The Long View: Interview with Sarah Friar, CFO of OpenAI

Close

The Long View: An Interview with Sarah Friar, CFO of
OpenAI

“The Angle” Music

Cold OPEN “I don't think any one company, any
one country even, can do what we're doing.”

Eric Veiel

Welcome back to “The Angle from T. Rowe Price”, a podcast
for curious investors. Just a reminder that outside of the U.S., this podcast
is for investment professionals only.

In this episode, I am very excited to welcome Sarah
Friar. Sarah is the CFO of Open AI, one of the companies at the heart of
the AI revolution, and specifically the company that brought us Chat GPT.

Sarah, welcome to “The Angle” Thank you for having us.

Sarah Friar

Such a pleasure to have you in our space. And such a
pleasure to get to do this conversation.

Eric Veiel

Yeah. The vibe in the office is, is really cool. Is it a lot
of excitement going on today, I guess, for Sora 2 launch.

Sarah Friar

The kingdom of Sora? That is a pretty funny meme that people
have all over the building today. So, exciting.

Eric Veiel

I have to admit, I usually feel a little bit old when I
visit, you know, companies in the Bay area. I felt exceptionally old walking
through the lobby today.

Sarah Friar

Not at all. Not at all. But, yes. No, it's, it has an
intensity about OpenAI. We, everyone wants to be in the office. It's very
research driven culture. It feels a lot like a university lab setting. But then
days like today, Sora, bring, like a whimsy. Like a fun to it, that I really
enjoy.

Eric Veiel

How do you keep that culture of, like, a research driven
environment with, you know, the commercialization so front and center as you
get bigger and deploy more capital, which we'll certainly talk about in a bit?

Sarah Friar

Yeah, it's very top of mind because I do think it is, huge
differentiator for us, that we are research first. And, you know, we grew up,
we're almost a ten-year-old company at this stage. But the first seven years of
life were only a research lab. And in fact, wasn't until November of 22. Right.
That ChatGPT kind of burst on the scene. And frankly, we put a subscription on
it to pay for compute, because we didn't really know what else to do, but it
wasn't really with the intent. We're a very mission driven company. So, the
making sure that as we are building, we're building AGI for the benefit of
humanity. And I think that also goes back to kind of a research like culture
where, you know, they're really going deep on things because they have a
passion for it, but because they believe it's gonna have a great outcome in the
world. So, models that might transform health care, models that might transfer
transform education, models that might transform the whole creativity, like art
and so on. And so to, to nurture that, you know, number one is our space, as
we've seen, we try to not be overly corporate, frankly. I think number two is
always being clear that it is a research first culture. And we're going to
optimize for that. And that can be hard for people. For example, coming from
the kind of normal, normal tech companies where engineering is always. The
kings and queens. And that's a shift.

And then I think the third thing is really keeping an
incredibly high bar, right? There's not that many researchers in the world that
can work on what we're working on, and we don't want to deplete that bar. At
the same time, I think we're the only company that has built kind of a farm
team system where we're taking folks fresh out of the university setting, often
post a PhD, so they might be a little further along in their journey, but
really turning them into the best AI researchers on the planet. And that's a
reason why you often see people come at us for talent, because we are the folks
that have this farm team effectively being grown. But it's but it's very
deliberate, that nurturing of that ecosystem.

Eric Veiel

It is a really unique company in so many ways. One of which
is. Right. So, it's private, unlike not all, but some of your biggest, you
know, compatriots in this space. But it's also part nonprofit, part for profit.
Just explain a little bit for people, you know, just briefly how that structure
is and how that benefits. You've kind of alluded to it a bit, but I'm just
yeah.

Sarah Friar

So, sitting atop OpenAI is the non-for-profit. And again, it
is mission aligned to how do we build AGI for the benefit of humanity. And it's
really important. It's often why researchers will come to work here and not go
to other tech companies writ large. Underneath we do have a for profit entity
today. So, an LLC. But as we move towards the recap that we've been working
towards we’ll flip into being a PBC, a public benefit corp. The importance of a
PBC versus a typical C Corp is really that you can maintain two things. You
have both fiduciary duty to your shareholders as a board member. But
importantly, you also have a fiduciary duty to the mission. And we felt that
that was important. Going back to the cultural question you just asked, it was
important to show researchers that that's who we would be, but I think it's
also important as a signal to shareholders. Right. In the end, we want to bring
on the shareholders who understand that importance and that we're going to
balance those two things and not just be shareholder value only, but we will do
both. And there are other examples of great PBCs out there like Lemonade. Warby
Parker I think is the PBC. Ben and Jerry's was before they were bought. And
today folks like Novo Nordisk are a good example of even a foundation that sits
atop, for profit. So, well, it's, it can be a little different. It's not, this
is not a place that we're wildly innovating.

Eric Veiel

So, let's talk a little bit about where you are. It's,
you're in such a unique seat, right? As you mentioned, ten-year-old company.
You're the CFO. And normally, if you're maybe the CFO of a ten-year-old
company, you've got a little bit of capital to play around with. You are
deploying capital at such an amazing pace. Maybe just update us a little bit
because it's changing in real time. A big announcement today out of Korea as
well. And in terms of further the announcement, just how you're thinking about
capital deployments and sort of where you are in the announcements that you've
done so far.

Sarah Friar

Yeah, let me back up for a second and say, just first of
all, we're building a backbone for a new era. In the same way that you built
the railways, you built the electricity pylons that lit up cities and towns and
ultimately rural populations, building the internet. So, we firmly believe
we're at the beginning of that new era, and as we bring on this age of
intelligence, right, we started with in the beginning with ChatGPT, kind of
more of a chat bot, moving to reasoning. So, the big breakthrough of last year
was the reinforcement learning that happens in a post training world. To this
year, really a world of agents and starting to see agents that can do really
long, we call them long horizon tasks for you. So, you can ask an agent to
code. It can go off and code for three, five, seven, 12 hours, by itself, to
create code that you will push to production. For a kind of non-coder, it could
go off and shop for me. I was just in the market for an advent calendar. I hope
my daughter's not listening, but it's been our thing since she was about two
years old.  And so rather than me trawling many sites, I sent my agent off
to do it. So, we're just starting to kind of scratch the surface of how this
will change our lives. And we think it will take a vast amount of compute. So
that's the backbone.

At the same time, we're building a strong company that can
have fundamentally sound business. So, 700 million weekly active users, I'm
sure we'll get more into the business, but with a goal. We've never seen a
company grow at this pace ever. And with a goal to clearly be able to finance a
lot of our future today with free cash flow, always the best way to fund a
business to do it. But really only recently also taking on equity investing. I
know, folks here are big numbers, but really, last year was the first external
equity round we did, aside from our Microsoft investment. And so this year, the
bigger round that we're completing right now, the $41 billion round, that's
really only the second round we've done.

And then I think the third piece is it's going to take an
ecosystem. I don't think any one company, any one country, even, can do what
we're doing. It's kind of a space race. And so that's why you see us right now
really trying to make the ecosystem rise with us. So, it might be the supply
chain, right? The people who build chips. Clearly our deal with Nvidia, the
people who build data centers, the people who build all of the other equipment
that sits inside these data centers. But it's also going to take governments
coming to the table, whether it's to relax some of the regulations to help us
build faster, to bring power on more quickly. But it might also be to think of
financing there. And then, of course, what we can do to finance our own
business directly.

Eric Veiel

Right. So, it’s really stunning to think about just so many
of these different components. So, I want to get to power in a second. But
before we do just the, the amount of capital that you all are deploying, some
coming from, as you mentioned, your first couple of rounds of of other
investors, the big influx from Nvidia, how are you as the CFO thinking about
the return on investment just from that, and over what time horizon is a
reasonable way to think about, you know, using your metaphor of the railroads,
of laying all of this track?  Before we start running the trains.

Sarah Friar

I mean, the good news is that unlike the railroads, where
you actually had to have a track that went from, say, New York to Philly, and
then you could put a train on it.

Eric Veiel

Let’s say New York to Baltimore.

Sarah Friar

Well to Baltimore. Good point, actually. Sorry. Wrong.
Wrong, major to choose. We are already showing value out of the gate, right? We
don't. The track can still be unfinished at the end. That's why you have 700
million weekly actives already using your technology. And I think ChatGPT has
revolutionized how many people just go about their day to day. Certainly, in a
personal setting, it doesn't matter who I talked to. They're like, oh, I use
chat for my workout routine. I use chat to help me plan dinner tonight. I use
chat to help me learn a new language. Had this wonderful story at dinner the
other night from a CFO. I was expecting a very CFO answer, and he was talking
about taking his daughter on a camping trip and how as they were lying there
looking up at the stars, she wanted to know where the Big Dipper was. So, he
pulled out chat and he was like, hey, chat, help me find the Big Dipper. And
then they, he told her all these stories about all of the different
constellations. It was such a cool story.

So, that's in your personal life, but in your professional
life, 500 million enterprise seats sold, like the fastest enterprise business
to get to $1 billion era. Today, we're putting up $1 billion in revenue per
month, right? That is already showing that the value is here. So, when you talk
about ROI, I look at things like the cohort of, for example, ChatGPT paid
users. What's the gross margin on that business? Actually, a very healthy gross
margin. Our enterprise gross margin already looks a lot like enterprise
software. So, I know that the early cohorts are already making a return, that
if I stopped the clock today, didn't do any more investment from a research
perspective. I actually think we'd already be throwing off quite a bit of cash.
So that's what investors see when they make an investment decision. But of
course, to your point, we are still investing for future, because we believe
we're only getting started.

Eric Veiel

So, let's talk a little bit about growing the top line and
some of the different evidence. So, 700 million weekly users, I think at the
beginning of the year was like 400 or maybe even a little less than that. And
we're sitting here and like, not even done the year. That's incredible. That's
just on the consumer side, as you said before, we jump to the enterprise just a
little bit about how you're thinking about the consumer opportunity and maybe
monetizing that. There's obviously a paid component to it. How are you thinking
about that?

Sarah Friar

Yeah. So today, again, just put a little parameters around
those numbers. Today about 90% of our users are outside the United States.
Which even that I find amazing? Because most tech companies where you're born
tends to be the predominant. Yeah. And if you look at it really follows
population. So, India, Indonesia, Vietnam, Philippines, Brazil. Right. Anywhere
with large populations of the world attracts. The second thing I would say is
today we only monetize about 5% of our user base. So, 95% is effectively free.
Now the free is a strategy. I mean, number one, it's mission aligned. So, we
want to make sure everyone benefits not just for people that can pay, but it is
also a way to let people edge into what is intelligence means to me. Right.
When I say to someone, you're going to get access to intelligence. They look at
me and they're not quite sure, and they might just move on, but when I say
you're going to get access, it can be in a very profound way, like something
that actually helps you do a task better, faster, or sometimes it can be in a
very fun way, like image gen or now creating a Sora video. Right. There's,
there's a whole gamut of how people experience it. If we don't let you try it,
then you don't know what you're missing out on. And the way our business model
works is on the pure subscription side. We want you to use until you run out of
tokens. Then you hit a paywall, and if you're getting real value, you will
subscribe.

And so, the more that we release, much more sophisticated
features that take a lot of tokens. So, video gen is a good example, but as is
deep research, the more likely you are to hit that moment where we can get you
to subscribe. And again, it's subscribing for value, not just because we're
tricking you into that. So, I actually love that the base of our business is
subscription. How that will grow is more users, more people jumping that wall.
And actually, the other thing I'm very proud of this year is releasing our low
end, price point, what I call GPT ChatGPT Go, so that in countries like India,
Indonesia, where a $20 per month subscription effectively is cost prohibitive.
Now we're unlocking those markets. So that is growing the subscription
business.

Clearly, though, we have a lot of people highly engaged with
our products. And so, you can start to see the the beginnings of other business
models. Just on Monday we released our commerce option. And so why we went
there first is we could see the use case of ChatGPT was often to help me
purchase something. Right. I'm, I'm a new mother. You know, I'm expecting
twins. What's the best stroller to buy in San Francisco? I'm thinking of
upgrading my car to an electric vehicle. What should I choose? So very much bottom
of funnel, often high consideration, high value purchases. And so, by getting
into that commerce flow, we're getting paid for the outcome, which frankly,
actually resonates really well with our researchers. It keeps ChatGPT as a true
north of giving you the best answer.

Eric Veiel

Not the paid answer.

Sarah Friar

Super important.

Eric Veiel

That's an important part of Sam's view on this whole.

Sarah Friar

Very important. But it's a massive market, right? If we can
take a chunk of the world's commerce or retail, that is a massive new business
model, which I get really excited about. So, I think we're just scratching the
surface on many of these things. Longer term, you know, I said, this year is
the year of agents. I think we're starting to experience what an agent could
feel like even within ChatGPT. You see it today, actually, with both deep
research, which originally was set up to be almost like a research agent. So,
if I need to go do, if I'm going on vacation, I want to do a deep look at these
are three destinations I could go to. Go and think about this for ten, 15
minutes. It'll come back with an amazing response that might take a travel
agent maybe days to come back with. So that's an example of an agent work. But
also, I like study mode because that's a good example of an agent working as a
teacher. So, if you ask a question, instead of giving you the immediate answer,
it may actually start by asking you a question back. So, it uses much more of a
Socratic method, like tell me about plans. So instead of like, here's the
answer, you know, it will instead say, oh, you wanted to learn about plants.
Tell me a little bit about what you know, or what's your interest in plants, is
it because you're going to plant something? Or are you interested in the oxygen
cycle or whatever, the rain cycle, pick it and as it goes, it will even say,
I'm going to tell you a little bit more, and then I'm going to ask you a few test
questions to see if what I'm saying is resonating. That is, it already
transforming into an agent. We just take away this word agent, which is not
very engaging for most humans and turn it into a teacher and how we go monetize
that in future. I think there's a lot of optionality on that front too, to get
to value.

Eric Veiel

So, you've also and that's absolutely fascinating
when you think about just the original chat bot and how you've already, in
such a short amount of time, thought about so many different avenues to engage
with the, the consumer. You've also really now started to think about moving up
the stack from an application layer perspective. Talk a little bit about how
you're thinking about moving into the application layer.

Sarah Friar

Yeah. So, if we were, if we actually were on video, as we
joked about earlier, I have a chart that I do love that is a classic kind of
stack diagram. And in the middle is a bar that’s our LOM, and it says where do
we start, we want to be the frontier model. Always research culture. Super
important. However, you come up a level out of the model layer, you had an API,
API layer, which is actually something that has really resonated with
enterprises and developers. Out of the API layer, you get to an application
layer and then ultimately to I need a better word, but substrate layer. I'm
sure you will come to work on consumer hardware but think of that as one
element of the stack. We're actually going down as well. We’re starting to
think about our own ability to build first party data center technology, maybe
down to the chip layer. So, OpenAI has become in just a two-ish year period,
two-and-a-half-year period, a very multi-dimensional business. I like that, you
know, all good investors love it because you can build moats in different parts
of that stack. You can actually trade economics around, because if you're
competitors only in one place, you can actually often be quite aggressive, say,
on pricing in that place, because you can subsidize with other revenue streams
elsewhere. So, I think getting to that multi-dimensional, look, field, shape of
the company has been really important.

You asked me about the application layer specifically. For
sure, and there I think it actually trended a little bit into how we think
about go to market on the enterprise side.

Eric Veiel

That's a good transition.

Sarah Friar

Today, we think about our both everything from the smallest
of businesses. So, people who come from being a consumer into being maybe a
prosumer into what we call our ChatGPT for business, which is our self-serve
product, which goes up to about 199 seats, up to about 200. And from there, of
course, we move straight into the world of enterprise. And today, a year ago, I
felt I showed up and every CEO, CFO, CIO, whatever C-level person would say,
what do we do with that? I know it's important, but what do we do? Today, I
think we've gotten into a better motion of, you know, step one is just deploy
ChatGPT wall to wall; sounds self-serving, but your people will probably more
likely do find the innovation than you will telling them. We try to make that
easy today by doing a more of a platform based price so that you don't have to
make tradeoffs like, okay, well, I'm okay paying this more expensive thing for
my coders, but I'm not going to have my finance people use it; be a terrible
decision, by the way. To now, being able to say, okay, everyone can have it. So
that allows you to actually create some horizontal outcomes as well, like
performance management. Right. Everyone does performance reviews. You can have
a custom GPT to help people write that performance review. And everyone in the
company can use that. Like sounds like a silly example. When you work at
companies, you know how painful that is.

Eric Veiel

On my flight out here today, one of my coworkers used it to,
to write his self eval. It was amazing what he put into this and what he got
out of it.

Sarah Friar

It's amazing. It’s the best. But now we're starting to say
to companies, okay, so that's great that got you started, but now we need kind
of it to link back to your top-down strategy. So, if you are, for example, the
world's largest retailer, you want to have a strategy for natural language
search because you want consumers to have better basket sizes. But to be able
to do more sophisticated things, right. If you think about search and retail
has made us turn into crazy people. We write in things like black shoes, high
heels. Like I walked into a store, I wouldn't say black shoes, high heels. I
would say I'm going to a party this weekend. I want to wear black. I would love
shoes that are comfortable right? You would have a conversation. Search, inside
chat, is now bringing back conversational, but that's great for a retailer
because that usually leads to a bigger basket size, more customer retention.
But then there's also, say my merchandizing strategy. And then there might be
my supply chain strategy. And so now there's a top-down push of where do I want
to inject this intelligence directly into my organization. And so that allows
you to go from kind of a base price that you're paying us now onto a very
specific, outcome driven decision making on what enterprise product am I going
to deploy here and what am I going to pay for it? And so that has been a really
major move I've seen just in the last year with the C-level suite.

Eric Veiel 

What is the, when you talk to different companies or more
importantly, when you observe different companies trying to figure out how to
deploy, because everybody does have this sort of panic, feel like I know my
competitors are doing it. There's this unbelievable paranoia in the corporate
world about it. What are the best practices that you've seen of, you know,
companies engaging with their employees in a positive, productive way to really
diffuse the technology into the the fabric of the companies.

Sarah Friar

And so, I do think it starts with this idea of just give
people access to Chat wall to wall, what we started to do, even in my own
finance departments, we do a hackathon about every 3 to 6 months. Sounds fancy.
We just brought in our sales team. We give people space. That might be a half
day, could be a full day. It often starts with just like a pen on a piece of
paper. Like what are some of the more routine tasks that you do. You do that
individually, then you set a grip. I might put the tax team together. The
procurement team, they share notes, they look at, okay, what are the places
where we're spending a lot of time repeating the same work? Or were there
mistakes that often occur and then that allows them to do this first step of
maybe a custom GPT. I'll give you an example. When we were raising that first
round last September, one of the more routine tasks you get when you raise
money is diligence, right? Investors all like the same thing.

Eric Veiel

Well familiar with it.

Sarah Friar

So, you get a lot of similar questions. And so, there was
like, a night we were all working late. And, you know, we were looking at a
question. I was like, we answered that over here, pull that in. But then that
slight different bit of nuance, pull it from over here. And then we were kind
of cobbling it together, and we recognized that if we created a custom GPT
where we fed it our corporate model, our corporate presentation, the other
diligence questions we'd already answered, we gave it a persona which included, you
know, remember, this information is confidential. Don't go outside to the web.
Just use internal information. Your tone should be professional. Like like make
it snappy. Investors don't have a ton of time. Get to the point. But also on the
edge, be a little bit salesy, right? We are all selling. And we put through the
first set of diligence questions. We were almost weeping with joy and how a
something that would have taken probably 2 to 3 hours took like a minute, maybe
less, in front of our very eyes. Now there's still a human in the loop to do a
final check, and maybe you just add a few extra notes because you understand
exactly what that investor was saying, but just the transformation.

Today, we're now moving down a path where we're doing things
like our Sox control process, where we're starting to think about our pc
readiness, for example, we can do that using voice, right? If you want to check
if someone's controls are correct, you sit with them, you have a conversation.
What do you do if you see this? You take that transcript back to your desk. You
can turn it into a Sox control, like what's called an RCM control (Risk Control
Matrix) in seconds. And instead of spending your time with all of the
transposing and transcribing and interviewing, you're moving to the inside.
Like, do you have a control problem or not? If we have a control problem, what
else should we do? And by the way, that's a great moment to even go back to
ChatGPT and say, help me think this through. So, just so many ways that we see
enterprises doing what I just described in my microcosm of my finance team at
OpenAI, but doing it inside their corporation, but in their marketing team,
their comms team, their product team. And then you start getting into worlds
like coding. They're already doing massive amounts, right? I know one of the
largest companies I get to spend my time with, over 40% of the code they're
creating today is coming agentic ally created.

Eric Veiel

Yeah, that's amazing. We, we actually just did a hackathon
using our subscription, just last week, and the results were really
interesting. Where we're actually going to do an expo next week where we're
bringing in a bunch of different vendors, obviously, including all of you in
your team to, to think about how we can push into the, into the organization.
But it really does feel like it has to be sort of a top down as well as bottoms
up led process to get, to get the whole enterprise moving.

Sarah Friar

The best companies I've seen, like the early adopters;
Morgan Stanley is a great example in their wealth management division. One of
our earliest adopters really figured out how to use ChatGPT in that moment
where a wealth advisors maybe doing a first conversation with a potential
client, so discovery. So instead of just doing the discovery, the chat bot was
actually giving them examples of like products to sell. So, the, the first call
went from just pure discovery to discovery plus sales. So, you took out a
meeting probably. The conversion rates went up. The advisors were happier. They
were bringing on bigger books. All good. But the top-down motion was actually
the board. The board said that we want to see this, either X amount more
revenue or Y amount cost reduction. And that is your go get executive team. And
so suddenly you had this really nice sandwich of like very top-down meets very
grassroots bottoms up. And I think Morgan Stanley, what they created with that
first wealth management installation. Now we've seen companies everyone in the
space implement effectively.

Eric Veiel

One of the other areas that I think everybody's really
anxious about is you guys acquired Johnny Ive's company not too long ago. So,
there's lots of talk and speculation about what the hardware approach might be
for OpenAI. Before you dodge my question, I will just tell you, I, when I, when
I had this podcast with Jensen a couple months back, he talked about the
marrying of robotics and agentic AI and how that's really not that far away.
It's happening now until it gets to be something like very common is probably
only a couple of years out. Is there a path forward for you all on the hardware
side? And just like, theoretically, how might you approach that?

Sarah Friar

Yeah. So, I mean, first of all, kind of start by just
talking about the multi-modality of models. So, when we think about a
definition of AGI, it's not just tech space, right? All arms are very good
today a text-based ingestion and ultimately output. And even though I know
ChatGPT has multi-model capabilities, I still tend to go back to talking with
my phone call it. My teenagers are experts, but we are also already quite good
at being able to speak, to listen, and to use visual, right, to be able to recognize
a plant, for example, by just showing it the camera from your phone. Think of
Sora as both a fun breakthrough, right? Incredible to be doing just natural
language and creating short form videos and all the fun you can have with it.
But it is also pulling in a lot of data about the world in a much more
three-dimensional, kind of human like way. As we get into devices, we think the
same way. That these devices, whether it's a humanoid robotic type outcome or
whether it's a Johnny IV device, which I cannot tell you about, and I'm going
to sit on my hands so I don't actually talk with my hands here. But these are
ways to think about them, first and foremost, as a way to collect data that
will help ultimately models be fully human-like in the way we experience the
world. Right? There's not a lot of collection of data today on just how we
experience the world in terms of space. And what I mean by that is not, you
know, the planet, but literally I move my hands when I talk. I'm actually
creating a lot of data right now as I sit in front of you, but nothing collects
that data. Maybe if we were being videoed there would be some bits and bytes.
Seeing my hands move. But no, nothing is really collecting that feel. And to
get to a moment of humanoids, and I think countries like China are actually
very far ahead of us right now. Like there's so much to learn, right? When I
pick this glass of water up, I know exactly how much pressure to put. So, I
don't like break the glass over glass but not drop it. When I set it on the
table, I know how to set it down without making a noise. That is an incredibly
sophisticated use of data.

And so as we think about the substrate, it's both data
collection, probably first and foremost, which most people wouldn't say first,
but then it is also, every era of computing has had a substrate that's really
brought it to the fore, right, when we went from, to the PC era until you had a
PC, didn't really unlock it for most people. When we got to the mobile era, if
you didn't have a phone, didn't recognize it had a GPS inside, how could you
have ever gotten to apps like Uber or this idea you would DoorDash something on
the fly, or that a map would just work for you? And so, we think similarly with
AI, we're not making use of all these multimodal capabilities, so how do we
kind of bring that to people in a much more kind of fun, but very utilitarian
type way?

Eric Veiel

That's, it's fascinating. It's going to change; the world is
going to change so much in such a short amount of time. And it already is. But
we're we're all anxious to see where where you all go with that. Let's talk a
little bit more about moving down into the data center, data factory,
intelligent factory creation aspect of your business. And how, you come to that
decision to, to move that way. I mean, historically, that's the more
capital-intensive business, potentially the lower return business. And certainly,
if we want to think about last cycles going back to the internet buildout, you
know, the stock charts of the companies that spent all the money to build that
infrastructure did not end well. So, what about that decision and what are you
all trying to accomplish there?

Sarah Friar

So, I think at it's, just in it's simplest, we think that
infrastructure is destiny, and that it is a moat to have access to maximum
compute. So that is very simple way to put it. If we could do all that just
through third parties, we would do that. But the question is, can we get enough
compute if we do that? I'll give you a more sophisticated answer. And remember,
I'm speaking as someone who in currently is, does not have enough compute to
meet the demand in front of us, which is a surefire way to not have as much
revenue as we should have. You can kind of see, even in our Sora launch,
there's a reason why it launched to Canada, the US, only on iOS, and it's an
invite only system because you can clearly see we're trying to balance the
excitement of finally getting this model out to learn with getting in front of
users who can benefit from it, but still understanding that we are dramatically
short GPUs, or however you want to talk about your compute metric of choice.

On the decision to go from third party to first party, some
of that is just about continuing to diversify out where we get compute from.
Not even a year ago, we really only had compute from Microsoft. And that
constrained our business because even Microsoft couldn't build it fast enough.
But I think on top of that, it didn't allow us to see the true kind of breadth
of what was available out there right. There, many other suppliers helps us see
who might do things differently, better, more cheaply, less latency, like less
risk, be more available to us in parts of the world, you know, one CSP versus
another. It just really opened our eyes to the the huge step up you get when
you bring that diversity to bear. So that was kind of a huge moment for us. It
also made us start to think about how much IP we are creating, because we are
the forefront company in AI, so we understand better than anyone. What does the
right build look for inferencing? What does the right build look like when
you're inferencing agents versus a chat bot? What does the right build look
like when you're inferencing a Sora real time video? What does it look like
when you're training? We think the laws of scaling are firmly intact. So,
we want very, very large fabrics, right? More compute means more data
means the smartest kind of models around, keeps giving us more intelligence.
So, we don't want to feel constrained even just in the in the scale of those
fabrics.

So, I think our debate internally is, right now, by just
moving into at least some first party, it probably carries over enough benefits
to the rest of our supply chain. You know, just pulling their socks up a little
bit more. That it's worth it for that alone. But I do think a lot about folks
like an Amazon. Right? When Amazon was starting, there was a lot of skepticism
about like, why were they building all these servers and building this thing
called we didn't have a word for it, but cloud computing, like, I remember when
investors talked about it as if it was the same server selling books at night
were somehow selling off. Do you remember when there was this whole thing?

Eric Veiel

Yeah, it was a seminal moment for our our analysts who saw
the value of that. And the stock was getting killed, and he was absolutely
brilliant and saw through it. And it was when we doubled down on that stock.

Sarah Friar

And look at AWS today, is a business in its own right. It
has a 40%, I think, EBITDA margin, 35, 40% EBITDA margin. I don't want us to
look back and feel like we missed a mega move to also be whatever the CSP of
the future automates, the AI service provider of the future. And so, we have to
start creating that capability as an option. And then I think we can decide, do
we want to really invest hard behind that option? Like, is that the right thing
to do to become a more capital-intensive business? Is there a way to do it with
partners so it moves it off our balance sheet, or are we better off just
staying as the layer above the CSPs, whatever we want to call them in the
future? I don't know what the exact right answer is, but I think it behooves us
to keep the optionality in place.

Eric Veiel

So, one of the big limiting factors of all of this
infrastructure build that we haven't talked about yet is power. And, we were
sort of, I was emailing with some of our team, on my flight out here, we were
talking about the announcement that you had around, Korea and just the amount
of, DRAM that you were, you know, setting up here and what that would translate
into, in terms of the Nvidia equipment that you could use and then ultimately
what that would mean for power. And basically, we got to it's like the state of
Texas in terms of power that you would need to to put all of that to work. So
how are you all thinking about real breakthroughs in power generation? Because
it's clearly a limiting factor. And it looks like in, from a US perspective,
it's a limiting factor for us as a country. But maybe take that second. But for
you as a company?

Sarah Friar

Yeah. We do think that that is becoming the next real choke
point in the supply chain is the access to power. We've certainly, as a
country, made a lot of moves just in the last year or so to speed that up. And
clearly, it's always a balance of thinking about kind of a more traditional
power sources to renewables. But even then, there's still not enough. And by
the way, we think a big part of intelligence and model like novel model
breakthroughs, maybe helping us solve this conundrum of power. Can we get to
nuclear fusion, for example? Like, I think we're only getting there through
this AI generation taking us because it's going to be both about the physics of
that, but also just even the materials to build small form nuclear reactors and
so on. Like so that is it's an interesting thing where the intelligence itself
is actually what helps the intelligence.

It doesn't frighten me when you say like the state of Texas,
because when I look at us as a company, we're already talking about, this year
we'll do about two gigawatts of, if I translate it back into power, about two
gigawatts. Two years ago, we were doing about 200MW. So, if you just think
about that, like a 10x in two years. So, if we ten x again in two years, could
we be doing 20GW in two years? Like that's already probably bigger than the
state of Texas? I know Ireland is where I go because I'm Northern Irish, like
the country of Ireland does about seven gigawatts of power. So, it's bigger
than a country.

And that sounds crazy. And yet like, you know, to hark back
to the last answer, when we started to tell our supply chain that we needed two
gigawatts in 2025, they thought we were crazy and would not build it. And yet
here we are, where we could probably use 2.5, maybe three gigawatts. I don't
know where it would stop this year, but it certainly is not enough at two
gigawatts. And that is another reason to want to control our own destiny.
That's why I started with that line. Infrastructure is destiny. Yeah, because
the only thing I know for sure is no compute equals no revenue. So that is a
sure thing in life. But on the power front, it is also taking us out there in
the world to look at a lot more sources of power, countries who are rich in
renewables, that can have a very different price point. I think it starts to
get into where you are going with geopolitical outcomes as well. I mean, China
is building a grid at a significantly faster pace than the United States. And,
you know, this is a place to for kind of we've used terms like democratic AI,
you know, whatever, whatever you want to call it. But like there is going to be
a set of allies that come together. And if we want AI to keep coming along in
an aligned way, and in our view, part of alignment is that democracies are a
path that we want to stay on. It's who are the allies we can bring with us on
that path. And many of them have access to a lot less expensive power.

Eric Veiel

So putting these dots together, it wouldn't be crazy to
expect some kind of announcements, partnerships around things like power gen
from you all in the maybe not too distant future, because it seems like if it
is going to be the binding constraint, you're going to get ahead of it as fast
as you can. And there's a business strategy perspective here too.

Sarah Friar

There is.

Eric Veiel

Securing that is going to give you a huge advantage over
everybody else who's racing to do what you're doing.

Sarah Friar

And all the hyperscalers are doing it already. And they’re
are our competitors in this AI race. The good/bad thing about taking that more
optionality there is it's also very fungible assets. Those are assets that can
be bought and sold. So going back to my seat as a CFO, they give me a lot less
indigestion. Like these things tend to be very finance able. They also tend to
be sellable as we decide where exactly we want to put our chips on the table.
So, you may see us taking a lot of options to then make decisions over the long
run. Right? You already see Microsoft doing that, Google doing that all the
time. They trade in and out of data center footprint, partially because we come
along and we say we need our next fabric to be x-size, so that all needs to be
interconnected. Therefore, that data center that you have in Wisconsin is not
as good as this data center footprint down here in Texas, where we already have
three other footprints and we want to interconnect them. So, you see people
doing a lot of trading across power, PPA land and then, you know, less on the
kit and so on.

Eric Veiel

So, let's transition a little bit because there's there's so
much to talk about. We could go for hours. And I want to be respectful of your
time. But one of the questions I wanted to ask you is are there any use cases,
or any scenarios that we just should not use AI for? 

Sarah Friar

Gosh, that is a good question. I've never been asked, it
quite that way.

Eric Veiel

Better questions, better outcomes. T. Rowe Price.

Sarah Friar

I think. Good job. In terms of like what not to use it for
it. I, I think we generally feel like AI has applicability across really
anything that humans touch. However, that doesn't mean to say we don't want to
have guardrails around things. So, safety is a really important criteria for us
as we build models, test models, put models out into the world. We had a lot of
debate around our most recent decision around open source, for example. And we
typically look at CBRN. So cyber, radiological, nuclear and bio risks. And we
score our models across those four criteria. And so, we will absolutely hold
back models where we think they are not, where there's more risk or danger
being injected into the system. Bio is the one I would say further along
because it's harder to do things, for example, in nuclear because you got a own
like weapons grade plutonium or whatever. The other area that we look a lot at
risk is just the, to how the models interacting with a human. So, for example,
we did pull back one of our models, I think at 4.5, ChatGPT 4.5 or in that
zone, because it was very sycophantic. So, it's different sort of risk
entirely. But it was absolutely causing people to get more and more engaged
because it was telling you lots of nice things like that. That performance
review would have been amazing.

Eric Veiel

So good.

Sarah Friar

Yeah. So, we just want to be really mindful of when we want
to slow people down. Or maybe insert something, if we can see that a behavior
is beginning to unfold. People are using ChatGPT for a lot of medical reasons,
for a lot of very personal reasons, more so than maybe any technology ever
before. People will talk to ChatGPT about relationship problems or, I want to
get a divorce. Should I get, I mean, things that you will not even share with
another human. And so, to your point, like, should ChatGPT be deployed
everywhere? It's a very good question, right? If it helps someone really think
through a tough problem that they cannot share elsewhere, I think that's only a
good thing. But I think we need to be very careful that in that moment where a
human needs to be pop back into the loop, maybe a therapist or a doctor or
someone talking to a teenager, that we are also allowing for avenues for that
to happen. And we we rolled out more parental controls this week actually for
exactly that reason.

Eric Veiel

Yeah, it, it actually is a good segue. I was going to ask
you, as we talk a little bit about, you know, these tools, just advice that you
give to younger people coming up, either at your company or, we both have kids
about the same age in sort of that college age about, you know, how to
think about their future in terms of it is really going to be probably
radically different in terms of the workforce, etc.. So, what do you tell
people when they come to you and say, like, what should I do?

Sarah Friar

Yeah, I mean, first and foremost, I like this quote. I've
used it a few times now, right? “I don't think your job is going to be taken by
AI, but it's probably going to be taken by someone using AI.” So, make sure
you're the one using AI. So, you know, experiment, listen, learn, be curious.
Ask other people what are they doing. For students, we just put out 100 things
like 100 questions for ChatGPT for students.

Eric Veiel

I saw it as I was waiting. Yeah.

Sarah Friar

It’s super fun. I mean, you kind of just look at some of the
ways your, your peers are using it. So, number one. Number two, like
within schools, I would love to get it even more deeply embedded into the
education system, starting early in the education system. But in the university
setting for sure. I think there was a lot of fear a year ago, two years ago.
And I think you've seen the flip of professors, tutors and so on. Just saying
this is a must. And we see it just in seat deployments, right? Cal State universities
just applied for 500,000 seats. ASU 180,000. Oxford University, my alma mater,
just deployed wall to wall. So, I think people are understanding that this is
the shape of the future. So, we need to bring our students along. Our staff,
faculty, tutors and so on need to be there.

We rolled out an AI job platform, with the idea of how can
we help people reskill. And we put together a set of AI certifications. So, we
are trying to prompt the market as well through that. And I think another
couple things I usually say to my kids like is number one, still, you have to
learn to think, and you have to learn to do hard stuff. Like my daughter's a
chemist. I'm like, it's got like, you still need a lot of help in chemistry,
but AI's now, like the company she worked in the summer doing small molecule
drug discovery. They're using AI to help do a lot of the early design. What
could this molecules look like? But a human still needs to put all the
ingredients together, test it, and then feed that back into kind of keep
iterating with the model. That to me is a perfect harmony of human plus model,
doing something to extend new novel outcomes.

And then I think the other thing we need to think about,
generally speaking, is reskilling ourselves throughout our lifespan. Like, the
jobs most of us do today, our parents could never have imagined. We are all
going to work much longer, assuming we want to. We're going to live longer. The
idea that you work to your 65/70, yeah, people are having to work because they
also want to save for retirement. But we're also healthier. We're also able to
stay in the workforce for longer, like how do we make use of that, and don't
have this kind of outdated concept that education ends at 22 or 23? It's crazy.
Right, why wouldn't we? Why wouldn’t you go back to college in your 40s or your
60s? Like, there's this wonderful quote from one of the Supreme Court judges who
at like 95, started to learn Greek and someone said, why are you learning
Greek? And his response was like, well, why wouldn't I learn Greek? And I just
feel like that sort of learning mindset is what I tell my kids all the time. Be
curious, learn. And there is no substitute for hard work.

Eric Veiel

Excellent. When we think about, the business itself and
where it is today, if we were sitting here, let's not say next year, let's say
in three years from now, what do you think would be the most surprising
development if AI hasn't taken off? What if it hasn't?

Sarah Friar

What is the reason?

Eric Veiel

That it hasn't taken off.

Sarah Friar

Such a shock. I don't really know if I can see a reason for
it not to take off because the value created already is so extreme to me. Like
perhaps that would be that we stay too superficial, and we don't get it deeply
enough embedded. We kind of allow ourselves like, usually what I find on my
team when I look at like, why there's an uneven distribution of usage, the
people not using it a lot, usually don't use a lot because they're too busy.
And so, I have to help them slow down to do the thing that will speed them up
on the other side. So, I think that that would be one reason why I would see a
company say fail is that the CEO, the people who can actually cause a pause or
cause a change of direction, don't don't push it fast enough.

And there's probably examples, like if you go back to the
internet age, coming upon us where the people who really didn't grab hold of,
say, e-commerce in the retail space. It was a very slow burn, but in the end,
it ended poorly. And this era is happening much faster, and so I think if the I
don't believe there's a world where AI isn’t successful, but I think there will
be lots of examples where companies are not successful or even individuals. And
I think it's because you're not grabbing hold of it quickly enough right now
and getting that hackathon done and getting that into your system. If you feel
behind, I always say to everyone “I feel behind”. Yeah, I come to work every
day and I'm like, I'm not using our product enough. I feel like I've missed the
feature we've rolled out, or I didn't know that the agent could go do the thing
that is a completely normal way to feel. But instead of making it, you feel
fearful, like it's a kind of a joyousness of like, oh, there's something new to
discover out there.

Eric Veiel

There’s so much out there. Yeah, so much more to do. Sarah,
this has been fantastic. Thank you so much for taking the time to talk to us.

Sarah Friar

Super fun. Thank you.

Eric Veiel

Thank you.

Again, I’m Eric Veiel. Thank you for listening to The Angle.
We look forward to your company on future episodes. You can find more
information about this and other topics on our website. Please rate and
subscribe wherever you get your podcasts. The Angle, better questions, better
insights. Only from T. Rowe Price.

This podcast episode was recorded in October of 2025 and is
for general information and educational purposes only.  Outside the United
States, it is for investment professional use only.  It is not intended to
be used by persons in jurisdictions which prohibit or restrict distribution of
the material herein.

This podcast does not give advice or recommendations of any
nature; or constitute an offer or solicitation to buy or sell any security in
any jurisdiction. Prospective investors should seek independent legal,
financial, and tax advice before making any investment decision. Past
performance is not a guarantee or a reliable indicator of future results. All
investments are subject to risk, including the possible loss of principal.

Discussions relating to specific securities are
informational only, and are not recommendations, and may or may not have been
held in any T. Rowe Price portfolio. Any forward-looking statements are for
discussion purposes only and are never guaranteed. There should be no
assumption that the securities were or will be profitable. T. Rowe Price is not
affiliated with, or a subsidiary of, any company discussed.

The views contained herein are those of the speakers as of
the date of the recording and are subject to change without notice. The views
may differ from those of other T. Rowe Price associates and/or affiliates.
Information is from sources deemed reliable but not guaranteed. Please visit
https://www.troweprice.com/en/uk/insights/the-long-view-OpenAI for full global
issuer disclosures.

This podcast is copyright by T. Rowe Price, 2025.

ID0008488

202510 - 4811214