:25
Open tabs, fill up forms, click on
01:03:26
things, upload stuff, all that stuff. If
01:03:29
you want to automate, you need a
01:03:31
browser. You need an AI that can
01:03:33
natively control the browser. So that is
01:03:35
comet. And that's why no matter how many
01:03:38
other tools in the market exist like
01:03:40
open claw or like claw co-work
01:03:43
>> executing tasks on a browser on the
01:03:45
server side along with all the other
01:03:47
things is something uniquely perplexity
01:03:49
can do.
01:03:50
>> Yeah. My dream is that you'll create an
01:03:53
Android app that roots my Android phone.
01:03:56
>> Yeah.
01:03:57
>> And that you just take over and see
01:03:58
everything because one of the blockers I
01:04:00
have now is some of the websites have
01:04:03
gotten a little pnicity.
01:04:05
>> Yeah. I don't want to mention too many,
01:04:07
but Reddit, LinkedIn.
01:04:09
>> Yeah.
01:04:10
>> And like they're just I I am a great
01:04:13
Reddit user. I'm a great LinkedIn
01:04:15
supporter, but sometimes like I need to
01:04:17
get my inmail.
01:04:18
>> Yeah.
01:04:19
>> From my LinkedIn and I just need to, you
01:04:22
know, find seven people at company. I is
01:04:24
there going to be a solution
01:04:27
>> between the LinkedIn and Reddits of the
01:04:29
world and the claws and perplexities? Is
01:04:32
how is that
01:04:33
>> I mean
01:04:33
>> negotiation going? You don't have to
01:04:35
speak about any specific ones unless you
01:04:36
want to,
01:04:37
>> but it feels like there's got to be a
01:04:39
solution
01:04:40
>> and I'm willing to pay for it as a user.
01:04:42
I'm willing to play Reddit to allow my
01:04:44
bot to show up and behave properly.
01:04:46
>> Well, I I I cannot speak about any
01:04:48
particular company, but we are happy to
01:04:51
work with anyone, right? So, um I think
01:04:54
with with Comet, our idea is to give
01:04:56
people the flexibility to set things up
01:04:58
on their own.
01:04:58
>> Yeah. and uh any um official APIs that
01:05:02
anyone's willing to offer, we're always
01:05:04
happy to put that as part of computer.
01:05:07
Here's what I think should happen. Let
01:05:08
me see if you agree. Um and this is for
01:05:11
Steve Huffman at Reddit.
01:05:14
I go on Reddit. I do a pro account for
01:05:18
20 bucks a month. And when I do that, I
01:05:20
can authenticate whatever tool I want um
01:05:24
to do a series of well- behaved things a
01:05:28
certain number of times a day.
01:05:30
>> Yeah.
01:05:30
>> So, it's not unlimited. I'm not going to
01:05:31
scrape the whole site, but I would like
01:05:33
it to just let Perplexi or computer go
01:05:37
and just tell me, hey,
01:05:39
>> what are people saying on the this
01:05:40
weekend startups and all-in subreddits?
01:05:42
Summarize it for me so I get the
01:05:44
customer feedback. And I would literally
01:05:46
name my
01:05:48
uh agent and I would say I it won't post
01:05:51
on my behalf. It won't vote on my
01:05:53
behalf. Just needed to do a couple of
01:05:54
little readonly things. This would be an
01:05:56
easy solution. Or LinkedIn I would like
01:05:58
if you I have I already pay LinkedIn
01:06:00
like 50 bucks a month. Like they should
01:06:01
just let the $50 a month one work with
01:06:04
computer.
01:06:04
>> Yeah, absolutely. I mean,
01:06:06
>> okay, this is for Satia Nadella. Let
01:06:09
LinkedIn work with Perplexity and the
01:06:11
other players and we'll pay you extra.
01:06:14
>> Perfect. It's a revenue stream. Don't
01:06:15
you think API access for our customers
01:06:17
is a revenue stream?
01:06:18
>> I think so. I think so. I think I think
01:06:20
fundamentally giving users a choice
01:06:23
>> and setting it up as a win-win for both
01:06:24
the business and the user
01:06:26
>> Yeah.
01:06:26
>> is where the world should head to.
01:06:28
>> And and I I would say the same thing
01:06:30
applies to any any website in the world.
01:06:32
Like if if you want an AI to use it on
01:06:34
your behalf, it should be okay for cuz
01:06:36
that's what the user wants.
01:06:37
>> I mean, I have a paid New York Times
01:06:40
subscription. like let me go in there
01:06:42
and do you know whatever 100 searches a
01:06:45
day, a week, a month, whatever they
01:06:47
choose, but that would make the
01:06:49
subscription that much more sticky.
01:06:51
>> Exactly.
01:06:51
>> Uh all right, Arvin, love the product.
01:06:55
Anybody at home,
01:06:56
>> it's just tremendous. Go learn computer
01:06:59
and get the Comet browser. It has
01:07:01
changed my business for the last two
01:07:03
years. Love the product and we'll have
01:07:06
you back soon when you launch your
01:07:07
operating system and come up with your
01:07:09
own server and desktop server but
01:07:11
business is the focus. Yeah.
01:07:13
>> Yes.
01:07:13
>> All right. Great seeing you.
01:07:15
>> We have an amazing guest Arthur
01:07:16
Manchester here the CEO of Mistral AI.
01:07:19
How are you doing sir?
01:07:21
>> Great. Thank you for having me.
01:07:22
>> And so you're here at Nvidia's big
01:07:26
conference,
01:07:28
big announcement. You're going to be
01:07:30
working with Nvidia to build models.
01:07:33
uh to open source them. What is the uh
01:07:36
big announcement here?
01:07:37
>> Well, we're announcing that we are going
01:07:39
to be training the next generation of
01:07:41
frontier models with uh with Nvidia. Um
01:07:44
it's something that we've been doing
01:07:45
before with Nvidia with MLMO, something
01:07:47
we did like 18 months ago. And the point
01:07:49
for us is really to be able to produce
01:07:51
the best open source models out there so
01:07:53
that we can actually use those assets to
01:07:55
specialize them through products that we
01:07:57
do for our customers like Forge that
01:08:00
helps us customize the models for the
01:08:02
enterprise we work with in engineering
01:08:04
in physics in science uh in making them
01:08:06
better at certain languages when we work
01:08:08
with governments etc.
01:08:10
>> And and Michel obviously based in uh
01:08:12
France you're the leading AI company
01:08:15
there. What's it like running the
01:08:17
company and building a large language
01:08:19
model in Europe? Obviously, there's
01:08:20
regulations and all kinds of
01:08:23
considerations. Privacy, the French are
01:08:24
known for protecting privacy. In the
01:08:26
United States, we're known for taking it
01:08:28
away. How is the landscape there and
01:08:31
what do you have to deal with there that
01:08:32
maybe you wouldn't have to deal with in
01:08:34
America? And what's the pros and the
01:08:35
cons? I'd say first, we have 25% of our
01:08:38
business in the US. Uh, and 25% of our
01:08:40
researchers are actually here. So I
01:08:42
actually spend a lot of time here as
01:08:44
well as in France as well as in the UK
01:08:46
in Singapore where we are. So of course
01:08:48
it's it's different markets. Uh it's
01:08:51
markets where you have language which is
01:08:52
a topic uh where there's much more
01:08:55
manufact manufacturing is a bigger piece
01:08:57
of the cake than it is here. uh and I'd
01:09:00
say the our strength has been to also
01:09:02
work with European companies that are a
01:09:04
bit lagging behind uh and that wants to
01:09:06
adopt the technology to to leap forward
01:09:09
and we've been able to do that through a
01:09:10
forward deployment engineering
01:09:11
engagement through our forge product for
01:09:13
our studio product that allows to deploy
01:09:15
agents that do end to end automation but
01:09:18
on top of that the thing that we have
01:09:19
announced today like forge is something
01:09:21
that is actually being used today uh
01:09:23
with customers in the US because they
01:09:25
come to us with uh needs for post
01:09:28
training for making mod specifically
01:09:29
good at financial services and what's
01:09:32
happening is that we have this product
01:09:33
and we can bring the models to
01:09:34
specialize them as well.
01:09:36
>> And so your belief is specialized
01:09:39
verticalized models healthcare finance
01:09:42
engineering different verticals will win
01:09:44
the day or a a global model will win the
01:09:47
day that does everything.
01:09:49
>> Well you need general purpose models to
01:09:51
do the orchestration parts etc. But at
01:09:53
some point you enterprises sits on a lot
01:09:56
of intellectual property on a lot of
01:09:58
signals coming from physical systems
01:10:00
from factories from tools and the it's
01:10:03
actually not trivial to connect those
01:10:05
systems to connect those data to models
01:10:07
that are closed source. If you have open
01:10:09
models you can actually add uh new
01:10:10
parameters you can make a lot of deeper
01:10:13
things that you cannot do with closed
01:10:14
models. You can also and that's what
01:10:16
something that we do. We don't we not
01:10:18
only do we work at the model side but
01:10:19
also at the orchestration side. We see
01:10:21
it with subject matter experts to
01:10:23
understand their needs and we build
01:10:24
business applications that are fully
01:10:26
bespoke to their needs by modifying the
01:10:28
models but also modifying the harness on
01:10:30
top etc. So we believe that eventually
01:10:33
building on open source technology is a
01:10:35
way to save cost is a way to have better
01:10:36
control because you can sit the thing on
01:10:38
every cloud that you want on your
01:10:40
hardware if you want you can deploy it
01:10:41
on the edge if you want and eventually
01:10:43
uh from a from a customization
01:10:45
perspective and from leveraging your
01:10:47
decades of IP that you've been acrewing
01:10:49
in financial services in heavy
01:10:51
manufacturing like companies like SML
01:10:53
for instance they do benefit from
01:10:54
working with us because we take their
01:10:56
data and we build models that are
01:10:57
specifically good for their
01:10:59
>> um just training data using experts to
01:11:03
come in and refine a model. Most people
01:11:05
don't know this business that well, but
01:11:07
this has become a very large part of the
01:11:10
industry. Obviously, scale AI was doing
01:11:12
it. They went to Facebook, lost a lot of
01:11:14
the customer base who didn't want to uh
01:11:17
send their data, I guess, over to Meta.
01:11:19
Uh we're investors in a company called
01:11:20
Micro One that's doing pretty well in
01:11:22
this space. There's other folks doing
01:11:23
it. explain to the audience what you're
01:11:27
doing specifically for companies and how
01:11:29
this training works in a verticalized
01:11:31
way and then how you silo that data
01:11:33
because if you're working with one
01:11:34
customer in aerospace or fintech they
01:11:37
might have a need set but they may not
01:11:39
want that training to go to a
01:11:42
competitor. I can use a few examples. I
01:11:45
think overall the data segregation is
01:11:47
super important and the way we have
01:11:48
solved that is through a portable
01:11:50
platform. So our technology is a set of
01:11:52
services, a set of training tools, a set
01:11:55
of data processing tools that I can take
01:11:57
and that I can put on the infrastructure
01:11:59
of my customers. So suddenly from an IT
01:12:01
perspective and when we talk to the
01:12:02
CIOS, they realize that from security
01:12:04
perspective, the flow of data doesn't go
01:12:07
there's no data flow coming back to
01:12:08
Mistral because everything stays there.
01:12:10
Now uh the way we we then use that
01:12:13
technology that has been deployed is
01:12:15
that we're going to be working with uh
01:12:17
the team that is doing uh image scanning
01:12:20
and default detection with ISML for
01:12:22
instance and we're going to be sending
01:12:24
forward deployment engineers scientists
01:12:26
they're all PhDs they know how to train
01:12:27
models and they spend some time with the
01:12:29
subject matter experts that can explain
01:12:31
how an image is being detected what how
01:12:33
do you def detect defaults etc and based
01:12:36
on that we're going to work out what
01:12:37
kind of data needs to be used to train
01:12:40
the models that it's going to solve the
01:12:41
task in itself. And so the we we send
01:12:44
the technology typically we send a
01:12:46
little bit of scientists because uh you
01:12:49
do need that expertise transfer and that
01:12:51
knowledge transfer in between our teams
01:12:53
and the vertical experts and then we
01:12:55
make sure that eventually our team no
01:12:57
longer needs to be there to retrain the
01:12:58
models to get more data access etc. So
01:13:00
that combination of data segregation,
01:13:03
expertise transfer, knowledge transfer
01:13:05
is the one thing that makes us quite
01:13:06
unique and allows us to serve the most
01:13:09
critical use cases, the most critical
01:13:10
processes in industries that actually
01:13:12
need to take their data and put it into
01:13:14
models for it to work. Yeah, this seems
01:13:17
to be once the entire open web, what was
01:13:21
available legally, gray market, etc. I
01:13:25
wouldn't have you comment on that
01:13:26
controversy. Uh but we we kind of
01:13:29
exhausted what's in the open crawl.
01:13:31
Yeah,
01:13:31
>> we have.
01:13:31
>> And and it's time to actually
01:13:34
either make synthetic data or actually
01:13:37
use experts. Do you believe in synthetic
01:13:39
data and where does that work and where
01:13:41
does it fail? We use synthetic data as a
01:13:44
way to warm up the models. It's a way to
01:13:46
actually be quite efficient at the
01:13:48
beginning. If you have a large model and
01:13:49
you want to train a small model, you
01:13:50
would you will use your large model to
01:13:52
pro to process and to produce a lot of
01:13:55
synthetic data at the beginning. uh and
01:13:57
then but eventually you do need to have
01:13:58
human signal. Uh so the human signal is
01:14:01
something that is always a bit costly to
01:14:02
acquire because you need to talk to the
01:14:04
experts they need to give feedback to
01:14:06
the machines and so at the beginning
01:14:08
synthetic data allows you to do the
01:14:10
compression to to further compress the
01:14:11
models. At the end you do need to go and
01:14:13
get data that is uh produced by humans.
01:14:16
So yeah, it's a it's a way to have uh
01:14:18
it's it's mostly an efficient way of
01:14:20
training models to have big bigger
01:14:22
models that are used as as teachers for
01:14:25
smaller models, but it's not enough. And
01:14:26
so you also need human signal. Arthur,
01:14:28
we've seen um an incredible explosion.
01:14:31
We're sitting here on AO52
01:14:34
after OpenClaw, the year of our Lord, 52
01:14:37
days.
01:14:39
when you first saw Open Claw and saw the
01:14:41
reaction of hackers,
01:14:45
founders, startups, CEOs, just the
01:14:47
amount of energy and it racing to the
01:14:49
top of GitHub with the most number of
01:14:51
stars and likes and and all these
01:14:54
contributors. What did that say to you
01:14:56
as an executive in the space who's been
01:14:59
grinding on this for many years? What
01:15:01
what does that openclaw moment mean?
01:15:04
Well, it resonated a lot with what we
01:15:05
were doing with our customers uh because
01:15:08
pretty quickly uh enterprises realized
01:15:10
that if they wanted to make some gains
01:15:12
with artificial intelligence geni, they
01:15:14
would need to automate full processes.
01:15:16
And to automate a full process as an
01:15:18
enterprise, well, you can use open
01:15:19
cloud, but it's going to be uh it's
01:15:21
actually not really enough because you
01:15:22
you have data problems, you have
01:15:24
governance problems, you can't observe
01:15:26
uh the process that is running and you
01:15:28
can't can't control it in um in many
01:15:30
cases when you run a KYC process. So if
01:15:32
you're HSBC for instance, one of our
01:15:34
customers, uh you will want to have
01:15:36
deterministic gates that are going to
01:15:38
always do the same thing in a way that
01:15:40
is observable and that you can guarantee
01:15:43
the CIO that it's always going to go
01:15:44
through these gates and that's not
01:15:46
something that Openflow is providing
01:15:48
because it doesn't have this the kind of
01:15:50
primitives that you need to work on
01:15:52
collective productivity, observable
01:15:54
productivity and to work on mission
01:15:55
critical systems. On the other hand, uh
01:15:58
the autonomy it gives and the autonomy
01:16:00
it brings to to people that are just
01:16:02
individuals that are hacking together
01:16:04
things is a way to also show to
01:16:05
enterprises that if you set up the right
01:16:07
control plane, if you set up the right
01:16:09
sandboxes, if you connect to the right
01:16:11
data sources, if you make sure that you
01:16:13
your access controls are well respected,
01:16:15
then you can actually unleash the power
01:16:17
of agents doing things for your
01:16:19
employees and that's going to work. Work
01:16:21
on the platform cuz otherwise you will
01:16:23
not be at ease when you're sleeping. It
01:16:24
is um definitely something you have to
01:16:26
be thoughtful about. When I installed
01:16:28
it, I gave it just for my agent root
01:16:31
access to my Google Docs and my G Suite,
01:16:35
my notion, my Zoom and uh my notion and
01:16:40
uh GCAL, everything. And then I
01:16:42
realized, wow, I can with my enterprise
01:16:45
edition of Gmail essentially, I can just
01:16:48
summarize for my entire 21 person
01:16:50
investment company every conversation
01:16:52
going on in Gmail and then correlate it
01:16:55
with every conversation in Slack. And
01:16:57
then I realized, oh my gosh, there's
01:16:59
compensation discussions going on.
01:17:01
There's a person on a PIP who we put
01:17:03
them on a perform performance
01:17:05
improvement plan perhaps or something
01:17:06
like that. Oh, I have to make sure
01:17:09
nobody else can access this because the
01:17:12
power comes from giving it access to
01:17:15
data. But with great power comes great
01:17:17
responsibility and I think people are
01:17:19
learning that in real time. Yeah, it's a
01:17:21
big problem because the enterprise data
01:17:23
is not a single thing that you want to
01:17:24
put into a single system that is going
01:17:26
to be accessible by by everyone and so
01:17:28
you need to have this layer that
01:17:30
actually understands what is the is what
01:17:32
is in the data. you need to have a
01:17:34
semantic of what can actually be
01:17:35
proposed to uh HR or what can be
01:17:39
proposed to uh engineering and typically
01:17:42
compensation is one of these things you
01:17:43
want to make sure that the compensation
01:17:45
data does not flow back to all of the
01:17:47
all of the enterprise because you're
01:17:48
going to have a lot of problems uh if if
01:17:50
that's the case and so what you actually
01:17:52
need and which is hard to do is what we
01:17:54
call context engine so a mapping of
01:17:56
where the data sits that comes with a
01:17:58
certain number of metadata that is
01:18:00
telling you that this data is actually
01:18:01
not accessible to part of the company
01:18:04
and if you actually have someone in
01:18:06
engineering that is asking for something
01:18:08
related to comp the thing is actually
01:18:10
going to tell you look you actually
01:18:11
can't access that data so so that's uh
01:18:14
that's hard it's actually hard you need
01:18:16
to rethink entirely the way your IT
01:18:17
systems are being connected and uh at
01:18:20
some point you also need to think about
01:18:21
your management because your influ your
01:18:23
information flow is completely different
01:18:25
today uh if you're connecting agents
01:18:28
together with your data sources than it
01:18:29
used to be and suddenly maybe you don't
01:18:31
need that manager whose only purpose was
01:18:33
to take information from the bottom and
01:18:35
put the information on top etc. So
01:18:37
there's some IT problems to solve and
01:18:39
you need the right primitives, you need
01:18:41
sandboxes, you need airback based access
01:18:44
control and these kind of things and uh
01:18:46
you have change to do. You you need to
01:18:48
rethink your entire customer service uh
01:18:50
department cuz suddenly you actually
01:18:52
don't need that much transfer of
01:18:53
information operated by humans.
01:18:55
>> All right. Uh you have to go. You got a
01:18:57
flight to catch. It is so great to see
01:18:58
you Arthur. Continued success with
01:19:00
Mishril.
01:19:01
>> Thank you very much. Cheers. I'm really
01:19:02
lucky to have Daniel Roberts here. He's
01:19:04
the co-CEO and co-founder along with his
01:19:07
brother of Iron. They are a publicly
01:19:10
traded company. They started in BTC.
01:19:12
Welcome to the All-In Interview program.
01:19:14
>> Thanks, Jason. Pleasure to be here.
01:19:16
>> Yeah. And so you started in Sydney. You
01:19:19
and your brother um was seven, eight
01:19:22
years ago. And you got in early on
01:19:25
Bitcoin and all these Bitcoin monitor uh
01:19:28
miners wanted to have data centers. Huh.
01:19:30
>> Yeah, that that's directionally right.
01:19:32
So the thesis we saw was this explosion
01:19:35
of the digital world, the growth in the
01:19:36
online and at some point the real world
01:19:39
was going to struggle. So we set about
01:19:41
to build out largecale data centers.
01:19:43
Yes, the first use case was Bitcoin
01:19:45
mining. But as we said to our seed
01:19:47
investors, use that to bootstrap the
01:19:49
platform, generate cash flow, layer in
01:19:52
higher and better use cases over time as
01:19:53
they emerge. Here we are today with AI,
01:19:56
we are swapping out all the Bitcoin for
01:19:57
AI chips. When did you first start
01:19:59
seeing the demand in the company shift
01:20:02
from hey Bitcoin miners we need some
01:20:06
H100s whatever it is uh to hey we're
01:20:10
this nonprofit open AAI hey we're this
01:20:13
research lab we need some AI compute
01:20:16
when did that start hitting
01:20:17
>> look we had a bit of a false dawn I
01:20:19
would say back in 2020 we signed anou
01:20:21
with Dell to start bringing out
01:20:22
customers and compute but in hindsight
01:20:25
it was too early so we went back to
01:20:26
Bitcoin kept bootstrapping in the
01:20:28
platform. Look, I would say about 2
01:20:30
years ago and month by month, the demand
01:20:33
just continues to escalate.
01:20:34
>> And you were in so early that when you
01:20:37
were looking at data center space in the
01:20:39
United States,
01:20:42
you were one of one looking at the
01:20:43
space, one of two or three people
01:20:45
looking at the space, they they were
01:20:47
trying to sell you on space. Yeah.
01:20:49
>> Yeah. So, we actually develop the data
01:20:51
centers ourselves. So, we go and find
01:20:53
the land, we go and get the permits, we
01:20:54
go and apply for grid connections. And
01:20:57
we were doing it at a scale that just
01:20:59
amazed people at the time. Like 750
01:21:01
megawatts is our flagship Texas site
01:21:04
four years ago was unheard of. In the
01:21:06
middle of the desert, we're building
01:21:07
these big data centers. The traditional
01:21:08
data center industry going what are you
01:21:10
guys doing? We're saying we believe in
01:21:12
the future digitization, high
01:21:14
performance computing and obviously now
01:21:16
today it's paying dividends.
01:21:17
>> Yeah. I don't think anybody could have
01:21:19
predicted when chat GPT came out, Open
01:21:22
Claw recently as a turning point. Um,
01:21:25
and then you know, Microsoft, Google,
01:21:28
and everybody embracing this. Uh, and
01:21:30
that's your big partner, Microsoft.
01:21:33
>> Yes, Microsoft's one of our early
01:21:34
partners. We signed a $9.7 billion
01:21:37
contract with them late last year, but
01:21:39
as I was explaining to you before the
01:21:41
show, that's 5% of our capacity. So,
01:21:44
things are busy at the moment.
01:21:45
>> Yeah.
01:21:46
>> And when you do these buildouts,
01:21:49
the big conversation today is not is no
01:21:52
longer the number of GPUs putting in.
01:21:54
It's just power. Power is the uh
01:21:58
constraint today. Yeah.
01:21:59
>> Look, for many of the industry it is,
01:22:01
but for us, because we started 8 years
01:22:03
ago tying up all this land and power,
01:22:05
it's not. So, we've got 4 1/2 gawatt.
01:22:08
For context, that's almost as much power
01:22:11
annually as the Bay Area uses in its
01:22:13
entirety each. Wow. It's huge. So, for
01:22:16
us, the hurdle or the constraint is
01:22:18
really time to compute. And that's
01:22:21
emerging across the industry as well.
01:22:23
And time to compute means trades people
01:22:27
coming to West Texas living in a a
01:22:30
trailer that you set up to then break
01:22:34
ground on a data center, build
01:22:36
foundations, build water cooling
01:22:38
systems. Like this is hard manual labor
01:22:41
going on. Yeah,
01:22:42
>> exactly. And this is the whole real
01:22:44
world challenge to respond to these
01:22:46
digital exponential demand curves that
01:22:48
are unconstrained by the real world in
01:22:50
terms of their appetite. And it just
01:22:52
compounds. You need thousands of people
01:22:54
out in these locations that haven't
01:22:56
supported it. You put stress on supply
01:22:58
chains. We're seeing what's happening
01:22:59
with the memory, every aspect of it. So,
01:23:02
it's just permanent whack-a-ole,
01:23:03
permanent solving fires to try and be
01:23:05
bring online this compute.
01:23:07
>> And you get to spend time there.
01:23:10
>> What's it like when you set up a town or
01:23:13
you bring a thousand people or 2,000
01:23:16
people to what's a pretty much remote
01:23:19
small town? you I'm assuming that like
01:23:21
when you bring a thousand there might
01:23:22
only be 500 living there right now. So
01:23:26
what are those towns like? I'm it sounds
01:23:28
to me like something out of like the
01:23:30
gold mining era when people first you
01:23:32
know uh went and and were prospectors
01:23:35
prospecting town
01:23:37
>> pretty pretty much. I mean the
01:23:38
barbecue's great that was a draw card
01:23:40
but apart from that uh look we've always
01:23:42
had a policy of hiring local supporting
01:23:44
the local community. Uh this year we're
01:23:46
hitting a million dollars in community
01:23:48
grants cumulatively. That's things like
01:23:50
local playgrounds, supporting the fire
01:23:52
departments, but we will hire locally.
01:23:54
Once we can't find that trade locally,
01:23:56
we will expand the radius by 20 mi and
01:23:58
hire out of that and so on and so on for
01:24:00
us.
01:24:00
>> That's very thoughtful. Yeah. And and
01:24:02
these folks are coming say an
01:24:04
electrician or a construction worker.
01:24:07
They're coming having built houses or
01:24:10
you know uh maybe building um corporate
01:24:15
offices and now they come for a tour of
01:24:18
duty here and the salaries go up
01:24:20
massively but they got to leave their
01:24:22
family for a 3-month tour or something.
01:24:24
>> Yeah. Yes and no. Because typically
01:24:26
where we locate is where there's heavy
01:24:28
electrical infrastructure. Where there's
01:24:31
heavy electrical infrastructure is
01:24:32
typically where old manufacturing and
01:24:35
industry has closed down. Ah,
01:24:37
>> so we go in, leverage that sunk capex,
01:24:40
rehire, retrain local workforces and
01:24:43
bring a new industry to town in these
01:24:44
data centers. H
01:24:45
>> has has that workforce now been
01:24:48
completely depleted and we need to train
01:24:50
another generation, a younger generation
01:24:53
to be generation tool belt and really
01:24:55
embrace the trades
01:24:57
>> 100%. We're partnering with
01:24:58
universities, trade colleges.
01:25:00
Absolutely. And you go to a trade
01:25:02
school, you got you go to a college,
01:25:05
people are getting degrees in philosophy
01:25:07
and English literature, they're going
01:25:10
50k a year in debt, 200k a year in debt.
01:25:13
What's the starting salary for a trades
01:25:15
person working on a data center doing
01:25:18
electrical or construction or HVAC?
01:25:20
What's the ballpark range?
01:25:22
>> Uh, look, I won't talk specifics, but
01:25:24
they they are going up. The price is
01:25:26
going up. Depends on the level, but yes,
01:25:28
there is a rush for good. hearing 150 to
01:25:31
like 300K. Am I in the ballpark?
01:25:34
>> The lower end directionally, you're
01:25:35
right. Yeah.
01:25:35
>> Yeah. I mean, it's incredible when you
01:25:37
think about it. There's concern about,
01:25:40
hey, AI taking jobs and then on this
01:25:42
other side of the ledger can't find
01:25:45
enough talent to to to service it. Talk
01:25:48
to me about energy sources and how you
01:25:51
think about that. Uh, President Trump,
01:25:53
Chris Wright, the administration that
01:25:56
kind of started with, hey, clean,
01:25:57
beautiful coal. Year two, they're like,
01:26:00
"All sources matter." Nuclear,
01:26:02
obviously, nack gas is plentiful in that
01:26:05
area. We obviously got a lot of oil.
01:26:06
People don't know this about Texas in
01:26:08
the United States, the number one uh
01:26:11
source of solar installations. Yeah.
01:26:14
>> Yeah. Talk to us about energy.
01:26:15
>> So, so our our philosophy has been
01:26:17
sustainability from day one. We have
01:26:18
used 100% renewable energy since
01:26:21
inception.
01:26:22
>> What?
01:26:22
>> 100%.
01:26:23
>> Wait, how is that possible? It's
01:26:25
>> We use hydro in British Columbia. We use
01:26:27
wind and solar in West Texas. In West
01:26:29
Texas, where we're located, there's
01:26:31
around 45 to 50 GW of wind and solar.
01:26:34
>> Yeah.
01:26:34
>> The transmission line to export that
01:26:36
down to the load centers in Dallas and
01:26:38
Houston is 12 GW.
01:26:40
>> Oh.
01:26:40
>> So you go and locate to the source of
01:26:42
lowcost excess renewable energy,
01:26:44
monetize it into this digital commodity,
01:26:46
export it at the speed of light as
01:26:48
token.
01:26:49
>> Great arbitrage. And the wind is
01:26:51
producing a lot, but it it's harder to
01:26:53
get from those areas where people are
01:26:56
willing to put up. I mean, people don't
01:26:57
understand how big West Texas is. It is
01:27:01
an incredible amount of land. And you're
01:27:02
coming from Australia
01:27:04
where also on the west side, people
01:27:06
don't understand exactly how much just
01:27:09
pure nature land there is. Yeah.
01:27:10
Undeveloped.
01:27:11
>> So much land. And the issue is distance.
01:27:13
You've got to spend billions of dollars
01:27:14
on this transmission connection
01:27:16
infrastructure to move that power to
01:27:18
where people actually want it. You can
01:27:19
build wind farms, you can build solar
01:27:21
farms, but if you build it in the desert
01:27:23
and no one can use it, then what's the
01:27:25
point? So the whole opportunity for our
01:27:26
industry is to go to the source of that
01:27:28
power and monetize it.
01:27:30
>> So the data centers follow the wind
01:27:33
turbines, the solar installations.
01:27:36
How do you think about batteries and are
01:27:38
you able to put those online? Because
01:27:40
obviously you're going to have periods
01:27:41
where, hey, it's not a windy day. In
01:27:43
Texas, we have very few days when it's
01:27:45
overcast, so that problem's pretty much
01:27:47
solved. But you're going to have 50 days
01:27:48
where the sun's not beating down. So,
01:27:51
how do you deal with the demand and and
01:27:53
and softening that duck curve?
01:27:55
>> We don't need to.
01:27:56
>> The utility does that on our behalf. So,
01:27:58
this is why these grid connections are
01:28:00
so scarce, so hard to get, and so highly
01:28:02
valued because once you get that grid
01:28:04
connection, the utility underwrites all
01:28:06
of that variability. They guarantee you
01:28:09
24/7 reliable power.
01:28:11
>> Got it. So, on their side, they're
01:28:13
figuring it out. something goes down and
01:28:15
they could fall back even though you're
01:28:18
100% committed to renewables if they
01:28:20
needed to fall back to gas or whatever
01:28:22
they have that ability out there. So you
01:28:24
have that as a backup.
01:28:26
>> A lot of talk about or a debate. Are we
01:28:29
getting ahead of our skis? Are people
01:28:32
slowing down? There was some talk about
01:28:34
the OpenAI project maybe downscaling a
01:28:36
little bit. Is OpenAI a partner as well
01:28:38
or
01:28:39
>> uh can't comment.
01:28:40
>> Can't comment. Okay. So we'll we'll read
01:28:42
into that whatever we want.
01:28:44
But
01:28:46
are there pockets where people are
01:28:47
saying, "Hey, let's slow down." Or is it
01:28:49
still gang busters?
01:28:51
>> It's right up the end of the spectrum.
01:28:53
It's gang busters. We we cannot meet
01:28:55
demand. That's why the whole industry
01:28:57
now is around time to compute. There are
01:28:59
no idle GPUs in the world sitting in a
01:29:01
data center.
01:29:02
>> Yeah. And what's your take on when
01:29:06
software makes and this is a big uh
01:29:10
discussion from Jensen himself during
01:29:12
his two and a half hour keynote
01:29:13
yesterday. Uh we're sitting here
01:29:15
Wednesday. I think he did his keynote on
01:29:16
Tuesday. He was talking about hey
01:29:18
software is going to make it 50 times
01:29:19
more uh you know lower the cost of
01:29:21
tokens 50x and then you have um
01:29:25
transport also contributing to that.
01:29:28
When do you think the curve goes from
01:29:30
parabolic to simply growing at a
01:29:33
ridiculous level? Is is there a slowdown
01:29:36
coming or how are you planning for the
01:29:38
future?
01:29:38
>> Look, I think it's actually the
01:29:39
opposite. I think it feeds on itself.
01:29:41
So, I'll give you one example. You go
01:29:43
into chat GPT today and you generate an
01:29:45
image. You enter to the prompt. It's
01:29:47
like the dialup internet days.
01:29:48
>> It is
01:29:49
>> right. It takes minutes and you're like,
01:29:50
I better get this prompt right.
01:29:52
>> Yeah.
01:29:52
>> Finally, 2 minutes later, it comes. Now,
01:29:54
I'll give you an example. If we 10x the
01:29:56
amount of compute available, which is an
01:29:58
enormous task from where we are today,
01:30:00
and those images take 5 to 10 seconds,
01:30:02
are we going to generate more or less
01:30:04
images?
01:30:05
>> Oh, many more. Uh, this is Jevans
01:30:07
paradox. This is the theory of induced
01:30:09
traffic. You know, you build a couple
01:30:11
more lanes, people start to think, well,
01:30:13
maybe the uh distance from Bondai Beach
01:30:16
to to the central business district in
01:30:18
Sydney terms would be an acceptable
01:30:20
commute.
01:30:20
>> Love the analogy.
01:30:21
>> Yeah. Uh, so what do you think about or
01:30:25
or what are you seeing? I mean, we're
01:30:26
here at Nvidia. Obviously, they make the
01:30:29
leading edge chips. They just bought
01:30:30
Grock, so now you've got, you know, two
01:30:33
of the leading edge chips uh coming out
01:30:35
of the same company, but custom silicon
01:30:38
becoming a big discussion. Has that
01:30:40
started to land in the data centers yet?
01:30:42
Obviously, Google, don't know if they're
01:30:44
a customer, you can tell us, but they're
01:30:46
making custom silicon. Amazon is making
01:30:48
custom silicon. Meta is making custom
01:30:51
silicon. Talk to me about that
01:30:53
revolution and is it actually making it
01:30:55
to the data centers yet?
01:30:56
>> Look, it to various degrees it is.
01:30:58
They're promoting their products.
01:30:59
They're trying to tie up data center
01:31:01
capacity. So yes, there's multiple
01:31:03
silicon looking for homes. I think I
01:31:05
think it's fair to say Nvidia has a
01:31:07
massive head start. The ecosystem
01:31:09
they've incubated the standards that
01:31:11
they're setting. So I would say the
01:31:13
safest pathway to build out at scale
01:31:14
early is to follow the Nvidia road maps.
01:31:17
But absolutely over time we are seeing
01:31:19
these chips emerge.
01:31:20
>> A and in terms of desktop computing I
01:31:25
don't know if you saw the announcement
01:31:27
that um Dell and Nvidia are making a
01:31:31
really powerful desktop 750 gigs of RAM
01:31:34
lot of power. You're going to be able to
01:31:35
run some local models open source and
01:31:39
with openclaw and open source coming
01:31:42
from uh Kimmy and a bunch of the models
01:31:45
out in China.
01:31:47
has the hacker group, which I think you
01:31:49
started in like I did probably in
01:31:51
similar time periods. People are
01:31:53
starting to get really obsessed with
01:31:55
having a 10 or $20,000 desktop setup and
01:31:58
running this local. What do you think of
01:31:59
that trend? I'm curious.
01:32:00
>> Yeah, I mean the breakthroughs we're
01:32:01
seeing in software, the way it's
01:32:03
distributing power to every man in every
01:32:06
and woman in every house and their
01:32:07
ability to code and use products like
01:32:09
open core, the generation of demand and
01:32:12
appetite for compute at a local level
01:32:14
all the way through to these mega data
01:32:15
centers. It's absolutely real and as we
01:32:18
see the emergence of agents using more
01:32:20
and more as we see autonomous vehicles
01:32:22
and other automation, robotics, it's
01:32:24
absolutely going to compound.
01:32:26
>> And what about nuclear? Uh the Trump
01:32:29
administration
01:32:31
really seemed to flip the switch on a
01:32:35
growing
01:32:36
uh belief that hey wait, nuclear is
01:32:39
pretty great. It's clean. It's the
01:32:41
original renewable in a way. Uh, and
01:32:43
these new modular reactors have nothing
01:32:45
to do with Chernobyl, Fukushima, or
01:32:48
Three-Mile Island. They're much safer.
01:32:50
They're a completely different
01:32:51
architecture. Have the Have those
01:32:54
started to land yet? And are you since
01:32:56
you followed correctly in in the great
01:32:58
state of Texas where I'm from, you
01:33:00
followed correctly that time, are you
01:33:02
following nuclear?
01:33:03
>> I I think you have to. I think the
01:33:05
reality is it's going to take a decade,
01:33:07
a bit longer by the time big projects
01:33:09
can come into commissioning, but now is
01:33:11
the time to start that conversation. and
01:33:13
put in place policies, mobilize capital,
01:33:15
and start that ball rolling.
01:33:17
>> Yeah. Have you do you have a data center
01:33:19
going up near nuclear?
01:33:21
>> No, not at the moment.
01:33:22
>> Not at the moment. But you're actively
01:33:24
tracking that activity cuz
01:33:26
>> Yeah, this seems uh pretty inevitable.
01:33:30
Yeah,
01:33:30
>> feels like it.
01:33:31
>> And if that happens, what impact does it
01:33:33
have on your industry? If if you could
01:33:35
because obviously it's happening in
01:33:37
China and people always put the Bitcoin
01:33:39
miners they were like the canary in the
01:33:41
coal mine near the hydro dams and near
01:33:44
the nuclear where there was excess
01:33:45
capacity. What impact do you think this
01:33:48
has if you could actually have small
01:33:50
modular reactors next to data centers?
01:33:52
>> Well, I I think it just opens up the
01:33:54
market and enhances the US's competitive
01:33:55
advantage in this space. Like AI is
01:33:57
inevitable, robotics is inevitable. The
01:34:00
reality is the correlation between human
01:34:02
progress and energy consumption is
01:34:05
really really high over a very long time
01:34:07
period. So if we can find a way to
01:34:09
unlock new generation, clean generation
01:34:11
as nuclear and locate that more at the
01:34:13
source and enable more compute on a
01:34:15
distributed basis, all those use cases
01:34:17
we just discussed become easier, more
01:34:19
fluid, faster and then you get that
01:34:22
positive flywheel around Jebans's
01:34:23
paradox and demand. Talk to me about the
01:34:26
architecture today of
01:34:29
Ethernet and data moving between data
01:34:33
centers within data centers. That
01:34:35
backbone is going through a paradigm
01:34:38
shift as well. Yeah.
01:34:39
>> Yeah. Yeah, it is. And Jensen coins
01:34:41
coins the term the data center is the
01:34:43
new computer.
01:34:44
>> So you need to step back and you say
01:34:46
right this big building is essentially
01:34:48
the old desktop PC we had under our desk
01:34:51
at home. You go right how does that
01:34:53
work? So all the cabling, the latency,
01:34:56
the number of hops between each GPU, how
01:34:58
they talk to each other, the fabric
01:34:59
around Infiniband, Ethernet, it's
01:35:02
absolutely critical because every
01:35:04
millisecond matters in terms of
01:35:05
performance of that cluster.
01:35:07
>> Yeah. And where do you think uh or or
01:35:11
what do you think of Elon's uh vision?
01:35:16
It's obviously a a longer term vision of
01:35:18
putting data centers in space and
01:35:20
there's a couple other people working on
01:35:21
it as well. Yeah, I mean it's very hard
01:35:24
to argue with Elon. He's been very right
01:35:27
on a number of things for a very long
01:35:28
time. I think sitting here today, it
01:35:30
feels exceptionally difficult given the
01:35:32
cost of moving things to space, the
01:35:34
challenges around radiation. There's a
01:35:37
huge amount of engineering challenges,
01:35:39
but that's never scared Elon before. So,
01:35:41
I'm not
01:35:41
>> qualified and he's he he's inevitably
01:35:44
right, but sometimes he's late. He might
01:35:48
be late to the party. He might be late
01:35:49
to the dinner party. you might show up
01:35:51
at dessert, but generally uh he nails
01:35:54
it. How much of an issue is getting the
01:35:56
data out of the data center to consumers
01:36:00
today? Is that not something people are
01:36:02
worried about when you're building
01:36:03
something out in West Texas, all that
01:36:06
data, fiber, all that's been taken care
01:36:09
of or does that become a gating issue at
01:36:11
some point? So this was one of the big
01:36:13
myths that we had to bust when we
01:36:15
started this business because everyone
01:36:16
said data centers must be located close
01:36:18
to population centers, metropolitan
01:36:20
areas. Latency is really important and
01:36:22
we say yeah that's right latency is
01:36:23
important but the reality is in the US
01:36:26
Texas especially there is fiber
01:36:28
everywhere underneath the ground lots
01:36:30
and lots and lots of it. And when you
01:36:32
look at latency from our site in the
01:36:34
middle of the desert in West Texas down
01:36:37
to Dallas, the big carrier hotel, six
01:36:39
millisecond roundtrip latency. What's
01:36:42
six milliseconds? There's a thousand
01:36:44
seconds milliseconds in a second. Yeah,
01:36:46
>> we're talking six.
01:36:47
>> It's it's adjacent.
01:36:49
>> Yeah, it's it's not even uh Yeah, it's
01:36:52
definitely not material. Uh listen,
01:36:54
continued success uh and uh you're
01:36:58
hiring
01:36:59
>> a lot of people.
01:37:00
>> Yeah. Yeah, I think we got 129 job
01:37:03
advertisements up at the moment.
01:37:04
>> All right, so everybody go to the Iran
01:37:06
website. Uh, and listen, company's doing
01:37:09
fantastic. Thanks for spending some time
01:37:11
with us here at Allin
01:37:13
GTC.
01:37:16
>> Thanks, Jason.
01:37:16
>> Appreciate it.
01:37:33
I'm going all in.
