OpenAI's annualized revenue run rate grew 70% quarter over quarter in the third quarter, exiting at close to $50 billion, according to CFO Sarah Friar. She tells Deirdre Bosa what drove the quarter, from new models to an enterprise business she says doubled, and how OpenAI plans to keep business customers from switching. Friar explains when its Dots assistant could reach consumers, how OpenAI plans to cover the cost of serving AI agents with subscriptions and ads, why she calls cheap Chinese open models a myth and how she thinks about AI sovereignty.
00:00 OpenAI's Blockbuster Quarter 00:20 Enterprise Doubles 01:09 Keeping Enterprise Customers 01:49 Cyber as a Hook 02:38 When Dots Reaches Consumers 04:01 Will Consumer Habits Stick? 05:17 The Cost of Serving Agents 07:07 Busting the Open Model Myth 08:11 What Customers Want From Open Models 09:41 The Sovereignty Question 11:06 Going to Market With AWS
AI Generated·May contain errors
Video Summary
AI Generated
The company reported a blockbuster Q3 with a 70% quarter-over-quarter growth in annualized revenue run rate, exiting at nearly $50 billion in ARR. Growth was driven by frontier research (models 5.6 and Astra) and a doubling of the enterprise business, which now accounts for 50% of total revenue. The company is balancing its massive consumer base with expanding enterprise and SMB offerings.
Q3 was a blockbuster quarter for us.
If you look at our actual annualized revenue run rate,
it grew 70% quarter over quarter.
So we exited close to about $50 billion in ARR.
What drove that, number one, frontier research.
So we had a great model at the beginning of the
quarter, 5.6. And then Astra, of course,
launched in early September. Our
Business, or enterprise business,
doubled quarter over quarter, and that leaves us now
about 50-50, a very unified platform and very balanced,
so 50% of revenue coming from enterprises.
What's driving that is products like coding,
cyber, starting to get more virtual assistants
like DOTS into the enterprise, and then finally,
the business skew, so that's our SMB offering.
That grew more in Q3 than it had done in the prior six quarters,
and that's because I think
we're seeing a huge opportunity in the market
with small businesses, and we want to make
sure we show up for you, whether you're a consumer,
a developer, a small business, a
large enterprise, or a government.
And increasingly, that is the enterprise side.
OpenAI started with a gigantic consumer business,
and I know you've shifted that a lot to enterprise,
which probably accounts for a lot of that growth.
How sticky is that business?
Do you find that switching is harder
because I know you mentioned models, but it feels
like still every week there's a new model.
How do you make sure those customers stay?
Yeah, so I think that's exactly the job now
as we move into being this balanced company.
So I don't want to leave our consumers behind.
We should talk about them because 1.2 billion of them show up.
Every single week, and that business grew 20% quarter over quarter.
But on the enterprise side,
what helps them keep coming back for more is you lift out
of just the frontier model layer
and into products where they're using their context
from their organization, they're using memory,
and they're building hooks.
And that means that you're going from a wedge in,
like, let's take a company like Visa.
First deal we've done with them is cyber.
So cyber is very top of mind right now, right?
We just saw the PHP vulnerability
that could have brought down 70% of the internet, which
we created a patch for.
We see a company like Amgen that might actually start
with us on something that's more around
life sciences and drug discovery,
but right now really cares about their IP and making
sure that is secure.
So cyber is also becoming a hook there.
So for us, it's the applied side of research.
How do we put more pieces into enterprises that keep them coming back?
for more and it's even
in something like dots virtual assistants right that comes
becomes sticky very fast
because people just get habituated to using it and so
I'm particularly excited about
that new rollout that's coming. Right I was
telling you before we got on camera that I've
been using dots and I've been using
a lot of the personal agents but I
actually happen to like dots the best
when are you able to roll that out to consumers is that in the plans?
Yeah, it's a great question.
So maybe the first question is,
why did we start in the enterprise?
So I usually call this training at altitude
because enterprise customers typically have...
More difficult problems to solve, more complex,
certainly, and there's more advantage to being
connected into many systems for an enterprise, right?
My dot connects to my email, connects to my calendar,
connects to even some of my financial
systems.
And so if we get that right,
then we have the confidence to start bringing it down into
the consumer.
I think the second thing is we have to have
the right model for the right task at the
right price, because if you're going
to take something to 1.2 billion subscribers,
It has to work from an overall margin perspective and cost perspective.
So I think we just announced a few days ago
that we're now using more of our Terra model
underneath DOTS, and that'll be another way
to bring it to the consumer.
And then I think the third piece is, look,
we live in a compute-constrained world, so
we have to go first to where we can maximize
from a business model perspective.
But over time, we absolutely see this coming down to consumers.
You'll see these moves even by the end of the year.
Do you worry, though, that habits will be harder
to break the longer you wait to introduce
it to consumers?
Obviously, you know,
MetaMuse has been very aggressive in rolling it out.
It's collecting context all the time
that you guys don't give it to consumers.
So how much does that?
concern you? Yeah, so there's absolutely habituation
that happens, right? We know
1.2 billion users. What are they doing?
So 300 million of them are doing health and wellness
every single week. They ask chat, a health and wellness question.
That actually gathers a lot
of context. 200 of them are asking personal finance questions.
So you see them digging
into areas or you see consumers digging
into areas that can feel a little complicated.
Sometimes you might be a little embarrassed
to go talk to someone about until you feel a little
little bit more read up on the subject and all of that is
context. So I'm not worried that we're going
to get left behind on that front because we've been
working with consumers now for multiple years
and there's a lot of memory and context already
stored within ChatGPT.
So by the time your dot shows up, it's actually able to take advantage
of all of that memory.
I also think too that this is the advantage of being this two-sided
platform in a way. Sometimes things start in the enterprise
and then people pull them down
into their personal life, like way back when email was really that.
And then sometimes it goes the other way,
like people got their iPhone back in the day
and then they wanted to bring it to work.
And so I think we benefit by that cross-pollination
that can happen across both the segments.
Right, I wanna get to the cost part of this,
which you mentioned, it's expensive to serve, right?
Because you're essentially giving everyone a computer
in the cloud to do tasks on their behalf.
When does that get – are you looking at a world,
especially post-midterms, post everything that we know now about
the challenges data centers could face, maybe 2027,
but more likely 2028, how are you building that into your strategy
and plans and the cost of compute?
Yeah. So maybe just starting, first of all,
there's a cost to serve, but there's also the
revenue you can gain from it.
And our expectation on DOTS is that we'll be able to continue with
more of a subscription revenue base,
certainly in the consumer, to kick it off. Why are we
confident there? We already see
that we hit 100 million subscribers in Q3. And so remember,
back in the day, people were like,
no one's ever going to subscribe for a consumer product, right?
This is why other platforms had moved more quickly to advertising.
I think we've proven that that's not the case, right?
When you give people real value, they're willing to pay.
That said, 90% of our base is free,
and that's where something like ads kicks in.
And by the way, we think that there can be ad models
and even commerce models for DOTS
as it comes down into the long tail of consumers.
But short run, we will launch that with a subscription pricing.
And then where you're going on the cost to serve,
it does involve a lot of compute.
Thank
The advantage of being a frontier model company
is the frontier model is more and more involved
in training and creating the lower end.
And more nano micro models that are very cheap
to serve for still very good intelligence.
And our entire strategy is predicated
you.
on being along that Pareto efficient frontier.
So a model like Luna.
...is actually about one-fifth, one-sixth the cost of,
say, a Chinese open-source model.
That is one of the myths of 2026 that I think we have to myth-bust.
When we start being able to put products like DOTS
on a Luna model or a Terra model,
you really start to see the cost to serve become much more reasonable.
It allows us to both drive upward...
Upward-growing gross margins, upward-trending gross margins,
sorry, while still being able
to get the product into people's hands.
Right.
So the myth of 2026 is that Chinese models
are almost as capable and a lot cheaper, but
OpenAI, increasingly in the frontier labs,
increasingly have models that are cheap and
open weight as well.
Is that right?
So how much do you see that adoption happening?
Because when you look on sort of third-party websites,
you see that the Chinese models
are still the ones accelerating in terms of at least developers.
Yeah, well, actually, if you look at the stats on Luna,
when we dropped the pricing on Luna,
we dropped pricing 80% about two weeks after we launched it,
and we saw use go up almost 10x. So
I think actually you're seeing a shift in tokens.
What I hear when a customer comes and talks to me
about open weight, they're really asking for three things.
One is definitely cost.
And so we have to keep a strategy that is you will always get
the right token for the right task at the right price.
And in fact, more and more, we're creating model families,
so not just a single model.
So if you're using, say, the Terra family or the Sol family,
the model will decide for you in that moment, which.
particular model it should pull on
to create the right cost profile.
The second thing I hear customers ask for
is they want to be able to do more reinforcement learning.
They want kind of that guardrail down situation
where they're confident that their data
is being used appropriately to reinforcement learn
for their environment, for their particular use case.
And so a big part of that is building trust
that when you wanna use your data like that,
it is safe and secure in your environment.
I think the second thing is also being mindful of,
like, when can our general purpose model
perhaps accelerate you?
Like, in areas of cyber, you could absolutely
do more training in your environment.
Like, this is something Amgen
is doing right now to build agents for cyber.
But they're benefiting from the breakthroughs we're seeing
with a very cyber-capable model
like Astra.
And then the third thing I think customers
are asking for is often sovereignty.
That is a more difficult question,
particularly when you get outside the U.S., when I travel
around the world.
People want to know they can depend on us to be there.
And I think that's a longer conversation.
But, you know, I had an interesting presentation in Ireland.
I was back there for the Irish government currently has
the presidency of the EU.
And so the economist of the EU spoke and said,
look, we can talk about sovereignty of models,
but let's be clear using Stanford data.
The U.S. had 52 frontier models in 2025.
China had 37 and Europe had one.
And so in a world.
So when you say that the
sovereignty question
is complicated, are you saying
that governments
Countries need their own AI, but they're using a mixture of models.
They need access to all of them.
I don't think saying that they need their own AI
is actually a sensible outcome today.
When I give you those stats that say,
here's where frontier research is happening, here's
where frontier outcomes are happening.
That said, it's like any customer you go work with.
First of all, you have to show up in a very collaborative way.
Like, what are the problems you're trying to solve?
Because if I solve your problems, you
You will be happy.
And number two is that we are trusted.
And I think that's been a big part of the build in enterprise
is that we can show up as a trusted partner.
Just one example of that,
going to market with Amazon and AWS Bedrock.
That's allowed us to get into government
and into financial services
because that's already a trusted platform
that now we are able to go to market with.
And that's only true as of the last couple of months.