00:00:00
Open AAI's CFO, Sarah Frier.
00:00:02
>> We got to get right to it. You have just
00:00:04
completed what I regard as the most
00:00:06
successful fundraising round in history.
00:00:10
>> We're going to raise actually north of
00:00:12
$120 billion. We think AI is the biggest
00:00:15
era that we've seen to date. We're just
00:00:17
starting to understand what it's going
00:00:19
to mean for global productivity and with
00:00:21
that, you know, hopefully more
00:00:22
affluence, better lives for everyone.
00:00:25
Luck is whatever the preparation meets
00:00:27
opportunity, but you got to grab it.
00:00:31
>> Longtime listener, first time caller.
00:00:34
Quite exciting to get to hang out with
00:00:37
all the bros here. Hello.
00:00:39
>> We weren't sure how to start this off,
00:00:41
but I thought the best thing was to
00:00:42
allow our URSTW cryptosar to maybe
00:00:46
>> save comments. I saw an article today, I
00:00:48
think it might have been in the Wall
00:00:49
Street Journal, that the perception is
00:00:52
that there's an advantage to IPOing
00:00:55
earlier if you're an AI company. So now
00:00:58
we know SpaceX is going and then the
00:01:00
question is when's when are open AI and
00:01:03
Anthropic going to go? And I'm curious,
00:01:06
how do you think about that? Do you
00:01:08
think there is a little bit of a race on
00:01:10
or you know, you haven't made a decision
00:01:11
about that yet? Like in the end, an IPO,
00:01:14
I say this to the team all the time,
00:01:16
it's a milestone. It is not a
00:01:18
destination. Do not run your company as
00:01:21
if that's some sort of destination. It's
00:01:23
just another way to fund raise. We just
00:01:25
did, you heard me on on the the s the
00:01:27
sizzle reel, raise 122 billion dollars
00:01:30
in March, and that was to give ourselves
00:01:33
maximum flexibility. I feel like my job
00:01:36
as a CFO is create optionality for this
00:01:41
not just this company but just this era
00:01:44
that we're living in.
00:01:44
>> So Sarah was that that point in
00:01:47
fundraising is that the biggest private
00:01:49
republic up until the SpaceX IPO?
00:01:51
>> It is.
00:01:52
>> It is by orders of magnitude. I think
00:01:54
the largest IPO to date was the Sidio
00:01:56
Ramco which was about $30 billion. So it
00:01:59
is actually incredible that you're going
00:02:00
to have potentially three IPOs at a
00:02:03
scale that will be bigger even than 2001
00:02:07
2000 that that time frame there was a
00:02:09
lot that went on in the market too but
00:02:11
the market has grown and by the way the
00:02:13
other thing going on in the market is
00:02:14
like if you look at buybacks M&A and so
00:02:17
on it's actually a lot of capital keeps
00:02:19
being returned back to shareholders and
00:02:21
cash so there is a lot of money sitting
00:02:23
on the sidelines but in the spirit of
00:02:26
like the question David, I think in the
00:02:30
end you want to you'll be measured,
00:02:32
right? It's the in the end the market is
00:02:34
a weighing machine, not a popularity
00:02:35
machine. No one remembers who went
00:02:37
first, Google or Yahoo, Lyft or Uber.
00:02:40
And I say that not because whether I
00:02:41
want to be first or second, but I just
00:02:43
think it, you know, the the press loves
00:02:45
a bit of drama, but in the end, we're
00:02:47
going to have to build big, sustainable,
00:02:49
durable companies, and fundraising will
00:02:51
be a key component of doing exactly
00:02:53
that.
00:02:53
>> Sarah, breaking news.
00:02:55
>> Oh my god, so many people coming at me.
00:02:57
Hi, Jason. No, it is it is it is hard
00:02:59
balancing four interviewers at the same
00:03:01
time.
00:03:01
>> It's okay. This is my world, by the way.
00:03:04
So, I'm good with this. Jason,
00:03:05
>> Anthropic just uh confidentially filed
00:03:09
their S1. So, does that mean you're
00:03:11
third place in terms of the filing?
00:03:13
>> It does not mean anything yet because
00:03:16
you have to run now the gauntlet of the
00:03:18
SEC and who knows how long that takes
00:03:20
for anyone.
00:03:21
>> Yeah. Is it is there though a benefit to
00:03:24
them going forward? And I think
00:03:26
unpacking the rivalry with Anthropic is
00:03:29
on everybody's minds. So just I guess
00:03:32
you can't talk too much about IPOs. So
00:03:34
I'll just pivot to Anthropic was far
00:03:38
behind and now they've really um I think
00:03:41
everybody would agree in the industry
00:03:43
now blown past OpenAI in terms of
00:03:45
developers and corporations and it seems
00:03:48
revenue. So did how did that happen at
00:03:52
OpenAI when you had such a tremendous
00:03:53
lead? How did Anthropic blow past you
00:03:57
guys?
00:03:57
>> So let's talk a little bit about a
00:03:59
strategy. Our strategy is different,
00:04:01
right? So we are building the AI layer,
00:04:04
the infrastructure and it's really
00:04:06
important that there's a single
00:04:07
foundation but then with many interfaces
00:04:10
out into the world. So chat GPT is one
00:04:13
to the consumer. Over 900 million people
00:04:16
use Chat GPT weekly and it's become the
00:04:19
noun and the verb. It's how most people
00:04:21
experience um AI for the first time.
00:04:24
Kind of fun fact, our economic research
00:04:26
team just showed me um the fastest
00:04:28
growing continents now are Africa.
00:04:30
Probably not totally surprising since it
00:04:32
started a small base. fastest growing
00:04:34
languages are um Azerbjani and um what
00:04:39
Kazakhstani what is it's Kazak um which
00:04:42
it's kind of incredible to talk about
00:04:44
where it's going so multiple interfaces
00:04:46
chatgpt of course there's um codeex um
00:04:49
just hit 5 million over the weekend
00:04:51
we're really proud of that coming from
00:04:53
almost zero in January 5 million users
00:04:55
go codeex um help me prepare for this
00:04:58
little special up here too um there's of
00:05:01
course Frontier our enterprise offering
00:05:03
ing and everything every other way that
00:05:05
we can get out there to reach businesses
00:05:07
of all sizes. That is a very different
00:05:09
strategy. We think that because it's
00:05:11
served up on one model, there's a
00:05:13
compounding element of advantage that
00:05:16
comes from that. More users, more data,
00:05:18
more ability to personalize chats as a
00:05:22
front door. As we as models get bigger,
00:05:25
there's more efficiency. That should
00:05:26
lower the overall cost to give you a
00:05:28
token in the world. That should compound
00:05:30
to higher gross margins, ultimately more
00:05:33
ways to pay for compute, and then access
00:05:35
to compute is one of the really big
00:05:37
competitive advantages at the moment.
00:05:39
So, you know, we have to all run our own
00:05:41
races, but we all have to recognize
00:05:43
we're part of an ecosystem that also
00:05:46
needs to bring people along
00:05:48
collectively.
00:05:49
>> Did you spread a little bit too then too
00:05:50
many projects? People were talking about
00:05:52
this new gadget, Sora, and and then
00:05:55
maybe not enough focus on enterprise. Is
00:05:57
that a fair assessment of if there was a
00:05:59
mistake in the last year that was it?
00:06:01
>> No, I I think that the world loves to go
00:06:03
to binaryisms like are you a consumer
00:06:06
company Sarah? Are you an enterprise
00:06:07
company? The reality is we're very much
00:06:10
both. We're not one or the other. Right
00:06:12
now our revenue is getting pretty
00:06:14
balanced about 50/50. We are incredibly
00:06:17
focused on the enterprise. Like I spend
00:06:19
so much of my time with I mean just even
00:06:22
in the last week I could tell you I've
00:06:23
been to see Thermopisher in Boston. I
00:06:25
was with a bunch of banks in New York. I
00:06:28
was on the phone with travelers on
00:06:30
Friday. I spent this morning on the
00:06:32
phone with a tech company. It doesn't
00:06:34
matter the vertical. People are really
00:06:36
moving on AI right now. Our new head of
00:06:38
revenue, Denise Dresser, in seats since
00:06:40
December. She is a force of nature. And
00:06:43
so I think the enterprise broadly
00:06:45
speaking is really firing on all
00:06:47
cylinders. But we don't want to leave
00:06:49
the consumer behind. Remember our
00:06:50
mission at OpenAI is AGI for the benefit
00:06:54
of humanity, not for the benefit of
00:06:56
humanity who can pay or for the benefit
00:06:58
of humanity who live in an enterprise,
00:07:00
but very broadbased. Um, it's why we
00:07:04
offer so much free because we want
00:07:06
people to get a taste. Once they get a
00:07:09
taste of intelligence, the ability to
00:07:11
come up a commitment curve is
00:07:12
incredible. Our free users do about
00:07:14
seven turns, seven questions a day. Our
00:07:17
first paid tier do double that about 15.
00:07:20
Our our real paid tier the plus 20
00:07:23
bucks. Hopefully you're all on it or
00:07:24
higher about 3x and pro about um 11x
00:07:30
over a free user. So remember when you
00:07:32
got your flip phone and you're like yeah
00:07:34
I don't know what it does make some
00:07:35
calls. Now that same phone think of all
00:07:38
the things it does for you. That's the
00:07:40
path we're on with intelligence right
00:07:41
now. Sorry.
00:07:42
>> You said something very influential. I
00:07:44
think it was about 18 months ago for a
00:07:46
lot of us in the industry where you
00:07:48
framed a very simple economic trade-off
00:07:50
which was gigawatts to cash and I think
00:07:53
you said one gigawatt is roughly
00:07:55
equivalent to about $10 billion a year
00:07:58
of revenue to open AAI. So common number
00:08:01
one was this one gigawatt equals10
00:08:03
billion a year of revenue for you. But
00:08:05
it's not just you because you can
00:08:07
probably extrapolate that to anthropic
00:08:09
and other folks Gemini
00:08:11
>> but then you were really at the
00:08:12
forefront of getting access to power and
00:08:14
data centers and powered land. It seemed
00:08:17
a little crazy but now it looks like
00:08:19
hold on there's a huge deficit of
00:08:21
supply. Can you just unpack all of that
00:08:23
and explain
00:08:25
>> both the spectrum of where we are and
00:08:27
then those specific economics and if
00:08:28
that's changed. So first of all, yes,
00:08:31
compute is a very scarce resource at the
00:08:34
moment. I mean, what we see in our
00:08:35
business, we're going up that kind of
00:08:37
vertical wall of demand right now and
00:08:39
there's just not enough tokens
00:08:40
available. So we I'm very grateful that
00:08:44
I got to work alongside Greg and Sam. I
00:08:47
think we're very precient on this and
00:08:49
last year we were definitely taking some
00:08:51
you know arrows in the back about why
00:08:53
are they out there buying all this
00:08:54
compute and I think thank god we did
00:08:56
because in 26 we still won't have enough
00:08:59
compute. Um where are we on the compute
00:09:02
continuum there's kind of choke points
00:09:04
everywhere and and I think they will
00:09:06
continue to move back and forth. I mean
00:09:07
you all talk about this and know this as
00:09:10
well as anyone um here whether it's
00:09:13
energy first and foremost um land power
00:09:16
we get regulatory um environment such
00:09:19
that we can build quickly um when you
00:09:22
get into the racks and chips themselves
00:09:24
clearly do we have enough um in that
00:09:26
supply chain memory spike is is on at
00:09:29
the moment access to great talent um do
00:09:32
we have enough people coming through our
00:09:33
education system I really worry about
00:09:35
this right now I'm a trustee at Stanford
00:09:38
and you know I see just that you know we
00:09:40
need to keep the focus on education and
00:09:43
science um and then trust I mean I
00:09:45
actually put that as part of the supply
00:09:47
chain um Sam right now is in Selen
00:09:50
Michigan he's going to be cutting the
00:09:51
ribbon in about two hours so you are
00:09:53
getting a sneak preview but they told me
00:09:55
it was okay to say it in the room um
00:09:57
that will be you know sticking shovels
00:09:59
in the ground on a 1 gigawatt data
00:10:01
center which is part of our or Oracle
00:10:03
complex really important there on the
00:10:05
trust side that we don't leave
00:10:07
communities behind. I spent seven years
00:10:10
of my life working at Next Door doing
00:10:12
the hard work of what it means to be
00:10:14
local and you cannot tell people from
00:10:17
top down what they need because they
00:10:19
will tell you thank you but no thank
00:10:20
you. I will tell you what I need. And so
00:10:23
in a data center like that actually
00:10:25
spending a lot of time in the community
00:10:26
saying number one, we're not going to
00:10:28
raise your electricity bills. We're
00:10:30
going to pay for our infrastructure and
00:10:31
our power. It will not be the rateayer
00:10:33
that has to pay. Number two, we're going
00:10:35
to bring jobs, 2500 union jobs, good
00:10:39
jobs like electricians, HVAC, and so on.
00:10:43
We are going to pay our taxes, a billion
00:10:46
dollars in taxes just for that data
00:10:48
center into Michigan. And on top of
00:10:50
that, we're going to invest $45 million
00:10:53
going into education for codeex credits
00:10:56
to do what you all talked about this
00:10:57
weekend is like anyone who's not like
00:11:00
coming in filled to their new job. I
00:11:02
have teenagers using Codeex. It would be
00:11:04
like I would never hire a finance person
00:11:06
didn't know how to use Excel. And I
00:11:08
pretty much probably wouldn't hire a
00:11:09
finance person today that doesn't know
00:11:11
how to use a tool like Codeex. Um so
00:11:14
that you know so when I think about
00:11:16
investment we're having to invest ahead
00:11:18
of demand. That means we need to both be
00:11:20
able to find all of the compute and all
00:11:23
the pieces and then pay for it. So that
00:11:25
goes back to your capital question on
00:11:26
IPO. And then on the other side on the
00:11:29
economics, look, the economics do
00:11:30
continue to get better. They're getting
00:11:32
better on multiple fronts. I think we
00:11:34
are doing a better job of actually
00:11:36
showing true value to our customers. And
00:11:38
I think you get beyond kind of a cost
00:11:40
plus type pricing into something that
00:11:42
feels more akin to the value being
00:11:45
created. Now, scarcity of tokens helps
00:11:47
because it's causing a bit of a
00:11:49
compression. Can you talk about that in
00:11:50
just like without specific names where
00:11:54
you know the landscape exists today in
00:11:57
terms of all the power that's available
00:11:59
and all the demand that exists across
00:12:01
everybody. Yep.
00:12:03
>> What's going to happen over the next
00:12:04
year just at the current course and
00:12:06
speed of what is available of the data
00:12:08
centers that's available of the tokens
00:12:10
that's available of the infrastructure
00:12:11
that's available for everybody because
00:12:13
you know I told this story last week but
00:12:16
you know I'll use anthropic and one of
00:12:17
the frustrating things is at some point
00:12:18
it just says you know 10:30 it's like
00:12:21
all right Jimoth see you at 2:30.
00:12:22
>> Yeah.
00:12:22
>> And that's not a viable experience
00:12:25
>> right. Um
00:12:26
>> and in fairness to chat GPT actually
00:12:28
I've never had that with
00:12:29
>> Yeah. We we're quite generous with our
00:12:32
tokens and again on purpose we're trying
00:12:34
to drive access so people understand
00:12:37
because if you're on that free tier not
00:12:39
actually getting the latest model but
00:12:40
we're trying to put it in your hands so
00:12:42
you get a sense for it by the way
00:12:44
because you know if you're a kid um
00:12:47
doing homework like I think about when I
00:12:49
grew up and the encyclopedia bratannicas
00:12:52
showed up at the front door in Northern
00:12:54
Ireland in a tiny little community in
00:12:56
the middle of the troubles. It was like
00:12:57
the clouds parted and so we want to make
00:12:59
sure that people get that feeling by the
00:13:02
way but the landscape right now in 26 if
00:13:05
you want to buy more compute good luck
00:13:07
to you like tell me cuz I don't know
00:13:09
where else to find it. I mean as you
00:13:10
know
00:13:11
>> well I was going to say Elon ironically
00:13:13
ended up being the one person that had
00:13:16
too much compute in a way but good job
00:13:19
on like figuring out how to sell that
00:13:21
off. Um in 27 it's pretty limited as
00:13:25
well frankly. Now there's a couple of
00:13:27
things shifting around when we talk
00:13:28
about compute. There's training that
00:13:31
mostly still all happens here in the
00:13:34
United States for USG reasons for making
00:13:37
sure that a national asset and effect is
00:13:39
happening on US soil. For inference, we
00:13:42
want that to be global. And I think
00:13:44
particularly in an agentic world, you
00:13:46
want much more kind of real time. Even
00:13:48
for things like Sora and video, which by
00:13:50
the way, yeah, we have, you know, we had
00:13:52
to make a really tough choice because we
00:13:54
didn't have enough compute and we said a
00:13:56
lot
00:13:57
>> right now. Yeah, video does, but video
00:13:59
is not over. Like in particular, when
00:14:02
you start to think about where AI is
00:14:04
taking us into more multimodality. So
00:14:07
remember, we've all been taught by the
00:14:10
last generation of technology to talk
00:14:12
with our thumbs. It's a disease. You
00:14:15
walk around, everyone's looking down.
00:14:16
They don't look up anymore. Teenagers
00:14:18
sit on my sofa at night and talk to each
00:14:21
other with their thumbs. I'm like, who
00:14:22
are you talking to? And my son will be
00:14:24
like him. I'm like, okay, talk.
00:14:27
Multimodality is here. Um hopefully I
00:14:30
think you all talked about it this
00:14:32
weekend. You're talking to your tool. I
00:14:33
talk to codeex every day. And so that is
00:14:36
changing rapidly, but that is going to
00:14:38
need much more kind of real time compute
00:14:40
because it's an odd experience if I was
00:14:42
talking to
00:14:43
>> Johnny I of this puck these earpieces.
00:14:46
So maybe tell us a little bit about that
00:14:47
problem if you've admitted it now.
00:14:49
>> If I if I tell you it's an earpiece,
00:14:50
Johnny will come and steal my teenage
00:14:52
son. I might give it to him. give them
00:14:54
to
00:14:54
>> but you do believe that there should be
00:14:56
some
00:14:57
>> we're changing into a consumer substrate
00:15:00
that I cannot tell you what it is but by
00:15:02
the end of this year we will unveil it
00:15:04
early next year I have seen it I've
00:15:07
tried it I am a hand talker right now
00:15:09
I'm sitting on my paradigm
00:15:12
>> shift when when yeah when you used it
00:15:13
was it like having an iPhone for the
00:15:15
first time
00:15:16
>> it's very what Johnny and team are
00:15:19
really good at is bringing humanity to
00:15:23
devices. And I don't really know how to
00:15:25
explain that well, but when you see it,
00:15:27
you feel it.
00:15:28
>> It feels natural in some way.
00:15:29
>> It feels very natural, but it feels very
00:15:32
lovable
00:15:33
>> really.
00:15:34
>> And I can't really explain what that
00:15:36
emotion is cuz
00:15:37
>> intimate in some way in terms of
00:15:39
>> technology,
00:15:40
>> not taking your phone out and it's it's
00:15:42
seamless is what I've heard from people
00:15:43
that played with it.
00:15:44
>> Technology is very um can be very
00:15:46
mechanistic, but we all know great
00:15:49
design just makes everything fade away,
00:15:52
right? what um at the time you know the
00:15:54
simple is hard.
00:15:56
>> Yeah.
00:15:56
>> But I I think this is a very this story
00:15:59
just going back to the earlier question.
00:16:00
So putting on
00:16:01
>> the CFO hat, help us understand the
00:16:05
capital allocation model that you use
00:16:07
cuz a lot of businesses over the last
00:16:09
decade, two decades that have kind of
00:16:11
been these outsized returners have found
00:16:14
some unique way to deploy capital at a
00:16:16
higher ROC than anyone else. and then
00:16:18
you end up plowing all your capital into
00:16:20
that higher ROC bucket.
00:16:23
>> What is that for you guys? And how do
00:16:25
you think about that le that portfolio
00:16:27
approach to having more of these kind of
00:16:28
big returner shots and is there an
00:16:30
engine where that gets better over time?
00:16:32
>> There has to be because in the end the
00:16:35
durable highvalue companies created in
00:16:38
this era, I don't think they're not
00:16:40
going to be magical. They're going to
00:16:42
look like the great companies of prior
00:16:43
eras. They're going to create customer
00:16:45
value. starts with a customer um and
00:16:49
really helps the customer do something
00:16:51
different, better, more revenue, more
00:16:54
efficiency, right? Thermmaisher wants to
00:16:57
be able to get um patient screening done
00:17:00
faster so they get FDA approval faster.
00:17:03
That's really important. Like if you
00:17:04
have a form of cancer where you have
00:17:07
weeks to live, the difference between a
00:17:09
breakthrough in four weeks and two weeks
00:17:11
can literally be life or death. They
00:17:13
also have, I'm going to misquote this,
00:17:15
but something like 30,000 38,000 people
00:17:19
in the field selling those amazing like
00:17:21
if you walk into any lab in the country,
00:17:23
you'll just see thermopisher plastered
00:17:25
all over every device. Those people want
00:17:27
to be more efficient going to work. Like
00:17:29
the the fastest takeoff of codecs within
00:17:31
OpenAI right now is actually in our go
00:17:34
to market team. Our devs are there, but
00:17:36
like if you look at the pace of growth
00:17:38
kind of month over month, it's all in
00:17:39
GTM. So they want more productivity out
00:17:42
of their GTM team and of course um
00:17:44
they're doing things in areas like
00:17:46
finance which I get really excited
00:17:47
about. So customer value first from that
00:17:50
now you need to get to great gross
00:17:52
margin. So how do you get to great gross
00:17:54
margin? You're looking at like the cost
00:17:55
of revenue. The main input is compute.
00:17:58
The good news on compute is that there
00:18:00
is a massive deflationary curve on cost
00:18:02
right from chat GBT uh five to 54 I
00:18:08
think the deprecation of cost was
00:18:09
something like 97%. It's like kind of an
00:18:12
amazing curve. Actually, I'm slightly
00:18:14
from four to 54, it was 97%. But that
00:18:18
happened in like two years.
00:18:20
>> That's kind of wowing, right?
00:18:22
>> Um, even our newest model, if you look
00:18:24
at 55 that we just released, we're
00:18:26
trying to now translate that back to the
00:18:28
customer. So, we actually raise prices
00:18:30
on 552x.
00:18:32
But if you look at what the cost of the
00:18:33
customer is, they're probably still
00:18:35
getting a break of about 20 to 30% cost
00:18:38
reduction per token because just much
00:18:40
more efficient per token. So there's a
00:18:42
lot to do in that envelope and and part
00:18:45
of making an alloc a capital allocation
00:18:47
decision is having to
00:18:49
>> if you make it on today's cost profile,
00:18:51
>> you actually might mispric the outcome.
00:18:54
So you have to lean in a little on the
00:18:55
cost profile. And then as we think about
00:18:58
like the builds, yeah, you are having to
00:19:00
make like really my focus today on
00:19:03
compute is what's the compute I can buy
00:19:05
for 28 onwards. Like that Michigan data
00:19:09
center in Seline, I don't think we will
00:19:11
be getting compute out of it until
00:19:13
probably end of 27, early 28. So that's
00:19:16
where you're starting to make your bets.
00:19:18
And in fact, where I feel most short of
00:19:20
compute right now is starting to look at
00:19:22
30, 31, 32. So you're having to create a
00:19:26
business model. Um now the good news is
00:19:28
each year goes by we get more confidence
00:19:31
in the build. We're seeing it massively
00:19:33
outperform and so that's giving us more
00:19:36
and more confidence and the market is
00:19:38
coming towards us much more.
00:19:40
>> All right. So how are you making the
00:19:42
compute need forecast multiple years out
00:19:46
accounting for all of the architectural
00:19:49
and model advancements that are
00:19:50
happening where call it value or utility
00:19:53
per unit of power is going up and help
00:19:57
us understand how you kind of estimate
00:20:00
that given that there's a lot of
00:20:02
technology development going on that has
00:20:03
a high kind of variance to it.
00:20:06
>> Yeah. Yeah. So we we do have to make
00:20:08
multiple um assumptions both on the
00:20:11
compute itself. So we assume right now
00:20:13
that compute actually on a per gigawatt
00:20:15
is getting more expensive because power
00:20:17
is getting more expensive, memory is
00:20:18
getting more expensive and so on.
00:20:20
However, the the intelligence that we
00:20:23
get on the other side out because of the
00:20:26
deprecation on the chip side is is more
00:20:28
than making up for that. So in in terms
00:20:30
of a per unit sold to a customer, it
00:20:32
should actually get a lot less expensive
00:20:34
for
00:20:34
>> no model improvement in that.
00:20:36
That's just
00:20:37
>> exactly that's just the chip itself. We
00:20:39
don't try to overestimate on the model
00:20:41
side because sometimes like 55 is an
00:20:44
incredibly good model on the efficiency
00:20:46
side, but if you look at something like
00:20:48
54, the prior model, it was a really
00:20:50
large pre-trained model. It was very
00:20:52
expensive. It was actually hard to
00:20:54
serve. And sometimes we want to do that
00:20:56
really big pre-train moment and then we
00:20:59
take multiple model turns to be able to
00:21:01
kind of drive down on the cost side. I
00:21:03
mean in the in the near term like in 26
00:21:06
and 27 I clearly build a model that's
00:21:09
bottoms up. So I know what my products
00:21:12
are. I have a sense of what the pricing
00:21:14
will be. Um you know P time, you know,
00:21:16
consumer P* Q. How many Wows do I think
00:21:19
I have? I can see what the shape of the
00:21:20
line is. How many of them will subscribe
00:21:23
advertising coming in is also still
00:21:25
related to how many weekly activives,
00:21:28
how many dailies, how many messages and
00:21:30
so on. So you can you can do actually a
00:21:32
pretty good model job in 26 and 27. That
00:21:36
said, the shape of the line keeps taking
00:21:38
us by surprise to the upside. When you
00:21:40
get into the outer years, you're
00:21:42
actually looking more at the compute
00:21:43
you've bought and almost just doing an
00:21:46
algorithm the other way that's saying
00:21:47
this amount of compute should equate
00:21:49
somewhat to this amount of revenue. I
00:21:51
don't know for certain exactly where it
00:21:53
will all come from. Like a year ago, I
00:21:55
built a model for investors that showed
00:21:57
agentic revenue. And the story was,
00:22:00
we're going to have this thing. We're
00:22:02
going to be in the agent era. We're
00:22:03
going to hand it to a developer with
00:22:05
natural language. They're going to be
00:22:07
able to build and we think they will pay
00:22:10
upwards of maybe $2,000 a month for it,
00:22:14
which is kind of laughable in hindsight,
00:22:16
but nobody believed. Um, they were like,
00:22:18
I don't even know what she's talking
00:22:19
about. There's no way that will happen.
00:22:21
And $2,000 a month, remember when people
00:22:23
were losing their minds over Chat GBT
00:22:25
Pro being at $200? Oh my god, no one
00:22:29
will ever pay for that.
00:22:30
>> Yeah.
00:22:30
>> So why 122 billion? Does it take you to
00:22:33
2031, 2032? Like how do you get the
00:22:35
calculus on the capital needs as you do
00:22:37
that modeling?
00:22:38
>> You maybe even more specific. So the
00:22:41
estimates I've seen is that to stand up
00:22:43
1 gawatt of AI compute costs about $50
00:22:47
billion in cap land, PowerShell chips,
00:22:51
everything all in around 50 billion.
00:22:54
Do you have to front all of that money
00:22:56
when you create a new data center or how
00:22:59
much of it do you do? How much of it can
00:23:01
you get debt for? Does a 100red billion
00:23:04
raise only get you two gigawatts or does
00:23:06
it get you five? Like what does it get
00:23:08
you?
00:23:08
>> It's it's a great question. So if you
00:23:10
look at our compute strategy um and it's
00:23:13
crazy how fast the world has changed. So
00:23:15
just two years ago we were literally one
00:23:18
we had one CSP we worked with Microsoft
00:23:20
Azure um we we sat on one chip Nvidia we
00:23:24
had one product chat GBT one price point
00:23:27
$20 a month so I often use a Rubik's
00:23:29
cube as kind of my metaphor so we were
00:23:31
like one cube in the bottom today if you
00:23:34
look at our strategy it's been to go
00:23:36
first of all multi multiple CSPs because
00:23:39
what CSPs do for us in effect is they
00:23:41
shift capex into opex so you pay as you
00:23:44
get the revenue so as you're actually
00:23:46
utilizing the data centers. So in effect
00:23:48
we are writing somewhat on their ability
00:23:50
to build and have capex and um
00:23:53
financing. So today we sit on top of
00:23:55
every CSP Oracle um coreweave um
00:23:59
Microsoft GCP AWS and a bunch of small
00:24:02
neocalars on the chip side. We've also
00:24:05
um gone for a program of being
00:24:07
multi-chip um because we want to make
00:24:09
sure you're always on the frontier. I
00:24:11
think if you're only on one chip,
00:24:12
there's just inherently a moment where
00:24:14
you can't be on the frontier because
00:24:16
there's some leaprogging that happens.
00:24:18
So today, Nvidia remains our absolute
00:24:20
priority partner. They have the Frontier
00:24:22
chip. Our next big trading run in the
00:24:24
fall will be done on Ver Rubin. We're
00:24:26
really excited about that. And now we're
00:24:28
plotting kind of the Fineman series
00:24:30
that's coming. But we also now have
00:24:32
chips in the pipeline from AMD. Um
00:24:35
Cerebrris is already online. It's been
00:24:37
an incredible low latency chip. Great
00:24:39
for devs, for example, that want
00:24:41
real-time coding. And there's our own
00:24:43
chip that we're working on with
00:24:44
Broadcom. And then beyond that, there's
00:24:46
other ways we've diversified. So now,
00:24:47
think about that Rubik cube. It's become
00:24:49
much more multi-dimensional. And it
00:24:52
allows us to effectively utilize
00:24:55
investment grade CSPs in order to be
00:24:58
able to go fast and push it back to be
00:25:00
more opex, not capex. Now, we are
00:25:03
starting to shift gears into more of a
00:25:05
built to suit type environment. We
00:25:08
announced a data center we're building
00:25:10
with SoftBank Energy um down in Texas.
00:25:13
That's the beginning of something that's
00:25:14
beyond a CSP. There's a little bit more
00:25:17
capex required there. And then finally,
00:25:19
I think as the world progresses,
00:25:21
remember we've done all that just in two
00:25:22
years. The reason I like a Rubik's cube
00:25:24
is again, please chat GBT this, but I
00:25:27
think a Rubik's cube has something like
00:25:28
a quintilion different um forms it can
00:25:31
come up with. And so it just gives us a
00:25:33
lot of optionality. So remember what I
00:25:35
said, my job is maximum optionality and
00:25:39
in a moment where I'm not yet an
00:25:40
investment grade type of entity where I
00:25:43
can go get lowerc cost debt financing,
00:25:45
being able to work with partners to do
00:25:47
that is really important.
00:25:48
>> So do you think that in 5 years from now
00:25:50
the stack is just merged together? What
00:25:53
do I mean? In traditional or historical
00:25:56
markets, you'd have Nvidia sell the
00:25:57
chips, but that's all they do. And then
00:25:59
you'd have, you know, Microsoft just run
00:26:01
a cloud. That's all they would do. And
00:26:03
then you would have a consumer app.
00:26:05
That's all you would do. But now we see
00:26:07
everybody doing everything. You know,
00:26:08
you guys have silicon that you're
00:26:10
spinning. You have models that you make.
00:26:13
You may or may not eventually decide
00:26:14
that you need to be some form of a
00:26:16
NeoCloud yourself. If you look at
00:26:18
Nvidia, they have incredible silicon,
00:26:20
but they also have their own open source
00:26:21
models. They're increasingly becoming an
00:26:23
offtaker. Google is a cloud company
00:26:26
first, but they also have a chip. Now
00:26:27
they have models. So it's all merging.
00:26:32
is if that continues to happen, does
00:26:34
that make the competitive landscape
00:26:36
simpler or easier?
00:26:38
>> I mean, I think where everyone is trying
00:26:40
to make sure they reside is the layer
00:26:42
that is closest to the customer where
00:26:45
usually you take the the largest portion
00:26:48
of the profits of the ecosystem, right?
00:26:50
No one wants to find themselves
00:26:54
away. Absolutely. And so that's why
00:26:57
today when I think about our position
00:26:59
and comes back to where I started why we
00:27:01
want to be that AI intelligence layer is
00:27:05
because a year ago people talked about
00:27:07
the commoditization of the LLMs. Um and
00:27:10
frankly it's gone the opposite because
00:27:13
as you start building an agentic layer
00:27:15
and we've all started to use this word
00:27:16
harness but the harness is what brings
00:27:19
the context the memory right it I I have
00:27:23
in my codeex I have a whole ginormous
00:27:26
memory file where it knows that I'm me
00:27:29
it knows I'm the CFO of openai it knows
00:27:31
how I like to write things how I like to
00:27:33
say things it knows what I'm interested
00:27:35
in it actually also knows that I'm a mom
00:27:38
teenagers I mean it just carries all
00:27:40
this memory and that makes the model
00:27:43
more powerful for me. Now think about
00:27:45
what happens when that memory in that
00:27:46
context is brought into an actual
00:27:49
enterprise environment. So now it's not
00:27:51
just even about the data that resides
00:27:54
there, but I always think about the the
00:27:56
intuition of like back when I worked on
00:27:57
Wall Street, right? There was all the
00:27:59
data in the world that told you what a
00:28:01
stock should do post an earnings call.
00:28:04
But give me one second. Then you called
00:28:07
your trader and the trader be like,
00:28:09
"Yeah, stock's not going up, Sarah." And
00:28:11
I'm like, "What are you talking about?
00:28:12
Like all the numbers say it did this, it
00:28:14
did this, it did this." He's like,
00:28:14
"Yeah, I know, but I know this fund is
00:28:17
under pressure and they need to sell
00:28:19
down their book and that is going to
00:28:21
kill the stock for the next week."
00:28:23
Right? That is the intuition
00:28:25
of an enterprise. Like it's the best
00:28:27
example I always think of because I came
00:28:29
out of a financing world, but there's
00:28:30
this intuition in every walk of life.
00:28:34
And that's where I think the models are
00:28:35
now getting very connected to the memory
00:28:38
and context and intuition of your
00:28:40
company. And that's what gets CEOs and
00:28:43
seuite really excited because they're
00:28:45
like okay now I really see how this is
00:28:47
going to add value to drive my revenue
00:28:49
line my top line but also you know I can
00:28:52
think about as an efficiency play as
00:28:54
well. And so back to what you're asking,
00:28:56
I think what people want to make sure is
00:28:57
they stay as close to that value as
00:28:59
possible
00:29:00
>> and be flexible enough to pivot as you
00:29:03
have to rap.
00:29:04
>> Yeah.
00:29:04
>> But
00:29:04
>> sorry Jason,
00:29:05
>> it's quite all right. Um it's been
00:29:07
wonderful and and you've been so great
00:29:09
with the details. One final detail
00:29:11
question. Rapid fire. Three great
00:29:13
greatest consumer businesses of our
00:29:15
lifetime. iPhone, Meta Advertising
00:29:18
Network, and Google's advertising
00:29:20
network. Two of those three are ad based
00:29:21
and and even Apple has a sprinkling of
00:29:23
ads.
00:29:24
>> Haven't heard you talk about ads much.
00:29:26
People tell me they're seeing some ads
00:29:28
in the experiment in the free version.
00:29:30
What is your commitment to the ad
00:29:32
version? You guys got a little uh
00:29:34
trolled by Anthropic during the Super
00:29:37
Bowl? Oh, you're going to have ads. But
00:29:38
is ads the solution to making this free
00:29:41
for the world?
00:29:42
>> Yeah. So, first of all, on the ad front,
00:29:45
you know, we we want to stick by our
00:29:46
principles. We want to make sure that
00:29:48
you know you're always getting the best
00:29:50
result based on the model, not by
00:29:52
something that was sponsored. So that
00:29:53
has to hold true. And I think the second
00:29:56
thing is that we'll always provide a
00:29:58
free a tier, sorry, an ad free tier for
00:30:00
people that just don't want ads. But
00:30:02
with that said, if you took if you took
00:30:06
Fiji says this really well, if you know
00:30:09
Google and Meta had a baby, it would be
00:30:11
chat GPT because what you have in Google
00:30:13
search and by the way we know we have at
00:30:16
least 11% of the search market. It's a
00:30:18
lot more because actually when you do a
00:30:20
Google search and the page refreshes
00:30:22
that counts as one in chat GBT when you
00:30:25
do a whole conversation where you might
00:30:26
ask 50 questions that also only counts
00:30:28
as one. So in reality, we have a much
00:30:30
higher portion, very high intent. That
00:30:33
is great for advertisers because I'm
00:30:35
effectively telling you what I'm doing,
00:30:36
right? I want really cool shoes to sit
00:30:39
on the stage. I'm telling you what I
00:30:41
want to go by. In Meta's case, right,
00:30:43
they use this like people like you sort
00:30:46
of intent. So they have the demographic.
00:30:48
We have more than that because we have
00:30:50
memory, right? I just told you it knows
00:30:52
who I am. So imagine putting memory and
00:30:54
context next to intent. You should have
00:30:56
a very potent ad platform which gives
00:30:59
you an ability to offer up massive
00:31:02
access to the world writ large because
00:31:05
now you can pay for it. And I think back
00:31:07
to a question you asked Freeberg like if
00:31:09
you look at um the revenue per token
00:31:12
right now. If I was optimizing only for
00:31:15
today, I would give every token to the
00:31:17
API,
00:31:17
>> right?
00:31:18
>> Every token to the API order of
00:31:20
magnitude more than to the consumer.
00:31:22
However, I told you we're playing our
00:31:24
own game. We have a strategy where we
00:31:26
believe there's an AI infrastructure
00:31:28
layer, a utility like electricity, and
00:31:30
in a future state, you'll want to be
00:31:32
able to serve the world at large,
00:31:34
consumers, small businesses, large
00:31:37
enterprises, governments. That's our
00:31:39
strategy.
00:31:39
>> Ladies and gentlemen, the CFO of Open
00:31:41
AI, Sarah Frier. Well done. Fabulous.
